summaryrefslogtreecommitdiffstats
path: root/meta-oe
diff options
context:
space:
mode:
Diffstat (limited to 'meta-oe')
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106759.patch545
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106761.patch188
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106762.patch1355
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106763.patch96
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106764.patch25
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106766.patch25
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106768.patch182
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106769.patch1294
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106770.patch138
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106771.patch211
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106772.patch350
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106773.patch119
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106775.patch67
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106776.patch46
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106777.patch192
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106778.patch225
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106781.patch741
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106782.patch27
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-4_6-branch-linaro-backports.inc18
-rw-r--r--meta-oe/recipes-devtools/gcc/gcc-common-4.6.inc2
20 files changed, 5845 insertions, 1 deletions
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106759.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106759.patch
new file mode 100644
index 000000000..c51576794
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106759.patch
@@ -0,0 +1,545 @@
12011-06-20 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
2
3 Backport from mainline.
4 2011-06-03 Julian Brown <julian@codesourcery.com>
5
6 * config/arm/arm-cores.def (strongarm, strongarm110, strongarm1100)
7 (strongarm1110): Use strongarm tuning.
8 * config/arm/arm-protos.h (tune_params): Add max_insns_skipped
9 field.
10 * config/arm/arm.c (arm_strongarm_tune): New.
11 (arm_slowmul_tune, arm_fastmul_tune, arm_xscale_tune, arm_9e_tune)
12 (arm_v6t2_tune, arm_cortex_tune, arm_cortex_a5_tune)
13 (arm_cortex_a9_tune, arm_fa726te_tune): Add max_insns_skipped field
14 setting, using previous defaults or 1 for Cortex-A5.
15 (arm_option_override): Set max_insns_skipped from current tuning.
16
172011-06-14 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
18
19 Backport from mainline.
20 2011-06-02 Julian Brown <julian@codesourcery.com>
21
22 * config/arm/arm-cores.def (cortex-a5): Use cortex_a5 tuning.
23 * config/arm/arm.c (arm_cortex_a5_branch_cost): New.
24 (arm_cortex_a5_tune): New.
25
26 2011-06-02 Julian Brown <julian@codesourcery.com>
27
28 * config/arm/arm-protos.h (tune_params): Add branch_cost hook.
29 * config/arm/arm.c (arm_default_branch_cost): New.
30 (arm_slowmul_tune, arm_fastmul_tune, arm_xscale_tune, arm_9e_tune)
31 (arm_v6t2_tune, arm_cortex_tune, arm_cortex_a9_tune)
32 (arm_fa726_tune): Set branch_cost field using
33 arm_default_branch_cost.
34 * config/arm/arm.h (BRANCH_COST): Use branch_cost hook from
35 current_tune structure.
36 * dojump.c (tm_p.h): Include file.
37
38 2011-06-02 Julian Brown <julian@codesourcery.com>
39
40 * config/arm/arm-cores.def (arm1156t2-s, arm1156t2f-s): Use v6t2
41 tuning.
42 (cortex-a5, cortex-a8, cortex-a15, cortex-r4, cortex-r4f, cortex-m4)
43 (cortex-m3, cortex-m1, cortex-m0): Use cortex tuning.
44 * config/arm/arm-protos.h (tune_params): Add prefer_constant_pool
45 field.
46 * config/arm/arm.c (arm_slowmul_tune, arm_fastmul_tune)
47 (arm_xscale_tune, arm_9e_tune, arm_cortex_a9_tune)
48 (arm_fa726te_tune): Add prefer_constant_pool setting.
49 (arm_v6t2_tune, arm_cortex_tune): New.
50 * config/arm/arm.h (TARGET_USE_MOVT): Make dependent on
51 prefer_constant_pool setting.
52
532011-06-14 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
54
55 Backport from mainline
56 2011-06-01 Paul Brook <paul@cpodesourcery.com>
57
58 * config/arm/arm-cores.def: Add cortex-r5. Add DIV flags to
59 Cortex-A15.
60 * config/arm/arm-tune.md: Regenerate.
61 * config/arm/arm.c (FL_DIV): Rename...
62 (FL_THUMB_DIV): ... to this.
63 (FL_ARM_DIV): Define.
64 (FL_FOR_ARCH7R, FL_FOR_ARCH7M): Use FL_THUMB_DIV.
65 (arm_arch_hwdiv): Remove.
66 (arm_arch_thumb_hwdiv, arm_arch_arm_hwdiv): New variables.
67 (arm_issue_rate): Add cortexr5.
68 * config/arm/arm.h (TARGET_CPU_CPP_BUILTINS): Set
69 __ARM_ARCH_EXT_IDIV__.
70 (TARGET_IDIV): Define.
71 (arm_arch_hwdiv): Remove.
72 (arm_arch_arm_hwdiv, arm_arch_thumb_hwdiv): New prototypes.
73 * config/arm/arm.md (tune_cortexr4): Add cortexr5.
74 (divsi3, udivsi3): New patterns.
75 * config/arm/thumb2.md (divsi3, udivsi3): Remove.
76 * doc/invoke.texi: Document ARM -mcpu=cortex-r5
77
78=== modified file 'gcc/config/arm/arm-cores.def'
79--- old/gcc/config/arm/arm-cores.def 2011-01-03 20:52:22 +0000
80+++ new/gcc/config/arm/arm-cores.def 2011-06-14 16:00:30 +0000
81@@ -70,10 +70,10 @@
82 /* V4 Architecture Processors */
83 ARM_CORE("arm8", arm8, 4, FL_MODE26 | FL_LDSCHED, fastmul)
84 ARM_CORE("arm810", arm810, 4, FL_MODE26 | FL_LDSCHED, fastmul)
85-ARM_CORE("strongarm", strongarm, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, fastmul)
86-ARM_CORE("strongarm110", strongarm110, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, fastmul)
87-ARM_CORE("strongarm1100", strongarm1100, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, fastmul)
88-ARM_CORE("strongarm1110", strongarm1110, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, fastmul)
89+ARM_CORE("strongarm", strongarm, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, strongarm)
90+ARM_CORE("strongarm110", strongarm110, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, strongarm)
91+ARM_CORE("strongarm1100", strongarm1100, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, strongarm)
92+ARM_CORE("strongarm1110", strongarm1110, 4, FL_MODE26 | FL_LDSCHED | FL_STRONG, strongarm)
93 ARM_CORE("fa526", fa526, 4, FL_LDSCHED, fastmul)
94 ARM_CORE("fa626", fa626, 4, FL_LDSCHED, fastmul)
95
96@@ -122,15 +122,16 @@
97 ARM_CORE("arm1176jzf-s", arm1176jzfs, 6ZK, FL_LDSCHED | FL_VFPV2, 9e)
98 ARM_CORE("mpcorenovfp", mpcorenovfp, 6K, FL_LDSCHED, 9e)
99 ARM_CORE("mpcore", mpcore, 6K, FL_LDSCHED | FL_VFPV2, 9e)
100-ARM_CORE("arm1156t2-s", arm1156t2s, 6T2, FL_LDSCHED, 9e)
101-ARM_CORE("arm1156t2f-s", arm1156t2fs, 6T2, FL_LDSCHED | FL_VFPV2, 9e)
102-ARM_CORE("cortex-a5", cortexa5, 7A, FL_LDSCHED, 9e)
103-ARM_CORE("cortex-a8", cortexa8, 7A, FL_LDSCHED, 9e)
104+ARM_CORE("arm1156t2-s", arm1156t2s, 6T2, FL_LDSCHED, v6t2)
105+ARM_CORE("arm1156t2f-s", arm1156t2fs, 6T2, FL_LDSCHED | FL_VFPV2, v6t2)
106+ARM_CORE("cortex-a5", cortexa5, 7A, FL_LDSCHED, cortex_a5)
107+ARM_CORE("cortex-a8", cortexa8, 7A, FL_LDSCHED, cortex)
108 ARM_CORE("cortex-a9", cortexa9, 7A, FL_LDSCHED, cortex_a9)
109-ARM_CORE("cortex-a15", cortexa15, 7A, FL_LDSCHED, 9e)
110-ARM_CORE("cortex-r4", cortexr4, 7R, FL_LDSCHED, 9e)
111-ARM_CORE("cortex-r4f", cortexr4f, 7R, FL_LDSCHED, 9e)
112-ARM_CORE("cortex-m4", cortexm4, 7EM, FL_LDSCHED, 9e)
113-ARM_CORE("cortex-m3", cortexm3, 7M, FL_LDSCHED, 9e)
114-ARM_CORE("cortex-m1", cortexm1, 6M, FL_LDSCHED, 9e)
115-ARM_CORE("cortex-m0", cortexm0, 6M, FL_LDSCHED, 9e)
116+ARM_CORE("cortex-a15", cortexa15, 7A, FL_LDSCHED | FL_THUMB_DIV | FL_ARM_DIV, cortex)
117+ARM_CORE("cortex-r4", cortexr4, 7R, FL_LDSCHED, cortex)
118+ARM_CORE("cortex-r4f", cortexr4f, 7R, FL_LDSCHED, cortex)
119+ARM_CORE("cortex-r5", cortexr5, 7R, FL_LDSCHED | FL_ARM_DIV, cortex)
120+ARM_CORE("cortex-m4", cortexm4, 7EM, FL_LDSCHED, cortex)
121+ARM_CORE("cortex-m3", cortexm3, 7M, FL_LDSCHED, cortex)
122+ARM_CORE("cortex-m1", cortexm1, 6M, FL_LDSCHED, cortex)
123+ARM_CORE("cortex-m0", cortexm0, 6M, FL_LDSCHED, cortex)
124
125=== modified file 'gcc/config/arm/arm-protos.h'
126--- old/gcc/config/arm/arm-protos.h 2011-05-03 15:17:25 +0000
127+++ new/gcc/config/arm/arm-protos.h 2011-06-14 16:00:30 +0000
128@@ -219,9 +219,14 @@
129 bool (*rtx_costs) (rtx, RTX_CODE, RTX_CODE, int *, bool);
130 bool (*sched_adjust_cost) (rtx, rtx, rtx, int *);
131 int constant_limit;
132+ /* Maximum number of instructions to conditionalise in
133+ arm_final_prescan_insn. */
134+ int max_insns_skipped;
135 int num_prefetch_slots;
136 int l1_cache_size;
137 int l1_cache_line_size;
138+ bool prefer_constant_pool;
139+ int (*branch_cost) (bool, bool);
140 };
141
142 extern const struct tune_params *current_tune;
143
144=== modified file 'gcc/config/arm/arm-tune.md'
145--- old/gcc/config/arm/arm-tune.md 2010-12-20 17:48:51 +0000
146+++ new/gcc/config/arm/arm-tune.md 2011-06-14 14:37:30 +0000
147@@ -1,5 +1,5 @@
148 ;; -*- buffer-read-only: t -*-
149 ;; Generated automatically by gentune.sh from arm-cores.def
150 (define_attr "tune"
151- "arm2,arm250,arm3,arm6,arm60,arm600,arm610,arm620,arm7,arm7d,arm7di,arm70,arm700,arm700i,arm710,arm720,arm710c,arm7100,arm7500,arm7500fe,arm7m,arm7dm,arm7dmi,arm8,arm810,strongarm,strongarm110,strongarm1100,strongarm1110,fa526,fa626,arm7tdmi,arm7tdmis,arm710t,arm720t,arm740t,arm9,arm9tdmi,arm920,arm920t,arm922t,arm940t,ep9312,arm10tdmi,arm1020t,arm9e,arm946es,arm966es,arm968es,arm10e,arm1020e,arm1022e,xscale,iwmmxt,iwmmxt2,fa606te,fa626te,fmp626,fa726te,arm926ejs,arm1026ejs,arm1136js,arm1136jfs,arm1176jzs,arm1176jzfs,mpcorenovfp,mpcore,arm1156t2s,arm1156t2fs,cortexa5,cortexa8,cortexa9,cortexa15,cortexr4,cortexr4f,cortexm4,cortexm3,cortexm1,cortexm0"
152+ "arm2,arm250,arm3,arm6,arm60,arm600,arm610,arm620,arm7,arm7d,arm7di,arm70,arm700,arm700i,arm710,arm720,arm710c,arm7100,arm7500,arm7500fe,arm7m,arm7dm,arm7dmi,arm8,arm810,strongarm,strongarm110,strongarm1100,strongarm1110,fa526,fa626,arm7tdmi,arm7tdmis,arm710t,arm720t,arm740t,arm9,arm9tdmi,arm920,arm920t,arm922t,arm940t,ep9312,arm10tdmi,arm1020t,arm9e,arm946es,arm966es,arm968es,arm10e,arm1020e,arm1022e,xscale,iwmmxt,iwmmxt2,fa606te,fa626te,fmp626,fa726te,arm926ejs,arm1026ejs,arm1136js,arm1136jfs,arm1176jzs,arm1176jzfs,mpcorenovfp,mpcore,arm1156t2s,arm1156t2fs,cortexa5,cortexa8,cortexa9,cortexa15,cortexr4,cortexr4f,cortexr5,cortexm4,cortexm3,cortexm1,cortexm0"
153 (const (symbol_ref "((enum attr_tune) arm_tune)")))
154
155=== modified file 'gcc/config/arm/arm.c'
156--- old/gcc/config/arm/arm.c 2011-05-11 14:49:48 +0000
157+++ new/gcc/config/arm/arm.c 2011-06-14 16:00:30 +0000
158@@ -255,6 +255,8 @@
159 static void arm_conditional_register_usage (void);
160 static reg_class_t arm_preferred_rename_class (reg_class_t rclass);
161 static unsigned int arm_autovectorize_vector_sizes (void);
162+static int arm_default_branch_cost (bool, bool);
163+static int arm_cortex_a5_branch_cost (bool, bool);
164
165
166 /* Table of machine attributes. */
167@@ -672,12 +674,13 @@
168 #define FL_THUMB2 (1 << 16) /* Thumb-2. */
169 #define FL_NOTM (1 << 17) /* Instructions not present in the 'M'
170 profile. */
171-#define FL_DIV (1 << 18) /* Hardware divide. */
172+#define FL_THUMB_DIV (1 << 18) /* Hardware divide (Thumb mode). */
173 #define FL_VFPV3 (1 << 19) /* Vector Floating Point V3. */
174 #define FL_NEON (1 << 20) /* Neon instructions. */
175 #define FL_ARCH7EM (1 << 21) /* Instructions present in the ARMv7E-M
176 architecture. */
177 #define FL_ARCH7 (1 << 22) /* Architecture 7. */
178+#define FL_ARM_DIV (1 << 23) /* Hardware divide (ARM mode). */
179
180 #define FL_IWMMXT (1 << 29) /* XScale v2 or "Intel Wireless MMX technology". */
181
182@@ -704,8 +707,8 @@
183 #define FL_FOR_ARCH6M (FL_FOR_ARCH6 & ~FL_NOTM)
184 #define FL_FOR_ARCH7 ((FL_FOR_ARCH6T2 & ~FL_NOTM) | FL_ARCH7)
185 #define FL_FOR_ARCH7A (FL_FOR_ARCH7 | FL_NOTM | FL_ARCH6K)
186-#define FL_FOR_ARCH7R (FL_FOR_ARCH7A | FL_DIV)
187-#define FL_FOR_ARCH7M (FL_FOR_ARCH7 | FL_DIV)
188+#define FL_FOR_ARCH7R (FL_FOR_ARCH7A | FL_THUMB_DIV)
189+#define FL_FOR_ARCH7M (FL_FOR_ARCH7 | FL_THUMB_DIV)
190 #define FL_FOR_ARCH7EM (FL_FOR_ARCH7M | FL_ARCH7EM)
191
192 /* The bits in this mask specify which
193@@ -791,7 +794,8 @@
194 int arm_arch_thumb2;
195
196 /* Nonzero if chip supports integer division instruction. */
197-int arm_arch_hwdiv;
198+int arm_arch_arm_hwdiv;
199+int arm_arch_thumb_hwdiv;
200
201 /* In case of a PRE_INC, POST_INC, PRE_DEC, POST_DEC memory reference,
202 we must report the mode of the memory reference from
203@@ -864,48 +868,117 @@
204 {
205 arm_slowmul_rtx_costs,
206 NULL,
207- 3,
208- ARM_PREFETCH_NOT_BENEFICIAL
209+ 3, /* Constant limit. */
210+ 5, /* Max cond insns. */
211+ ARM_PREFETCH_NOT_BENEFICIAL,
212+ true, /* Prefer constant pool. */
213+ arm_default_branch_cost
214 };
215
216 const struct tune_params arm_fastmul_tune =
217 {
218 arm_fastmul_rtx_costs,
219 NULL,
220- 1,
221- ARM_PREFETCH_NOT_BENEFICIAL
222+ 1, /* Constant limit. */
223+ 5, /* Max cond insns. */
224+ ARM_PREFETCH_NOT_BENEFICIAL,
225+ true, /* Prefer constant pool. */
226+ arm_default_branch_cost
227+};
228+
229+/* StrongARM has early execution of branches, so a sequence that is worth
230+ skipping is shorter. Set max_insns_skipped to a lower value. */
231+
232+const struct tune_params arm_strongarm_tune =
233+{
234+ arm_fastmul_rtx_costs,
235+ NULL,
236+ 1, /* Constant limit. */
237+ 3, /* Max cond insns. */
238+ ARM_PREFETCH_NOT_BENEFICIAL,
239+ true, /* Prefer constant pool. */
240+ arm_default_branch_cost
241 };
242
243 const struct tune_params arm_xscale_tune =
244 {
245 arm_xscale_rtx_costs,
246 xscale_sched_adjust_cost,
247- 2,
248- ARM_PREFETCH_NOT_BENEFICIAL
249+ 2, /* Constant limit. */
250+ 3, /* Max cond insns. */
251+ ARM_PREFETCH_NOT_BENEFICIAL,
252+ true, /* Prefer constant pool. */
253+ arm_default_branch_cost
254 };
255
256 const struct tune_params arm_9e_tune =
257 {
258 arm_9e_rtx_costs,
259 NULL,
260- 1,
261- ARM_PREFETCH_NOT_BENEFICIAL
262+ 1, /* Constant limit. */
263+ 5, /* Max cond insns. */
264+ ARM_PREFETCH_NOT_BENEFICIAL,
265+ true, /* Prefer constant pool. */
266+ arm_default_branch_cost
267+};
268+
269+const struct tune_params arm_v6t2_tune =
270+{
271+ arm_9e_rtx_costs,
272+ NULL,
273+ 1, /* Constant limit. */
274+ 5, /* Max cond insns. */
275+ ARM_PREFETCH_NOT_BENEFICIAL,
276+ false, /* Prefer constant pool. */
277+ arm_default_branch_cost
278+};
279+
280+/* Generic Cortex tuning. Use more specific tunings if appropriate. */
281+const struct tune_params arm_cortex_tune =
282+{
283+ arm_9e_rtx_costs,
284+ NULL,
285+ 1, /* Constant limit. */
286+ 5, /* Max cond insns. */
287+ ARM_PREFETCH_NOT_BENEFICIAL,
288+ false, /* Prefer constant pool. */
289+ arm_default_branch_cost
290+};
291+
292+/* Branches can be dual-issued on Cortex-A5, so conditional execution is
293+ less appealing. Set max_insns_skipped to a low value. */
294+
295+const struct tune_params arm_cortex_a5_tune =
296+{
297+ arm_9e_rtx_costs,
298+ NULL,
299+ 1, /* Constant limit. */
300+ 1, /* Max cond insns. */
301+ ARM_PREFETCH_NOT_BENEFICIAL,
302+ false, /* Prefer constant pool. */
303+ arm_cortex_a5_branch_cost
304 };
305
306 const struct tune_params arm_cortex_a9_tune =
307 {
308 arm_9e_rtx_costs,
309 cortex_a9_sched_adjust_cost,
310- 1,
311- ARM_PREFETCH_BENEFICIAL(4,32,32)
312+ 1, /* Constant limit. */
313+ 5, /* Max cond insns. */
314+ ARM_PREFETCH_BENEFICIAL(4,32,32),
315+ false, /* Prefer constant pool. */
316+ arm_default_branch_cost
317 };
318
319 const struct tune_params arm_fa726te_tune =
320 {
321 arm_9e_rtx_costs,
322 fa726te_sched_adjust_cost,
323- 1,
324- ARM_PREFETCH_NOT_BENEFICIAL
325+ 1, /* Constant limit. */
326+ 5, /* Max cond insns. */
327+ ARM_PREFETCH_NOT_BENEFICIAL,
328+ true, /* Prefer constant pool. */
329+ arm_default_branch_cost
330 };
331
332
333@@ -1711,7 +1784,8 @@
334 arm_tune_wbuf = (tune_flags & FL_WBUF) != 0;
335 arm_tune_xscale = (tune_flags & FL_XSCALE) != 0;
336 arm_arch_iwmmxt = (insn_flags & FL_IWMMXT) != 0;
337- arm_arch_hwdiv = (insn_flags & FL_DIV) != 0;
338+ arm_arch_thumb_hwdiv = (insn_flags & FL_THUMB_DIV) != 0;
339+ arm_arch_arm_hwdiv = (insn_flags & FL_ARM_DIV) != 0;
340 arm_tune_cortex_a9 = (arm_tune == cortexa9) != 0;
341
342 /* If we are not using the default (ARM mode) section anchor offset
343@@ -1991,12 +2065,7 @@
344 max_insns_skipped = 6;
345 }
346 else
347- {
348- /* StrongARM has early execution of branches, so a sequence
349- that is worth skipping is shorter. */
350- if (arm_tune_strongarm)
351- max_insns_skipped = 3;
352- }
353+ max_insns_skipped = current_tune->max_insns_skipped;
354
355 /* Hot/Cold partitioning is not currently supported, since we can't
356 handle literal pool placement in that case. */
357@@ -8211,6 +8280,21 @@
358 return cost;
359 }
360
361+static int
362+arm_default_branch_cost (bool speed_p, bool predictable_p ATTRIBUTE_UNUSED)
363+{
364+ if (TARGET_32BIT)
365+ return (TARGET_THUMB2 && !speed_p) ? 1 : 4;
366+ else
367+ return (optimize > 0) ? 2 : 0;
368+}
369+
370+static int
371+arm_cortex_a5_branch_cost (bool speed_p, bool predictable_p)
372+{
373+ return speed_p ? 0 : arm_default_branch_cost (speed_p, predictable_p);
374+}
375+
376 static int fp_consts_inited = 0;
377
378 /* Only zero is valid for VFP. Other values are also valid for FPA. */
379@@ -23123,6 +23207,7 @@
380 {
381 case cortexr4:
382 case cortexr4f:
383+ case cortexr5:
384 case cortexa5:
385 case cortexa8:
386 case cortexa9:
387
388=== modified file 'gcc/config/arm/arm.h'
389--- old/gcc/config/arm/arm.h 2011-06-02 12:12:00 +0000
390+++ new/gcc/config/arm/arm.h 2011-06-14 14:53:07 +0000
391@@ -101,6 +101,8 @@
392 builtin_define ("__ARM_PCS"); \
393 builtin_define ("__ARM_EABI__"); \
394 } \
395+ if (TARGET_IDIV) \
396+ builtin_define ("__ARM_ARCH_EXT_IDIV__"); \
397 } while (0)
398
399 /* The various ARM cores. */
400@@ -282,7 +284,8 @@
401 (TARGET_32BIT && arm_arch6 && (arm_arch_notm || arm_arch7em))
402
403 /* Should MOVW/MOVT be used in preference to a constant pool. */
404-#define TARGET_USE_MOVT (arm_arch_thumb2 && !optimize_size)
405+#define TARGET_USE_MOVT \
406+ (arm_arch_thumb2 && !optimize_size && !current_tune->prefer_constant_pool)
407
408 /* We could use unified syntax for arm mode, but for now we just use it
409 for Thumb-2. */
410@@ -303,6 +306,10 @@
411 /* Nonzero if this chip supports ldrex{bhd} and strex{bhd}. */
412 #define TARGET_HAVE_LDREXBHD ((arm_arch6k && TARGET_ARM) || arm_arch7)
413
414+/* Nonzero if integer division instructions supported. */
415+#define TARGET_IDIV ((TARGET_ARM && arm_arch_arm_hwdiv) \
416+ || (TARGET_THUMB2 && arm_arch_thumb_hwdiv))
417+
418 /* True iff the full BPABI is being used. If TARGET_BPABI is true,
419 then TARGET_AAPCS_BASED must be true -- but the converse does not
420 hold. TARGET_BPABI implies the use of the BPABI runtime library,
421@@ -487,8 +494,11 @@
422 /* Nonzero if chip supports Thumb 2. */
423 extern int arm_arch_thumb2;
424
425-/* Nonzero if chip supports integer division instruction. */
426-extern int arm_arch_hwdiv;
427+/* Nonzero if chip supports integer division instruction in ARM mode. */
428+extern int arm_arch_arm_hwdiv;
429+
430+/* Nonzero if chip supports integer division instruction in Thumb mode. */
431+extern int arm_arch_thumb_hwdiv;
432
433 #ifndef TARGET_DEFAULT
434 #define TARGET_DEFAULT (MASK_APCS_FRAME)
435@@ -2018,8 +2028,8 @@
436 /* Try to generate sequences that don't involve branches, we can then use
437 conditional instructions */
438 #define BRANCH_COST(speed_p, predictable_p) \
439- (TARGET_32BIT ? (TARGET_THUMB2 && !speed_p ? 1 : 4) \
440- : (optimize > 0 ? 2 : 0))
441+ (current_tune->branch_cost (speed_p, predictable_p))
442+
443
444 /* Position Independent Code. */
445 /* We decide which register to use based on the compilation options and
446
447=== modified file 'gcc/config/arm/arm.md'
448--- old/gcc/config/arm/arm.md 2011-06-02 15:58:33 +0000
449+++ new/gcc/config/arm/arm.md 2011-06-14 14:37:30 +0000
450@@ -490,7 +490,7 @@
451
452 (define_attr "tune_cortexr4" "yes,no"
453 (const (if_then_else
454- (eq_attr "tune" "cortexr4,cortexr4f")
455+ (eq_attr "tune" "cortexr4,cortexr4f,cortexr5")
456 (const_string "yes")
457 (const_string "no"))))
458
459@@ -3738,6 +3738,28 @@
460 (set_attr "predicable" "yes")]
461 )
462
463+
464+;; Division instructions
465+(define_insn "divsi3"
466+ [(set (match_operand:SI 0 "s_register_operand" "=r")
467+ (div:SI (match_operand:SI 1 "s_register_operand" "r")
468+ (match_operand:SI 2 "s_register_operand" "r")))]
469+ "TARGET_IDIV"
470+ "sdiv%?\t%0, %1, %2"
471+ [(set_attr "predicable" "yes")
472+ (set_attr "insn" "sdiv")]
473+)
474+
475+(define_insn "udivsi3"
476+ [(set (match_operand:SI 0 "s_register_operand" "=r")
477+ (udiv:SI (match_operand:SI 1 "s_register_operand" "r")
478+ (match_operand:SI 2 "s_register_operand" "r")))]
479+ "TARGET_IDIV"
480+ "udiv%?\t%0, %1, %2"
481+ [(set_attr "predicable" "yes")
482+ (set_attr "insn" "udiv")]
483+)
484+
485
486 ;; Unary arithmetic insns
487
488
489=== modified file 'gcc/config/arm/thumb2.md'
490--- old/gcc/config/arm/thumb2.md 2011-05-11 07:15:47 +0000
491+++ new/gcc/config/arm/thumb2.md 2011-06-14 14:37:30 +0000
492@@ -779,26 +779,6 @@
493 (set_attr "length" "2")]
494 )
495
496-(define_insn "divsi3"
497- [(set (match_operand:SI 0 "s_register_operand" "=r")
498- (div:SI (match_operand:SI 1 "s_register_operand" "r")
499- (match_operand:SI 2 "s_register_operand" "r")))]
500- "TARGET_THUMB2 && arm_arch_hwdiv"
501- "sdiv%?\t%0, %1, %2"
502- [(set_attr "predicable" "yes")
503- (set_attr "insn" "sdiv")]
504-)
505-
506-(define_insn "udivsi3"
507- [(set (match_operand:SI 0 "s_register_operand" "=r")
508- (udiv:SI (match_operand:SI 1 "s_register_operand" "r")
509- (match_operand:SI 2 "s_register_operand" "r")))]
510- "TARGET_THUMB2 && arm_arch_hwdiv"
511- "udiv%?\t%0, %1, %2"
512- [(set_attr "predicable" "yes")
513- (set_attr "insn" "udiv")]
514-)
515-
516 (define_insn "*thumb2_subsi_short"
517 [(set (match_operand:SI 0 "low_register_operand" "=l")
518 (minus:SI (match_operand:SI 1 "low_register_operand" "l")
519
520=== modified file 'gcc/doc/invoke.texi'
521--- old/gcc/doc/invoke.texi 2011-05-11 07:15:47 +0000
522+++ new/gcc/doc/invoke.texi 2011-06-14 14:37:30 +0000
523@@ -10208,7 +10208,8 @@
524 @samp{arm1136j-s}, @samp{arm1136jf-s}, @samp{mpcore}, @samp{mpcorenovfp},
525 @samp{arm1156t2-s}, @samp{arm1156t2f-s}, @samp{arm1176jz-s}, @samp{arm1176jzf-s},
526 @samp{cortex-a5}, @samp{cortex-a8}, @samp{cortex-a9}, @samp{cortex-a15},
527-@samp{cortex-r4}, @samp{cortex-r4f}, @samp{cortex-m4}, @samp{cortex-m3},
528+@samp{cortex-r4}, @samp{cortex-r4f}, @samp{cortex-r5},
529+@samp{cortex-m4}, @samp{cortex-m3},
530 @samp{cortex-m1},
531 @samp{cortex-m0},
532 @samp{xscale}, @samp{iwmmxt}, @samp{iwmmxt2}, @samp{ep9312}.
533
534=== modified file 'gcc/dojump.c'
535--- old/gcc/dojump.c 2010-05-19 19:09:57 +0000
536+++ new/gcc/dojump.c 2011-06-14 14:53:07 +0000
537@@ -36,6 +36,7 @@
538 #include "ggc.h"
539 #include "basic-block.h"
540 #include "output.h"
541+#include "tm_p.h"
542
543 static bool prefer_and_bit_test (enum machine_mode, int);
544 static void do_jump_by_parts_greater (tree, tree, int, rtx, rtx, int);
545
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106761.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106761.patch
new file mode 100644
index 000000000..4374e7ed6
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106761.patch
@@ -0,0 +1,188 @@
1 gcc/
2 Backport from mainline:
3
4 Chung-Lin Tang <cltang@codesourcery.com>
5 Richard Earnshaw <rearnsha@arm.com>
6
7 PR target/48250
8 * config/arm/arm.c (arm_legitimize_reload_address): Update cases
9 to use sign-magnitude offsets. Reject unsupported unaligned
10 cases. Add detailed description in comments.
11 * config/arm/arm.md (reload_outdf): Disable for ARM mode; change
12 condition from TARGET_32BIT to TARGET_ARM.
13
14 Chung-Lin Tang <cltang@codesourcery.com>
15
16 * config/arm/arm.c (arm_legitimize_reload_address): For NEON
17 quad-word modes, reduce to 9-bit index range when above 1016
18 limit.
19
20=== modified file 'gcc/config/arm/arm.c'
21--- old/gcc/config/arm/arm.c 2011-06-14 16:00:30 +0000
22+++ new/gcc/config/arm/arm.c 2011-06-27 22:14:07 +0000
23@@ -6488,23 +6488,134 @@
24 HOST_WIDE_INT val = INTVAL (XEXP (*p, 1));
25 HOST_WIDE_INT low, high;
26
27- if (mode == DImode || (mode == DFmode && TARGET_SOFT_FLOAT))
28- low = ((val & 0xf) ^ 0x8) - 0x8;
29- else if (TARGET_MAVERICK && TARGET_HARD_FLOAT)
30- /* Need to be careful, -256 is not a valid offset. */
31- low = val >= 0 ? (val & 0xff) : -((-val) & 0xff);
32- else if (mode == SImode
33- || (mode == SFmode && TARGET_SOFT_FLOAT)
34- || ((mode == HImode || mode == QImode) && ! arm_arch4))
35- /* Need to be careful, -4096 is not a valid offset. */
36- low = val >= 0 ? (val & 0xfff) : -((-val) & 0xfff);
37- else if ((mode == HImode || mode == QImode) && arm_arch4)
38- /* Need to be careful, -256 is not a valid offset. */
39- low = val >= 0 ? (val & 0xff) : -((-val) & 0xff);
40- else if (GET_MODE_CLASS (mode) == MODE_FLOAT
41- && TARGET_HARD_FLOAT && TARGET_FPA)
42- /* Need to be careful, -1024 is not a valid offset. */
43- low = val >= 0 ? (val & 0x3ff) : -((-val) & 0x3ff);
44+ /* Detect coprocessor load/stores. */
45+ bool coproc_p = ((TARGET_HARD_FLOAT
46+ && (TARGET_VFP || TARGET_FPA || TARGET_MAVERICK)
47+ && (mode == SFmode || mode == DFmode
48+ || (mode == DImode && TARGET_MAVERICK)))
49+ || (TARGET_REALLY_IWMMXT
50+ && VALID_IWMMXT_REG_MODE (mode))
51+ || (TARGET_NEON
52+ && (VALID_NEON_DREG_MODE (mode)
53+ || VALID_NEON_QREG_MODE (mode))));
54+
55+ /* For some conditions, bail out when lower two bits are unaligned. */
56+ if ((val & 0x3) != 0
57+ /* Coprocessor load/store indexes are 8-bits + '00' appended. */
58+ && (coproc_p
59+ /* For DI, and DF under soft-float: */
60+ || ((mode == DImode || mode == DFmode)
61+ /* Without ldrd, we use stm/ldm, which does not
62+ fair well with unaligned bits. */
63+ && (! TARGET_LDRD
64+ /* Thumb-2 ldrd/strd is [-1020,+1020] in steps of 4. */
65+ || TARGET_THUMB2))))
66+ return false;
67+
68+ /* When breaking down a [reg+index] reload address into [(reg+high)+low],
69+ of which the (reg+high) gets turned into a reload add insn,
70+ we try to decompose the index into high/low values that can often
71+ also lead to better reload CSE.
72+ For example:
73+ ldr r0, [r2, #4100] // Offset too large
74+ ldr r1, [r2, #4104] // Offset too large
75+
76+ is best reloaded as:
77+ add t1, r2, #4096
78+ ldr r0, [t1, #4]
79+ add t2, r2, #4096
80+ ldr r1, [t2, #8]
81+
82+ which post-reload CSE can simplify in most cases to eliminate the
83+ second add instruction:
84+ add t1, r2, #4096
85+ ldr r0, [t1, #4]
86+ ldr r1, [t1, #8]
87+
88+ The idea here is that we want to split out the bits of the constant
89+ as a mask, rather than as subtracting the maximum offset that the
90+ respective type of load/store used can handle.
91+
92+ When encountering negative offsets, we can still utilize it even if
93+ the overall offset is positive; sometimes this may lead to an immediate
94+ that can be constructed with fewer instructions.
95+ For example:
96+ ldr r0, [r2, #0x3FFFFC]
97+
98+ This is best reloaded as:
99+ add t1, r2, #0x400000
100+ ldr r0, [t1, #-4]
101+
102+ The trick for spotting this for a load insn with N bits of offset
103+ (i.e. bits N-1:0) is to look at bit N; if it is set, then chose a
104+ negative offset that is going to make bit N and all the bits below
105+ it become zero in the remainder part.
106+
107+ The SIGN_MAG_LOW_ADDR_BITS macro below implements this, with respect
108+ to sign-magnitude addressing (i.e. separate +- bit, or 1's complement),
109+ used in most cases of ARM load/store instructions. */
110+
111+#define SIGN_MAG_LOW_ADDR_BITS(VAL, N) \
112+ (((VAL) & ((1 << (N)) - 1)) \
113+ ? (((VAL) & ((1 << ((N) + 1)) - 1)) ^ (1 << (N))) - (1 << (N)) \
114+ : 0)
115+
116+ if (coproc_p)
117+ {
118+ low = SIGN_MAG_LOW_ADDR_BITS (val, 10);
119+
120+ /* NEON quad-word load/stores are made of two double-word accesses,
121+ so the valid index range is reduced by 8. Treat as 9-bit range if
122+ we go over it. */
123+ if (TARGET_NEON && VALID_NEON_QREG_MODE (mode) && low >= 1016)
124+ low = SIGN_MAG_LOW_ADDR_BITS (val, 9);
125+ }
126+ else if (GET_MODE_SIZE (mode) == 8)
127+ {
128+ if (TARGET_LDRD)
129+ low = (TARGET_THUMB2
130+ ? SIGN_MAG_LOW_ADDR_BITS (val, 10)
131+ : SIGN_MAG_LOW_ADDR_BITS (val, 8));
132+ else
133+ /* For pre-ARMv5TE (without ldrd), we use ldm/stm(db/da/ib)
134+ to access doublewords. The supported load/store offsets are
135+ -8, -4, and 4, which we try to produce here. */
136+ low = ((val & 0xf) ^ 0x8) - 0x8;
137+ }
138+ else if (GET_MODE_SIZE (mode) < 8)
139+ {
140+ /* NEON element load/stores do not have an offset. */
141+ if (TARGET_NEON_FP16 && mode == HFmode)
142+ return false;
143+
144+ if (TARGET_THUMB2)
145+ {
146+ /* Thumb-2 has an asymmetrical index range of (-256,4096).
147+ Try the wider 12-bit range first, and re-try if the result
148+ is out of range. */
149+ low = SIGN_MAG_LOW_ADDR_BITS (val, 12);
150+ if (low < -255)
151+ low = SIGN_MAG_LOW_ADDR_BITS (val, 8);
152+ }
153+ else
154+ {
155+ if (mode == HImode || mode == HFmode)
156+ {
157+ if (arm_arch4)
158+ low = SIGN_MAG_LOW_ADDR_BITS (val, 8);
159+ else
160+ {
161+ /* The storehi/movhi_bytes fallbacks can use only
162+ [-4094,+4094] of the full ldrb/strb index range. */
163+ low = SIGN_MAG_LOW_ADDR_BITS (val, 12);
164+ if (low == 4095 || low == -4095)
165+ return false;
166+ }
167+ }
168+ else
169+ low = SIGN_MAG_LOW_ADDR_BITS (val, 12);
170+ }
171+ }
172 else
173 return false;
174
175
176=== modified file 'gcc/config/arm/arm.md'
177--- old/gcc/config/arm/arm.md 2011-06-14 14:37:30 +0000
178+++ new/gcc/config/arm/arm.md 2011-06-27 22:14:07 +0000
179@@ -6267,7 +6267,7 @@
180 [(match_operand:DF 0 "arm_reload_memory_operand" "=o")
181 (match_operand:DF 1 "s_register_operand" "r")
182 (match_operand:SI 2 "s_register_operand" "=&r")]
183- "TARGET_32BIT"
184+ "TARGET_THUMB2"
185 "
186 {
187 enum rtx_code code = GET_CODE (XEXP (operands[0], 0));
188
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106762.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106762.patch
new file mode 100644
index 000000000..bbf9819ec
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106762.patch
@@ -0,0 +1,1355 @@
12011-06-28 Ira Rosen <ira.rosen@linaro.org>
2
3 Backport from FSF:
4
5 2011-06-07 Ira Rosen <ira.rosen@linaro.org>
6
7 gcc/
8 * tree-vectorizer.h (vect_recog_func_ptr): Make last argument to be
9 a pointer.
10 * tree-vect-patterns.c (vect_recog_widen_sum_pattern,
11 vect_recog_widen_mult_pattern, vect_recog_dot_prod_pattern,
12 vect_recog_pow_pattern): Likewise.
13 (vect_pattern_recog_1): Remove declaration.
14 (widened_name_p): Remove declaration. Add new argument to specify
15 whether to check that both types are either signed or unsigned.
16 (vect_recog_widen_mult_pattern): Update documentation. Handle
17 unsigned patterns and multiplication by constants.
18 (vect_pattern_recog_1): Update vect_recog_func references. Use
19 statement information from the statement returned from pattern
20 detection functions.
21 (vect_pattern_recog): Update vect_recog_func reference.
22 * tree-vect-stmts.c (vectorizable_type_promotion): For widening
23 multiplication by a constant use the type of the other operand.
24
25 gcc/testsuite
26 * lib/target-supports.exp
27 (check_effective_target_vect_widen_mult_qi_to_hi):
28 Add NEON as supporting target.
29 (check_effective_target_vect_widen_mult_hi_to_si): Likewise.
30 (check_effective_target_vect_widen_mult_qi_to_hi_pattern): New.
31 (check_effective_target_vect_widen_mult_hi_to_si_pattern): New.
32 * gcc.dg/vect/vect-widen-mult-u8.c: Expect to be vectorized
33 using widening multiplication on targets that support it.
34 * gcc.dg/vect/vect-widen-mult-u16.c: Likewise.
35 * gcc.dg/vect/vect-widen-mult-const-s16.c: New test.
36 * gcc.dg/vect/vect-widen-mult-const-u16.c: New test.
37
38 and
39
40 2011-06-15 Ira Rosen <ira.rosen@linaro.org>
41
42 gcc/
43 * tree-vect-loop-manip.c (remove_dead_stmts_from_loop): Remove.
44 (slpeel_tree_peel_loop_to_edge): Don't call
45 remove_dead_stmts_from_loop.
46 * tree-vect-loop.c (vect_determine_vectorization_factor): Don't
47 remove irrelevant pattern statements. For irrelevant statements
48 check if it is the last statement of a detected pattern, use
49 corresponding pattern statement instead.
50 (destroy_loop_vec_info): No need to remove pattern statements,
51 only free stmt_vec_info.
52 (vect_transform_loop): For irrelevant statements check if it is
53 the last statement of a detected pattern, use corresponding
54 pattern statement instead.
55 * tree-vect-patterns.c (vect_pattern_recog_1): Don't insert
56 pattern statements. Set basic block for the new statement.
57 (vect_pattern_recog): Update documentation.
58 * tree-vect-stmts.c (vect_mark_stmts_to_be_vectorized): Scan
59 operands of pattern statements.
60 (vectorizable_call): Fix printing. In case of a pattern statement
61 use the lhs of the original statement when creating a dummy
62 statement to replace the original call.
63 (vect_analyze_stmt): For irrelevant statements check if it is
64 the last statement of a detected pattern, use corresponding
65 pattern statement instead.
66 * tree-vect-slp.c (vect_schedule_slp_instance): For pattern
67 statements use gsi of the original statement.
68
69 and
70 2011-06-21 Ira Rosen <ira.rosen@linaro.org>
71
72 PR tree-optimization/49478
73 gcc/
74
75 * tree-vect-loop.c (vectorizable_reduction): Handle DOT_PROD_EXPR
76 with constant operand.
77
78=== added file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-s16.c'
79--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-s16.c 1970-01-01 00:00:00 +0000
80+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-s16.c 2011-06-19 10:59:13 +0000
81@@ -0,0 +1,60 @@
82+/* { dg-require-effective-target vect_int } */
83+
84+#include "tree-vect.h"
85+#include <stdlib.h>
86+
87+#define N 32
88+
89+__attribute__ ((noinline)) void
90+foo (int *__restrict a,
91+ short *__restrict b,
92+ int n)
93+{
94+ int i;
95+
96+ for (i = 0; i < n; i++)
97+ a[i] = b[i] * 2333;
98+
99+ for (i = 0; i < n; i++)
100+ if (a[i] != b[i] * 2333)
101+ abort ();
102+}
103+
104+__attribute__ ((noinline)) void
105+bar (int *__restrict a,
106+ short *__restrict b,
107+ int n)
108+{
109+ int i;
110+
111+ for (i = 0; i < n; i++)
112+ a[i] = b[i] * (short) 2333;
113+
114+ for (i = 0; i < n; i++)
115+ if (a[i] != b[i] * (short) 2333)
116+ abort ();
117+}
118+
119+int main (void)
120+{
121+ int i;
122+ int a[N];
123+ short b[N];
124+
125+ for (i = 0; i < N; i++)
126+ {
127+ a[i] = 0;
128+ b[i] = i;
129+ __asm__ volatile ("");
130+ }
131+
132+ foo (a, b, N);
133+ bar (a, b, N);
134+ return 0;
135+}
136+
137+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" { target vect_widen_mult_hi_to_si } } } */
138+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
139+/* { dg-final { scan-tree-dump-times "pattern recognized" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
140+/* { dg-final { cleanup-tree-dump "vect" } } */
141+
142
143=== added file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-u16.c'
144--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-u16.c 1970-01-01 00:00:00 +0000
145+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-const-u16.c 2011-06-19 10:59:13 +0000
146@@ -0,0 +1,77 @@
147+/* { dg-require-effective-target vect_int } */
148+
149+#include "tree-vect.h"
150+#include <stdlib.h>
151+
152+#define N 32
153+
154+__attribute__ ((noinline)) void
155+foo (unsigned int *__restrict a,
156+ unsigned short *__restrict b,
157+ int n)
158+{
159+ int i;
160+
161+ for (i = 0; i < n; i++)
162+ a[i] = b[i] * 2333;
163+
164+ for (i = 0; i < n; i++)
165+ if (a[i] != b[i] * 2333)
166+ abort ();
167+}
168+
169+__attribute__ ((noinline)) void
170+bar (unsigned int *__restrict a,
171+ unsigned short *__restrict b,
172+ int n)
173+{
174+ int i;
175+
176+ for (i = 0; i < n; i++)
177+ a[i] = (unsigned short) 2333 * b[i];
178+
179+ for (i = 0; i < n; i++)
180+ if (a[i] != b[i] * (unsigned short) 2333)
181+ abort ();
182+}
183+
184+__attribute__ ((noinline)) void
185+baz (unsigned int *__restrict a,
186+ unsigned short *__restrict b,
187+ int n)
188+{
189+ int i;
190+
191+ for (i = 0; i < n; i++)
192+ a[i] = b[i] * 233333333;
193+
194+ for (i = 0; i < n; i++)
195+ if (a[i] != b[i] * 233333333)
196+ abort ();
197+}
198+
199+
200+int main (void)
201+{
202+ int i;
203+ unsigned int a[N];
204+ unsigned short b[N];
205+
206+ for (i = 0; i < N; i++)
207+ {
208+ a[i] = 0;
209+ b[i] = i;
210+ __asm__ volatile ("");
211+ }
212+
213+ foo (a, b, N);
214+ bar (a, b, N);
215+ baz (a, b, N);
216+ return 0;
217+}
218+
219+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 3 "vect" { target vect_widen_mult_hi_to_si } } } */
220+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
221+/* { dg-final { scan-tree-dump-times "pattern recognized" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
222+/* { dg-final { cleanup-tree-dump "vect" } } */
223+
224
225=== modified file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-u16.c'
226--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-u16.c 2010-05-27 12:23:45 +0000
227+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-u16.c 2011-06-19 10:59:13 +0000
228@@ -9,13 +9,11 @@
229 unsigned short Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
230 unsigned int result[N];
231
232-/* short->int widening-mult */
233+/* unsigned short->unsigned int widening-mult. */
234 __attribute__ ((noinline)) int
235 foo1(int len) {
236 int i;
237
238- /* Not vectorized because X[i] and Y[i] are casted to 'int'
239- so the widening multiplication pattern is not recognized. */
240 for (i=0; i<len; i++) {
241 result[i] = (unsigned int)(X[i] * Y[i]);
242 }
243@@ -43,8 +41,8 @@
244 return 0;
245 }
246
247-/*The induction loop is vectorized */
248-/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" { xfail *-*-* } } } */
249-/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_pack_trunc } } } */
250+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { vect_widen_mult_hi_to_si || vect_unpack } } } } */
251+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 1 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
252+/* { dg-final { scan-tree-dump-times "pattern recognized" 1 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
253 /* { dg-final { cleanup-tree-dump "vect" } } */
254
255
256=== modified file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8.c'
257--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8.c 2009-05-08 12:39:01 +0000
258+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8.c 2011-06-19 10:59:13 +0000
259@@ -9,7 +9,7 @@
260 unsigned char Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
261 unsigned short result[N];
262
263-/* char->short widening-mult */
264+/* unsigned char-> unsigned short widening-mult. */
265 __attribute__ ((noinline)) int
266 foo1(int len) {
267 int i;
268@@ -28,8 +28,7 @@
269 for (i=0; i<N; i++) {
270 X[i] = i;
271 Y[i] = 64-i;
272- if (i%4 == 0)
273- X[i] = 5;
274+ __asm__ volatile ("");
275 }
276
277 foo1 (N);
278@@ -43,5 +42,7 @@
279 }
280
281 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { vect_widen_mult_qi_to_hi || vect_unpack } } } } */
282+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 1 "vect" { target vect_widen_mult_qi_to_hi_pattern } } } */
283+/* { dg-final { scan-tree-dump-times "pattern recognized" 1 "vect" { target vect_widen_mult_qi_to_hi_pattern } } } */
284 /* { dg-final { cleanup-tree-dump "vect" } } */
285
286
287=== modified file 'gcc/testsuite/lib/target-supports.exp'
288--- old/gcc/testsuite/lib/target-supports.exp 2011-06-02 12:12:00 +0000
289+++ new/gcc/testsuite/lib/target-supports.exp 2011-06-19 10:59:13 +0000
290@@ -2663,7 +2663,8 @@
291 } else {
292 set et_vect_widen_mult_qi_to_hi_saved 0
293 }
294- if { [istarget powerpc*-*-*] } {
295+ if { [istarget powerpc*-*-*]
296+ || ([istarget arm*-*-*] && [check_effective_target_arm_neon]) } {
297 set et_vect_widen_mult_qi_to_hi_saved 1
298 }
299 }
300@@ -2696,7 +2697,8 @@
301 || [istarget spu-*-*]
302 || [istarget ia64-*-*]
303 || [istarget i?86-*-*]
304- || [istarget x86_64-*-*] } {
305+ || [istarget x86_64-*-*]
306+ || ([istarget arm*-*-*] && [check_effective_target_arm_neon]) } {
307 set et_vect_widen_mult_hi_to_si_saved 1
308 }
309 }
310@@ -2705,6 +2707,52 @@
311 }
312
313 # Return 1 if the target plus current options supports a vector
314+# widening multiplication of *char* args into *short* result, 0 otherwise.
315+#
316+# This won't change for different subtargets so cache the result.
317+
318+proc check_effective_target_vect_widen_mult_qi_to_hi_pattern { } {
319+ global et_vect_widen_mult_qi_to_hi_pattern
320+
321+ if [info exists et_vect_widen_mult_qi_to_hi_pattern_saved] {
322+ verbose "check_effective_target_vect_widen_mult_qi_to_hi_pattern: using cached result" 2
323+ } else {
324+ set et_vect_widen_mult_qi_to_hi_pattern_saved 0
325+ if { [istarget powerpc*-*-*]
326+ || ([istarget arm*-*-*] && [check_effective_target_arm_neon]) } {
327+ set et_vect_widen_mult_qi_to_hi_pattern_saved 1
328+ }
329+ }
330+ verbose "check_effective_target_vect_widen_mult_qi_to_hi_pattern: returning $et_vect_widen_mult_qi_to_hi_pattern_saved" 2
331+ return $et_vect_widen_mult_qi_to_hi_pattern_saved
332+}
333+
334+# Return 1 if the target plus current options supports a vector
335+# widening multiplication of *short* args into *int* result, 0 otherwise.
336+#
337+# This won't change for different subtargets so cache the result.
338+
339+proc check_effective_target_vect_widen_mult_hi_to_si_pattern { } {
340+ global et_vect_widen_mult_hi_to_si_pattern
341+
342+ if [info exists et_vect_widen_mult_hi_to_si_pattern_saved] {
343+ verbose "check_effective_target_vect_widen_mult_hi_to_si_pattern: using cached result" 2
344+ } else {
345+ set et_vect_widen_mult_hi_to_si_pattern_saved 0
346+ if { [istarget powerpc*-*-*]
347+ || [istarget spu-*-*]
348+ || [istarget ia64-*-*]
349+ || [istarget i?86-*-*]
350+ || [istarget x86_64-*-*]
351+ || ([istarget arm*-*-*] && [check_effective_target_arm_neon]) } {
352+ set et_vect_widen_mult_hi_to_si_pattern_saved 1
353+ }
354+ }
355+ verbose "check_effective_target_vect_widen_mult_hi_to_si_pattern: returning $et_vect_widen_mult_hi_to_si_pattern_saved" 2
356+ return $et_vect_widen_mult_hi_to_si_pattern_saved
357+}
358+
359+# Return 1 if the target plus current options supports a vector
360 # dot-product of signed chars, 0 otherwise.
361 #
362 # This won't change for different subtargets so cache the result.
363
364=== modified file 'gcc/tree-vect-loop-manip.c'
365--- old/gcc/tree-vect-loop-manip.c 2011-05-18 13:24:05 +0000
366+++ new/gcc/tree-vect-loop-manip.c 2011-06-19 10:59:13 +0000
367@@ -1105,35 +1105,6 @@
368 first_niters = PHI_RESULT (newphi);
369 }
370
371-
372-/* Remove dead assignments from loop NEW_LOOP. */
373-
374-static void
375-remove_dead_stmts_from_loop (struct loop *new_loop)
376-{
377- basic_block *bbs = get_loop_body (new_loop);
378- unsigned i;
379- for (i = 0; i < new_loop->num_nodes; ++i)
380- {
381- gimple_stmt_iterator gsi;
382- for (gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi);)
383- {
384- gimple stmt = gsi_stmt (gsi);
385- if (is_gimple_assign (stmt)
386- && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME
387- && has_zero_uses (gimple_assign_lhs (stmt)))
388- {
389- gsi_remove (&gsi, true);
390- release_defs (stmt);
391- }
392- else
393- gsi_next (&gsi);
394- }
395- }
396- free (bbs);
397-}
398-
399-
400 /* Function slpeel_tree_peel_loop_to_edge.
401
402 Peel the first (last) iterations of LOOP into a new prolog (epilog) loop
403@@ -1445,13 +1416,6 @@
404 BITMAP_FREE (definitions);
405 delete_update_ssa ();
406
407- /* Remove all pattern statements from the loop copy. They will confuse
408- the expander if DCE is disabled.
409- ??? The pattern recognizer should be split into an analysis and
410- a transformation phase that is then run only on the loop that is
411- going to be transformed. */
412- remove_dead_stmts_from_loop (new_loop);
413-
414 adjust_vec_debug_stmts ();
415
416 return new_loop;
417
418=== modified file 'gcc/tree-vect-loop.c'
419--- old/gcc/tree-vect-loop.c 2011-03-01 13:18:25 +0000
420+++ new/gcc/tree-vect-loop.c 2011-06-22 06:21:13 +0000
421@@ -244,7 +244,7 @@
422 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
423 {
424 tree vf_vectype;
425- gimple stmt = gsi_stmt (si);
426+ gimple stmt = gsi_stmt (si), pattern_stmt;
427 stmt_info = vinfo_for_stmt (stmt);
428
429 if (vect_print_dump_info (REPORT_DETAILS))
430@@ -259,9 +259,25 @@
431 if (!STMT_VINFO_RELEVANT_P (stmt_info)
432 && !STMT_VINFO_LIVE_P (stmt_info))
433 {
434- if (vect_print_dump_info (REPORT_DETAILS))
435- fprintf (vect_dump, "skip.");
436- continue;
437+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
438+ && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
439+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
440+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
441+ {
442+ stmt = pattern_stmt;
443+ stmt_info = vinfo_for_stmt (pattern_stmt);
444+ if (vect_print_dump_info (REPORT_DETAILS))
445+ {
446+ fprintf (vect_dump, "==> examining pattern statement: ");
447+ print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
448+ }
449+ }
450+ else
451+ {
452+ if (vect_print_dump_info (REPORT_DETAILS))
453+ fprintf (vect_dump, "skip.");
454+ continue;
455+ }
456 }
457
458 if (gimple_get_lhs (stmt) == NULL_TREE)
459@@ -816,25 +832,17 @@
460
461 if (stmt_info)
462 {
463- /* Check if this is a "pattern stmt" (introduced by the
464- vectorizer during the pattern recognition pass). */
465- bool remove_stmt_p = false;
466- gimple orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
467- if (orig_stmt)
468- {
469- stmt_vec_info orig_stmt_info = vinfo_for_stmt (orig_stmt);
470- if (orig_stmt_info
471- && STMT_VINFO_IN_PATTERN_P (orig_stmt_info))
472- remove_stmt_p = true;
473- }
474+ /* Check if this statement has a related "pattern stmt"
475+ (introduced by the vectorizer during the pattern recognition
476+ pass). Free pattern's stmt_vec_info. */
477+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
478+ && vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info)))
479+ free_stmt_vec_info (STMT_VINFO_RELATED_STMT (stmt_info));
480
481 /* Free stmt_vec_info. */
482 free_stmt_vec_info (stmt);
483+ }
484
485- /* Remove dead "pattern stmts". */
486- if (remove_stmt_p)
487- gsi_remove (&si, true);
488- }
489 gsi_next (&si);
490 }
491 }
492@@ -4262,6 +4270,25 @@
493 return false;
494 }
495
496+ /* In case of widenning multiplication by a constant, we update the type
497+ of the constant to be the type of the other operand. We check that the
498+ constant fits the type in the pattern recognition pass. */
499+ if (code == DOT_PROD_EXPR
500+ && !types_compatible_p (TREE_TYPE (ops[0]), TREE_TYPE (ops[1])))
501+ {
502+ if (TREE_CODE (ops[0]) == INTEGER_CST)
503+ ops[0] = fold_convert (TREE_TYPE (ops[1]), ops[0]);
504+ else if (TREE_CODE (ops[1]) == INTEGER_CST)
505+ ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
506+ else
507+ {
508+ if (vect_print_dump_info (REPORT_DETAILS))
509+ fprintf (vect_dump, "invalid types in dot-prod");
510+
511+ return false;
512+ }
513+ }
514+
515 if (!vec_stmt) /* transformation not required. */
516 {
517 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
518@@ -4796,7 +4823,7 @@
519
520 for (si = gsi_start_bb (bb); !gsi_end_p (si);)
521 {
522- gimple stmt = gsi_stmt (si);
523+ gimple stmt = gsi_stmt (si), pattern_stmt;
524 bool is_store;
525
526 if (vect_print_dump_info (REPORT_DETAILS))
527@@ -4821,14 +4848,25 @@
528
529 if (!STMT_VINFO_RELEVANT_P (stmt_info)
530 && !STMT_VINFO_LIVE_P (stmt_info))
531- {
532- gsi_next (&si);
533- continue;
534+ {
535+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
536+ && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
537+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
538+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
539+ {
540+ stmt = pattern_stmt;
541+ stmt_info = vinfo_for_stmt (stmt);
542+ }
543+ else
544+ {
545+ gsi_next (&si);
546+ continue;
547+ }
548 }
549
550 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
551- nunits =
552- (unsigned int) TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
553+ nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (
554+ STMT_VINFO_VECTYPE (stmt_info));
555 if (!STMT_SLP_TYPE (stmt_info)
556 && nunits != (unsigned int) vectorization_factor
557 && vect_print_dump_info (REPORT_DETAILS))
558
559=== modified file 'gcc/tree-vect-patterns.c'
560--- old/gcc/tree-vect-patterns.c 2010-12-02 11:47:12 +0000
561+++ new/gcc/tree-vect-patterns.c 2011-06-22 12:10:44 +0000
562@@ -38,16 +38,11 @@
563 #include "recog.h"
564 #include "diagnostic-core.h"
565
566-/* Function prototypes */
567-static void vect_pattern_recog_1
568- (gimple (* ) (gimple, tree *, tree *), gimple_stmt_iterator);
569-static bool widened_name_p (tree, gimple, tree *, gimple *);
570-
571 /* Pattern recognition functions */
572-static gimple vect_recog_widen_sum_pattern (gimple, tree *, tree *);
573-static gimple vect_recog_widen_mult_pattern (gimple, tree *, tree *);
574-static gimple vect_recog_dot_prod_pattern (gimple, tree *, tree *);
575-static gimple vect_recog_pow_pattern (gimple, tree *, tree *);
576+static gimple vect_recog_widen_sum_pattern (gimple *, tree *, tree *);
577+static gimple vect_recog_widen_mult_pattern (gimple *, tree *, tree *);
578+static gimple vect_recog_dot_prod_pattern (gimple *, tree *, tree *);
579+static gimple vect_recog_pow_pattern (gimple *, tree *, tree *);
580 static vect_recog_func_ptr vect_vect_recog_func_ptrs[NUM_PATTERNS] = {
581 vect_recog_widen_mult_pattern,
582 vect_recog_widen_sum_pattern,
583@@ -61,10 +56,12 @@
584 is a result of a type-promotion, such that:
585 DEF_STMT: NAME = NOP (name0)
586 where the type of name0 (HALF_TYPE) is smaller than the type of NAME.
587-*/
588+ If CHECK_SIGN is TRUE, check that either both types are signed or both are
589+ unsigned. */
590
591 static bool
592-widened_name_p (tree name, gimple use_stmt, tree *half_type, gimple *def_stmt)
593+widened_name_p (tree name, gimple use_stmt, tree *half_type, gimple *def_stmt,
594+ bool check_sign)
595 {
596 tree dummy;
597 gimple dummy_gimple;
598@@ -98,7 +95,7 @@
599
600 *half_type = TREE_TYPE (oprnd0);
601 if (!INTEGRAL_TYPE_P (type) || !INTEGRAL_TYPE_P (*half_type)
602- || (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (*half_type))
603+ || ((TYPE_UNSIGNED (type) != TYPE_UNSIGNED (*half_type)) && check_sign)
604 || (TYPE_PRECISION (type) < (TYPE_PRECISION (*half_type) * 2)))
605 return false;
606
607@@ -168,12 +165,12 @@
608 inner-loop nested in an outer-loop that us being vectorized). */
609
610 static gimple
611-vect_recog_dot_prod_pattern (gimple last_stmt, tree *type_in, tree *type_out)
612+vect_recog_dot_prod_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
613 {
614 gimple stmt;
615 tree oprnd0, oprnd1;
616 tree oprnd00, oprnd01;
617- stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
618+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (*last_stmt);
619 tree type, half_type;
620 gimple pattern_stmt;
621 tree prod_type;
622@@ -181,10 +178,10 @@
623 struct loop *loop = LOOP_VINFO_LOOP (loop_info);
624 tree var, rhs;
625
626- if (!is_gimple_assign (last_stmt))
627+ if (!is_gimple_assign (*last_stmt))
628 return NULL;
629
630- type = gimple_expr_type (last_stmt);
631+ type = gimple_expr_type (*last_stmt);
632
633 /* Look for the following pattern
634 DX = (TYPE1) X;
635@@ -210,7 +207,7 @@
636 /* Starting from LAST_STMT, follow the defs of its uses in search
637 of the above pattern. */
638
639- if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
640+ if (gimple_assign_rhs_code (*last_stmt) != PLUS_EXPR)
641 return NULL;
642
643 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
644@@ -231,14 +228,14 @@
645
646 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
647 return NULL;
648- oprnd0 = gimple_assign_rhs1 (last_stmt);
649- oprnd1 = gimple_assign_rhs2 (last_stmt);
650+ oprnd0 = gimple_assign_rhs1 (*last_stmt);
651+ oprnd1 = gimple_assign_rhs2 (*last_stmt);
652 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
653 || !types_compatible_p (TREE_TYPE (oprnd1), type))
654 return NULL;
655- stmt = last_stmt;
656+ stmt = *last_stmt;
657
658- if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt))
659+ if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt, true))
660 {
661 stmt = def_stmt;
662 oprnd0 = gimple_assign_rhs1 (stmt);
663@@ -293,10 +290,10 @@
664 if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type)
665 || !types_compatible_p (TREE_TYPE (oprnd1), prod_type))
666 return NULL;
667- if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt))
668+ if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt, true))
669 return NULL;
670 oprnd00 = gimple_assign_rhs1 (def_stmt);
671- if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt))
672+ if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt, true))
673 return NULL;
674 oprnd01 = gimple_assign_rhs1 (def_stmt);
675 if (!types_compatible_p (half_type0, half_type1))
676@@ -322,7 +319,7 @@
677
678 /* We don't allow changing the order of the computation in the inner-loop
679 when doing outer-loop vectorization. */
680- gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
681+ gcc_assert (!nested_in_vect_loop_p (loop, *last_stmt));
682
683 return pattern_stmt;
684 }
685@@ -342,24 +339,47 @@
686
687 where type 'TYPE' is at least double the size of type 'type'.
688
689- Input:
690-
691- * LAST_STMT: A stmt from which the pattern search begins. In the example,
692- when this function is called with S5, the pattern {S3,S4,S5} is be detected.
693-
694- Output:
695-
696- * TYPE_IN: The type of the input arguments to the pattern.
697-
698- * TYPE_OUT: The type of the output of this pattern.
699-
700- * Return value: A new stmt that will be used to replace the sequence of
701- stmts that constitute the pattern. In this case it will be:
702- WIDEN_MULT <a_t, b_t>
703-*/
704+ Also detect unsgigned cases:
705+
706+ unsigned type a_t, b_t;
707+ unsigned TYPE u_prod_T;
708+ TYPE a_T, b_T, prod_T;
709+
710+ S1 a_t = ;
711+ S2 b_t = ;
712+ S3 a_T = (TYPE) a_t;
713+ S4 b_T = (TYPE) b_t;
714+ S5 prod_T = a_T * b_T;
715+ S6 u_prod_T = (unsigned TYPE) prod_T;
716+
717+ and multiplication by constants:
718+
719+ type a_t;
720+ TYPE a_T, prod_T;
721+
722+ S1 a_t = ;
723+ S3 a_T = (TYPE) a_t;
724+ S5 prod_T = a_T * CONST;
725+
726+ Input:
727+
728+ * LAST_STMT: A stmt from which the pattern search begins. In the example,
729+ when this function is called with S5, the pattern {S3,S4,S5,(S6)} is
730+ detected.
731+
732+ Output:
733+
734+ * TYPE_IN: The type of the input arguments to the pattern.
735+
736+ * TYPE_OUT: The type of the output of this pattern.
737+
738+ * Return value: A new stmt that will be used to replace the sequence of
739+ stmts that constitute the pattern. In this case it will be:
740+ WIDEN_MULT <a_t, b_t>
741+ */
742
743 static gimple
744-vect_recog_widen_mult_pattern (gimple last_stmt,
745+vect_recog_widen_mult_pattern (gimple *last_stmt,
746 tree *type_in,
747 tree *type_out)
748 {
749@@ -367,39 +387,112 @@
750 tree oprnd0, oprnd1;
751 tree type, half_type0, half_type1;
752 gimple pattern_stmt;
753- tree vectype, vectype_out;
754+ tree vectype, vectype_out = NULL_TREE;
755 tree dummy;
756 tree var;
757 enum tree_code dummy_code;
758 int dummy_int;
759 VEC (tree, heap) *dummy_vec;
760+ bool op0_ok, op1_ok;
761
762- if (!is_gimple_assign (last_stmt))
763+ if (!is_gimple_assign (*last_stmt))
764 return NULL;
765
766- type = gimple_expr_type (last_stmt);
767+ type = gimple_expr_type (*last_stmt);
768
769 /* Starting from LAST_STMT, follow the defs of its uses in search
770 of the above pattern. */
771
772- if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
773+ if (gimple_assign_rhs_code (*last_stmt) != MULT_EXPR)
774 return NULL;
775
776- oprnd0 = gimple_assign_rhs1 (last_stmt);
777- oprnd1 = gimple_assign_rhs2 (last_stmt);
778+ oprnd0 = gimple_assign_rhs1 (*last_stmt);
779+ oprnd1 = gimple_assign_rhs2 (*last_stmt);
780 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
781 || !types_compatible_p (TREE_TYPE (oprnd1), type))
782 return NULL;
783
784- /* Check argument 0 */
785- if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0))
786- return NULL;
787- oprnd0 = gimple_assign_rhs1 (def_stmt0);
788-
789- /* Check argument 1 */
790- if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1))
791- return NULL;
792- oprnd1 = gimple_assign_rhs1 (def_stmt1);
793+ /* Check argument 0. */
794+ op0_ok = widened_name_p (oprnd0, *last_stmt, &half_type0, &def_stmt0, false);
795+ /* Check argument 1. */
796+ op1_ok = widened_name_p (oprnd1, *last_stmt, &half_type1, &def_stmt1, false);
797+
798+ /* In case of multiplication by a constant one of the operands may not match
799+ the pattern, but not both. */
800+ if (!op0_ok && !op1_ok)
801+ return NULL;
802+
803+ if (op0_ok && op1_ok)
804+ {
805+ oprnd0 = gimple_assign_rhs1 (def_stmt0);
806+ oprnd1 = gimple_assign_rhs1 (def_stmt1);
807+ }
808+ else if (!op0_ok)
809+ {
810+ if (CONSTANT_CLASS_P (oprnd0)
811+ && TREE_CODE (half_type1) == INTEGER_TYPE
812+ && tree_int_cst_lt (oprnd0, TYPE_MAXVAL (half_type1))
813+ && tree_int_cst_lt (TYPE_MINVAL (half_type1), oprnd0))
814+ {
815+ /* OPRND0 is a constant of HALF_TYPE1. */
816+ half_type0 = half_type1;
817+ oprnd1 = gimple_assign_rhs1 (def_stmt1);
818+ }
819+ else
820+ return NULL;
821+ }
822+ else if (!op1_ok)
823+ {
824+ if (CONSTANT_CLASS_P (oprnd1)
825+ && TREE_CODE (half_type0) == INTEGER_TYPE
826+ && tree_int_cst_lt (oprnd1, TYPE_MAXVAL (half_type0))
827+ && tree_int_cst_lt (TYPE_MINVAL (half_type0), oprnd1))
828+ {
829+ /* OPRND1 is a constant of HALF_TYPE0. */
830+ half_type1 = half_type0;
831+ oprnd0 = gimple_assign_rhs1 (def_stmt0);
832+ }
833+ else
834+ return NULL;
835+ }
836+
837+ /* Handle unsigned case. Look for
838+ S6 u_prod_T = (unsigned TYPE) prod_T;
839+ Use unsigned TYPE as the type for WIDEN_MULT_EXPR. */
840+ if (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (half_type0))
841+ {
842+ tree lhs = gimple_assign_lhs (*last_stmt), use_lhs;
843+ imm_use_iterator imm_iter;
844+ use_operand_p use_p;
845+ int nuses = 0;
846+ gimple use_stmt = NULL;
847+ tree use_type;
848+
849+ if (TYPE_UNSIGNED (type) == TYPE_UNSIGNED (half_type1))
850+ return NULL;
851+
852+ FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs)
853+ {
854+ if (is_gimple_debug (USE_STMT (use_p)))
855+ continue;
856+ use_stmt = USE_STMT (use_p);
857+ nuses++;
858+ }
859+
860+ if (nuses != 1 || !is_gimple_assign (use_stmt)
861+ || gimple_assign_rhs_code (use_stmt) != NOP_EXPR)
862+ return NULL;
863+
864+ use_lhs = gimple_assign_lhs (use_stmt);
865+ use_type = TREE_TYPE (use_lhs);
866+ if (!INTEGRAL_TYPE_P (use_type)
867+ || (TYPE_UNSIGNED (type) == TYPE_UNSIGNED (use_type))
868+ || (TYPE_PRECISION (type) != TYPE_PRECISION (use_type)))
869+ return NULL;
870+
871+ type = use_type;
872+ *last_stmt = use_stmt;
873+ }
874
875 if (!types_compatible_p (half_type0, half_type1))
876 return NULL;
877@@ -413,7 +506,7 @@
878 vectype_out = get_vectype_for_scalar_type (type);
879 if (!vectype
880 || !vectype_out
881- || !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt,
882+ || !supportable_widening_operation (WIDEN_MULT_EXPR, *last_stmt,
883 vectype_out, vectype,
884 &dummy, &dummy, &dummy_code,
885 &dummy_code, &dummy_int, &dummy_vec))
886@@ -462,16 +555,16 @@
887 */
888
889 static gimple
890-vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out)
891+vect_recog_pow_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
892 {
893 tree fn, base, exp = NULL;
894 gimple stmt;
895 tree var;
896
897- if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
898+ if (!is_gimple_call (*last_stmt) || gimple_call_lhs (*last_stmt) == NULL)
899 return NULL;
900
901- fn = gimple_call_fndecl (last_stmt);
902+ fn = gimple_call_fndecl (*last_stmt);
903 if (fn == NULL_TREE || DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)
904 return NULL;
905
906@@ -481,8 +574,8 @@
907 case BUILT_IN_POWI:
908 case BUILT_IN_POWF:
909 case BUILT_IN_POW:
910- base = gimple_call_arg (last_stmt, 0);
911- exp = gimple_call_arg (last_stmt, 1);
912+ base = gimple_call_arg (*last_stmt, 0);
913+ exp = gimple_call_arg (*last_stmt, 1);
914 if (TREE_CODE (exp) != REAL_CST
915 && TREE_CODE (exp) != INTEGER_CST)
916 return NULL;
917@@ -574,21 +667,21 @@
918 inner-loop nested in an outer-loop that us being vectorized). */
919
920 static gimple
921-vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out)
922+vect_recog_widen_sum_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
923 {
924 gimple stmt;
925 tree oprnd0, oprnd1;
926- stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
927+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (*last_stmt);
928 tree type, half_type;
929 gimple pattern_stmt;
930 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
931 struct loop *loop = LOOP_VINFO_LOOP (loop_info);
932 tree var;
933
934- if (!is_gimple_assign (last_stmt))
935+ if (!is_gimple_assign (*last_stmt))
936 return NULL;
937
938- type = gimple_expr_type (last_stmt);
939+ type = gimple_expr_type (*last_stmt);
940
941 /* Look for the following pattern
942 DX = (TYPE) X;
943@@ -600,25 +693,25 @@
944 /* Starting from LAST_STMT, follow the defs of its uses in search
945 of the above pattern. */
946
947- if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
948+ if (gimple_assign_rhs_code (*last_stmt) != PLUS_EXPR)
949 return NULL;
950
951 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
952 return NULL;
953
954- oprnd0 = gimple_assign_rhs1 (last_stmt);
955- oprnd1 = gimple_assign_rhs2 (last_stmt);
956+ oprnd0 = gimple_assign_rhs1 (*last_stmt);
957+ oprnd1 = gimple_assign_rhs2 (*last_stmt);
958 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
959 || !types_compatible_p (TREE_TYPE (oprnd1), type))
960 return NULL;
961
962- /* So far so good. Since last_stmt was detected as a (summation) reduction,
963+ /* So far so good. Since *last_stmt was detected as a (summation) reduction,
964 we know that oprnd1 is the reduction variable (defined by a loop-header
965 phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
966 Left to check that oprnd0 is defined by a cast from type 'type' to type
967 'TYPE'. */
968
969- if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt))
970+ if (!widened_name_p (oprnd0, *last_stmt, &half_type, &stmt, true))
971 return NULL;
972
973 oprnd0 = gimple_assign_rhs1 (stmt);
974@@ -639,7 +732,7 @@
975
976 /* We don't allow changing the order of the computation in the inner-loop
977 when doing outer-loop vectorization. */
978- gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
979+ gcc_assert (!nested_in_vect_loop_p (loop, *last_stmt));
980
981 return pattern_stmt;
982 }
983@@ -669,23 +762,27 @@
984
985 static void
986 vect_pattern_recog_1 (
987- gimple (* vect_recog_func) (gimple, tree *, tree *),
988+ gimple (* vect_recog_func) (gimple *, tree *, tree *),
989 gimple_stmt_iterator si)
990 {
991 gimple stmt = gsi_stmt (si), pattern_stmt;
992- stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
993+ stmt_vec_info stmt_info;
994 stmt_vec_info pattern_stmt_info;
995- loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
996+ loop_vec_info loop_vinfo;
997 tree pattern_vectype;
998 tree type_in, type_out;
999 enum tree_code code;
1000 int i;
1001 gimple next;
1002
1003- pattern_stmt = (* vect_recog_func) (stmt, &type_in, &type_out);
1004+ pattern_stmt = (* vect_recog_func) (&stmt, &type_in, &type_out);
1005 if (!pattern_stmt)
1006 return;
1007
1008+ si = gsi_for_stmt (stmt);
1009+ stmt_info = vinfo_for_stmt (stmt);
1010+ loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1011+
1012 if (VECTOR_MODE_P (TYPE_MODE (type_in)))
1013 {
1014 /* No need to check target support (already checked by the pattern
1015@@ -736,9 +833,9 @@
1016 }
1017
1018 /* Mark the stmts that are involved in the pattern. */
1019- gsi_insert_before (&si, pattern_stmt, GSI_SAME_STMT);
1020 set_vinfo_for_stmt (pattern_stmt,
1021 new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL));
1022+ gimple_set_bb (pattern_stmt, gimple_bb (stmt));
1023 pattern_stmt_info = vinfo_for_stmt (pattern_stmt);
1024
1025 STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt;
1026@@ -761,8 +858,8 @@
1027 LOOP_VINFO - a struct_loop_info of a loop in which we want to look for
1028 computation idioms.
1029
1030- Output - for each computation idiom that is detected we insert a new stmt
1031- that provides the same functionality and that can be vectorized. We
1032+ Output - for each computation idiom that is detected we create a new stmt
1033+ that provides the same functionality and that can be vectorized. We
1034 also record some information in the struct_stmt_info of the relevant
1035 stmts, as explained below:
1036
1037@@ -777,52 +874,48 @@
1038 S5: ... = ..use(a_0).. - - -
1039
1040 Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be
1041- represented by a single stmt. We then:
1042- - create a new stmt S6 that will replace the pattern.
1043- - insert the new stmt S6 before the last stmt in the pattern
1044+ represented by a single stmt. We then:
1045+ - create a new stmt S6 equivalent to the pattern (the stmt is not
1046+ inserted into the code)
1047 - fill in the STMT_VINFO fields as follows:
1048
1049 in_pattern_p related_stmt vec_stmt
1050 S1: a_i = .... - - -
1051 S2: a_2 = ..use(a_i).. - - -
1052 S3: a_1 = ..use(a_2).. - - -
1053- > S6: a_new = .... - S4 -
1054 S4: a_0 = ..use(a_1).. true S6 -
1055+ '---> S6: a_new = .... - S4 -
1056 S5: ... = ..use(a_0).. - - -
1057
1058 (the last stmt in the pattern (S4) and the new pattern stmt (S6) point
1059- to each other through the RELATED_STMT field).
1060+ to each other through the RELATED_STMT field).
1061
1062 S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead
1063 of S4 because it will replace all its uses. Stmts {S1,S2,S3} will
1064 remain irrelevant unless used by stmts other than S4.
1065
1066 If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
1067- (because they are marked as irrelevant). It will vectorize S6, and record
1068+ (because they are marked as irrelevant). It will vectorize S6, and record
1069 a pointer to the new vector stmt VS6 both from S6 (as usual), and also
1070- from S4. We do that so that when we get to vectorizing stmts that use the
1071+ from S4. We do that so that when we get to vectorizing stmts that use the
1072 def of S4 (like S5 that uses a_0), we'll know where to take the relevant
1073- vector-def from. S4 will be skipped, and S5 will be vectorized as usual:
1074+ vector-def from. S4 will be skipped, and S5 will be vectorized as usual:
1075
1076 in_pattern_p related_stmt vec_stmt
1077 S1: a_i = .... - - -
1078 S2: a_2 = ..use(a_i).. - - -
1079 S3: a_1 = ..use(a_2).. - - -
1080 > VS6: va_new = .... - - -
1081- S6: a_new = .... - S4 VS6
1082 S4: a_0 = ..use(a_1).. true S6 VS6
1083+ '---> S6: a_new = .... - S4 VS6
1084 > VS5: ... = ..vuse(va_new).. - - -
1085 S5: ... = ..use(a_0).. - - -
1086
1087- DCE could then get rid of {S1,S2,S3,S4,S5,S6} (if their defs are not used
1088+ DCE could then get rid of {S1,S2,S3,S4,S5} (if their defs are not used
1089 elsewhere), and we'll end up with:
1090
1091 VS6: va_new = ....
1092- VS5: ... = ..vuse(va_new)..
1093-
1094- If vectorization does not succeed, DCE will clean S6 away (its def is
1095- not used), and we'll end up with the original sequence.
1096-*/
1097+ VS5: ... = ..vuse(va_new).. */
1098
1099 void
1100 vect_pattern_recog (loop_vec_info loop_vinfo)
1101@@ -832,7 +925,7 @@
1102 unsigned int nbbs = loop->num_nodes;
1103 gimple_stmt_iterator si;
1104 unsigned int i, j;
1105- gimple (* vect_recog_func_ptr) (gimple, tree *, tree *);
1106+ gimple (* vect_recog_func_ptr) (gimple *, tree *, tree *);
1107
1108 if (vect_print_dump_info (REPORT_DETAILS))
1109 fprintf (vect_dump, "=== vect_pattern_recog ===");
1110
1111=== modified file 'gcc/tree-vect-slp.c'
1112--- old/gcc/tree-vect-slp.c 2011-05-05 15:43:06 +0000
1113+++ new/gcc/tree-vect-slp.c 2011-06-19 10:59:13 +0000
1114@@ -2510,6 +2510,8 @@
1115 && STMT_VINFO_STRIDED_ACCESS (stmt_info)
1116 && !REFERENCE_CLASS_P (gimple_get_lhs (stmt)))
1117 si = gsi_for_stmt (SLP_INSTANCE_FIRST_LOAD_STMT (instance));
1118+ else if (is_pattern_stmt_p (stmt_info))
1119+ si = gsi_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
1120 else
1121 si = gsi_for_stmt (stmt);
1122
1123
1124=== modified file 'gcc/tree-vect-stmts.c'
1125--- old/gcc/tree-vect-stmts.c 2011-06-02 12:12:00 +0000
1126+++ new/gcc/tree-vect-stmts.c 2011-06-22 06:21:13 +0000
1127@@ -605,15 +605,76 @@
1128 break;
1129 }
1130
1131- FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE)
1132- {
1133- tree op = USE_FROM_PTR (use_p);
1134- if (!process_use (stmt, op, loop_vinfo, live_p, relevant, &worklist))
1135- {
1136- VEC_free (gimple, heap, worklist);
1137- return false;
1138- }
1139- }
1140+ if (is_pattern_stmt_p (vinfo_for_stmt (stmt)))
1141+ {
1142+ /* Pattern statements are not inserted into the code, so
1143+ FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we
1144+ have to scan the RHS or function arguments instead. */
1145+ if (is_gimple_assign (stmt))
1146+ {
1147+ tree rhs = gimple_assign_rhs1 (stmt);
1148+ if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1149+ == GIMPLE_SINGLE_RHS)
1150+ {
1151+ unsigned int op_num = TREE_OPERAND_LENGTH (gimple_assign_rhs1
1152+ (stmt));
1153+ for (i = 0; i < op_num; i++)
1154+ {
1155+ tree op = TREE_OPERAND (rhs, i);
1156+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1157+ &worklist))
1158+ {
1159+ VEC_free (gimple, heap, worklist);
1160+ return false;
1161+ }
1162+ }
1163+ }
1164+ else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1165+ == GIMPLE_BINARY_RHS)
1166+ {
1167+ tree op = gimple_assign_rhs1 (stmt);
1168+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1169+ &worklist))
1170+ {
1171+ VEC_free (gimple, heap, worklist);
1172+ return false;
1173+ }
1174+ op = gimple_assign_rhs2 (stmt);
1175+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1176+ &worklist))
1177+ {
1178+ VEC_free (gimple, heap, worklist);
1179+ return false;
1180+ }
1181+ }
1182+ else
1183+ return false;
1184+ }
1185+ else if (is_gimple_call (stmt))
1186+ {
1187+ for (i = 0; i < gimple_call_num_args (stmt); i++)
1188+ {
1189+ tree arg = gimple_call_arg (stmt, i);
1190+ if (!process_use (stmt, arg, loop_vinfo, live_p, relevant,
1191+ &worklist))
1192+ {
1193+ VEC_free (gimple, heap, worklist);
1194+ return false;
1195+ }
1196+ }
1197+ }
1198+ }
1199+ else
1200+ FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE)
1201+ {
1202+ tree op = USE_FROM_PTR (use_p);
1203+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1204+ &worklist))
1205+ {
1206+ VEC_free (gimple, heap, worklist);
1207+ return false;
1208+ }
1209+ }
1210 } /* while worklist */
1211
1212 VEC_free (gimple, heap, worklist);
1213@@ -1405,6 +1466,7 @@
1214 VEC(tree, heap) *vargs = NULL;
1215 enum { NARROW, NONE, WIDEN } modifier;
1216 size_t i, nargs;
1217+ tree lhs;
1218
1219 /* FORNOW: unsupported in basic block SLP. */
1220 gcc_assert (loop_vinfo);
1221@@ -1542,7 +1604,7 @@
1222 /** Transform. **/
1223
1224 if (vect_print_dump_info (REPORT_DETAILS))
1225- fprintf (vect_dump, "transform operation.");
1226+ fprintf (vect_dump, "transform call.");
1227
1228 /* Handle def. */
1229 scalar_dest = gimple_call_lhs (stmt);
1230@@ -1661,8 +1723,11 @@
1231 rhs of the statement with something harmless. */
1232
1233 type = TREE_TYPE (scalar_dest);
1234- new_stmt = gimple_build_assign (gimple_call_lhs (stmt),
1235- build_zero_cst (type));
1236+ if (is_pattern_stmt_p (stmt_info))
1237+ lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info));
1238+ else
1239+ lhs = gimple_call_lhs (stmt);
1240+ new_stmt = gimple_build_assign (lhs, build_zero_cst (type));
1241 set_vinfo_for_stmt (new_stmt, stmt_info);
1242 set_vinfo_for_stmt (stmt, NULL);
1243 STMT_VINFO_STMT (stmt_info) = new_stmt;
1244@@ -3231,6 +3296,33 @@
1245 fprintf (vect_dump, "use not simple.");
1246 return false;
1247 }
1248+
1249+ op_type = TREE_CODE_LENGTH (code);
1250+ if (op_type == binary_op)
1251+ {
1252+ bool ok;
1253+
1254+ op1 = gimple_assign_rhs2 (stmt);
1255+ if (code == WIDEN_MULT_EXPR)
1256+ {
1257+ /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of
1258+ OP1. */
1259+ if (CONSTANT_CLASS_P (op0))
1260+ ok = vect_is_simple_use_1 (op1, loop_vinfo, NULL,
1261+ &def_stmt, &def, &dt[1], &vectype_in);
1262+ else
1263+ ok = vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def,
1264+ &dt[1]);
1265+
1266+ if (!ok)
1267+ {
1268+ if (vect_print_dump_info (REPORT_DETAILS))
1269+ fprintf (vect_dump, "use not simple.");
1270+ return false;
1271+ }
1272+ }
1273+ }
1274+
1275 /* If op0 is an external or constant def use a vector type with
1276 the same size as the output vector type. */
1277 if (!vectype_in)
1278@@ -3263,18 +3355,6 @@
1279
1280 gcc_assert (ncopies >= 1);
1281
1282- op_type = TREE_CODE_LENGTH (code);
1283- if (op_type == binary_op)
1284- {
1285- op1 = gimple_assign_rhs2 (stmt);
1286- if (!vect_is_simple_use (op1, loop_vinfo, NULL, &def_stmt, &def, &dt[1]))
1287- {
1288- if (vect_print_dump_info (REPORT_DETAILS))
1289- fprintf (vect_dump, "use not simple.");
1290- return false;
1291- }
1292- }
1293-
1294 /* Supportable by target? */
1295 if (!supportable_widening_operation (code, stmt, vectype_out, vectype_in,
1296 &decl1, &decl2, &code1, &code2,
1297@@ -3300,6 +3380,14 @@
1298 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
1299 ncopies);
1300
1301+ if (code == WIDEN_MULT_EXPR)
1302+ {
1303+ if (CONSTANT_CLASS_P (op0))
1304+ op0 = fold_convert (TREE_TYPE (op1), op0);
1305+ else if (CONSTANT_CLASS_P (op1))
1306+ op1 = fold_convert (TREE_TYPE (op0), op1);
1307+ }
1308+
1309 /* Handle def. */
1310 /* In case of multi-step promotion, we first generate promotion operations
1311 to the intermediate types, and then from that types to the final one.
1312@@ -4824,10 +4912,26 @@
1313 if (!STMT_VINFO_RELEVANT_P (stmt_info)
1314 && !STMT_VINFO_LIVE_P (stmt_info))
1315 {
1316- if (vect_print_dump_info (REPORT_DETAILS))
1317- fprintf (vect_dump, "irrelevant.");
1318+ gimple pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
1319+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
1320+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
1321+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
1322+ {
1323+ stmt = pattern_stmt;
1324+ stmt_info = vinfo_for_stmt (pattern_stmt);
1325+ if (vect_print_dump_info (REPORT_DETAILS))
1326+ {
1327+ fprintf (vect_dump, "==> examining pattern statement: ");
1328+ print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
1329+ }
1330+ }
1331+ else
1332+ {
1333+ if (vect_print_dump_info (REPORT_DETAILS))
1334+ fprintf (vect_dump, "irrelevant.");
1335
1336- return true;
1337+ return true;
1338+ }
1339 }
1340
1341 switch (STMT_VINFO_DEF_TYPE (stmt_info))
1342
1343=== modified file 'gcc/tree-vectorizer.h'
1344--- old/gcc/tree-vectorizer.h 2011-05-05 15:43:06 +0000
1345+++ new/gcc/tree-vectorizer.h 2011-06-19 10:59:13 +0000
1346@@ -884,7 +884,7 @@
1347 /* Pattern recognition functions.
1348 Additional pattern recognition functions can (and will) be added
1349 in the future. */
1350-typedef gimple (* vect_recog_func_ptr) (gimple, tree *, tree *);
1351+typedef gimple (* vect_recog_func_ptr) (gimple *, tree *, tree *);
1352 #define NUM_PATTERNS 4
1353 void vect_pattern_recog (loop_vec_info);
1354
1355
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106763.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106763.patch
new file mode 100644
index 000000000..8d2ce2176
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106763.patch
@@ -0,0 +1,96 @@
12011-06-28 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
2
3 Backport from mainline.
4 LP 791327
5 gcc/
6 2011-06-09 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
7
8 PR target/49335
9 * config/arm/predicates.md (add_operator): New.
10 * config/arm/arm.md ("*arith_shiftsi"): Fix for SP reg usage
11 in Thumb2.
12
132011-06-28 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
14
15 Backport from mainline.
16 gcc/
17 2011-06-24 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
18
19 PR target/49385
20 * config/arm/thumb2.md (*thumb2_movhi_insn): Make sure atleast
21 one of the operands is a register.
22
23=== modified file 'gcc/config/arm/arm.md'
24--- old/gcc/config/arm/arm.md 2011-06-27 22:14:07 +0000
25+++ new/gcc/config/arm/arm.md 2011-06-28 12:02:27 +0000
26@@ -8584,18 +8584,22 @@
27 ;; Patterns to allow combination of arithmetic, cond code and shifts
28
29 (define_insn "*arith_shiftsi"
30- [(set (match_operand:SI 0 "s_register_operand" "=r,r")
31+ [(set (match_operand:SI 0 "s_register_operand" "=r,r,r,r")
32 (match_operator:SI 1 "shiftable_operator"
33 [(match_operator:SI 3 "shift_operator"
34- [(match_operand:SI 4 "s_register_operand" "r,r")
35- (match_operand:SI 5 "shift_amount_operand" "M,r")])
36- (match_operand:SI 2 "s_register_operand" "rk,rk")]))]
37+ [(match_operand:SI 4 "s_register_operand" "r,r,r,r")
38+ (match_operand:SI 5 "shift_amount_operand" "M,M,M,r")])
39+ (match_operand:SI 2 "s_register_operand" "rk,rk,r,rk")]))]
40 "TARGET_32BIT"
41 "%i1%?\\t%0, %2, %4%S3"
42 [(set_attr "predicable" "yes")
43 (set_attr "shift" "4")
44- (set_attr "arch" "32,a")
45- ;; We have to make sure to disable the second alternative if
46+ (set_attr "arch" "a,t2,t2,a")
47+ ;; Thumb2 doesn't allow the stack pointer to be used for
48+ ;; operand1 for all operations other than add and sub. In this case
49+ ;; the minus operation is a candidate for an rsub and hence needs
50+ ;; to be disabled.
51+ ;; We have to make sure to disable the fourth alternative if
52 ;; the shift_operator is MULT, since otherwise the insn will
53 ;; also match a multiply_accumulate pattern and validate_change
54 ;; will allow a replacement of the constant with a register
55@@ -8603,9 +8607,13 @@
56 (set_attr_alternative "insn_enabled"
57 [(const_string "yes")
58 (if_then_else
59+ (match_operand:SI 1 "add_operator" "")
60+ (const_string "yes") (const_string "no"))
61+ (const_string "yes")
62+ (if_then_else
63 (match_operand:SI 3 "mult_operator" "")
64 (const_string "no") (const_string "yes"))])
65- (set_attr "type" "alu_shift,alu_shift_reg")])
66+ (set_attr "type" "alu_shift,alu_shift,alu_shift,alu_shift_reg")])
67
68 (define_split
69 [(set (match_operand:SI 0 "s_register_operand" "")
70
71=== modified file 'gcc/config/arm/predicates.md'
72--- old/gcc/config/arm/predicates.md 2011-05-03 15:14:56 +0000
73+++ new/gcc/config/arm/predicates.md 2011-06-22 15:50:23 +0000
74@@ -687,3 +687,6 @@
75 (define_special_predicate "neon_struct_operand"
76 (and (match_code "mem")
77 (match_test "TARGET_32BIT && neon_vector_mem_operand (op, 2)")))
78+
79+(define_special_predicate "add_operator"
80+ (match_code "plus"))
81
82=== modified file 'gcc/config/arm/thumb2.md'
83--- old/gcc/config/arm/thumb2.md 2011-06-14 14:37:30 +0000
84+++ new/gcc/config/arm/thumb2.md 2011-06-20 12:18:27 +0000
85@@ -207,7 +207,9 @@
86 (define_insn "*thumb2_movhi_insn"
87 [(set (match_operand:HI 0 "nonimmediate_operand" "=r,r,m,r")
88 (match_operand:HI 1 "general_operand" "rI,n,r,m"))]
89- "TARGET_THUMB2"
90+ "TARGET_THUMB2
91+ && (register_operand (operands[0], HImode)
92+ || register_operand (operands[1], HImode))"
93 "@
94 mov%?\\t%0, %1\\t%@ movhi
95 movw%?\\t%0, %L1\\t%@ movhi
96
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106764.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106764.patch
new file mode 100644
index 000000000..a548b1b68
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106764.patch
@@ -0,0 +1,25 @@
12011-06-30 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
2
3 Backport from mainline.
4 LP 744754
5 2011-04-17 Chung-Lin Tang <cltang@codesourcery.com>
6
7 * config/arm/arm.c (neon_struct_mem_operand):
8 Support POST_INC/PRE_DEC memory operands.
9
10=== modified file 'gcc/config/arm/arm.c'
11--- old/gcc/config/arm/arm.c 2011-06-27 22:14:07 +0000
12+++ new/gcc/config/arm/arm.c 2011-06-29 09:13:17 +0000
13@@ -9357,6 +9357,11 @@
14 if (GET_CODE (ind) == REG)
15 return arm_address_register_rtx_p (ind, 0);
16
17+ /* vldm/vstm allows POST_INC (ia) and PRE_DEC (db). */
18+ if (GET_CODE (ind) == POST_INC
19+ || GET_CODE (ind) == PRE_DEC)
20+ return arm_address_register_rtx_p (XEXP (ind, 0), 0);
21+
22 return FALSE;
23 }
24
25
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106766.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106766.patch
new file mode 100644
index 000000000..15046a766
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106766.patch
@@ -0,0 +1,25 @@
12011-07-03 Ira Rosen <ira.rosen@linaro.org>
2
3 Backport from FSF:
4 2011-06-12 Ira Rosen <ira.rosen@linaro.org>
5
6 gcc/
7 * tree-vect-data-refs.c (vect_peeling_hash_get_most_frequent):
8 Take number of iterations to peel into account for equally frequent
9 misalignment values.
10
11=== modified file 'gcc/tree-vect-data-refs.c'
12--- old/gcc/tree-vect-data-refs.c 2011-06-02 12:12:00 +0000
13+++ new/gcc/tree-vect-data-refs.c 2011-06-29 11:20:24 +0000
14@@ -1256,7 +1256,9 @@
15 vect_peel_info elem = (vect_peel_info) *slot;
16 vect_peel_extended_info max = (vect_peel_extended_info) data;
17
18- if (elem->count > max->peel_info.count)
19+ if (elem->count > max->peel_info.count
20+ || (elem->count == max->peel_info.count
21+ && max->peel_info.npeel > elem->npeel))
22 {
23 max->peel_info.npeel = elem->npeel;
24 max->peel_info.count = elem->count;
25
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106768.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106768.patch
new file mode 100644
index 000000000..f1f7718eb
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106768.patch
@@ -0,0 +1,182 @@
12011-07-07 Richard Sandiford <richard.sandiford@linaro.org>
2
3 gcc/
4 * builtins.c (get_object_alignment): Fix comment.
5 * fold-const.c (get_pointer_modulus_and_residue): Remove
6 allow_func_align. Use get_object_alignment.
7 (fold_binary_loc): Update caller.
8
92011-07-07 Richard Sandiford <richard.sandiford@linaro.org>
10
11 gcc/
12 Backport from mainline:
13
14 2011-06-29 Richard Sandiford <richard.sandiford@linaro.org>
15
16 PR tree-optimization/49545
17 * builtins.c (get_object_alignment_1): Update function comment.
18 Do not use DECL_ALIGN for functions, but test
19 TARGET_PTRMEMFUNC_VBIT_LOCATION instead.
20 * fold-const.c (get_pointer_modulus_and_residue): Don't check
21 for functions here.
22 * tree-ssa-ccp.c (get_value_from_alignment): Likewise.
23
24 gcc/testsuite/
25 Backport from mainline:
26
27 2011-06-29 Richard Sandiford <richard.sandiford@linaro.org>
28
29 * gcc.dg/torture/pr49169.c: Restrict to ARM and MIPS targets.
30
312011-07-07 Richard Sandiford <richard.sandiford@linaro.org>
32
33 gcc/
34 Backport from mainline:
35
36 2011-07-27 Richard Guenther <rguenther@suse.de>
37
38 PR tree-optimization/49169
39 * fold-const.c (get_pointer_modulus_and_residue): Don't rely on
40 the alignment of function decls.
41
42 gcc/testsuite/
43 Backport from mainline:
44
45 2011-07-27 Michael Hope <michael.hope@linaro.org>
46 Richard Sandiford <richard.sandiford@linaro.org>
47
48 PR tree-optimization/49169
49 * gcc.dg/torture/pr49169.c: New test.
50
51=== modified file 'gcc/builtins.c'
52--- old/gcc/builtins.c 2011-03-03 21:56:58 +0000
53+++ new/gcc/builtins.c 2011-07-04 09:52:27 +0000
54@@ -264,7 +264,14 @@
55 }
56
57 /* Return the alignment in bits of EXP, an object.
58- Don't return more than MAX_ALIGN no matter what. */
59+ Don't return more than MAX_ALIGN no matter what.
60+
61+ Note that the address (and thus the alignment) computed here is based
62+ on the address to which a symbol resolves, whereas DECL_ALIGN is based
63+ on the address at which an object is actually located. These two
64+ addresses are not always the same. For example, on ARM targets,
65+ the address &foo of a Thumb function foo() has the lowest bit set,
66+ whereas foo() itself starts on an even address. */
67
68 unsigned int
69 get_object_alignment (tree exp, unsigned int max_align)
70@@ -286,7 +293,21 @@
71 exp = DECL_INITIAL (exp);
72 if (DECL_P (exp)
73 && TREE_CODE (exp) != LABEL_DECL)
74- align = DECL_ALIGN (exp);
75+ {
76+ if (TREE_CODE (exp) == FUNCTION_DECL)
77+ {
78+ /* Function addresses can encode extra information besides their
79+ alignment. However, if TARGET_PTRMEMFUNC_VBIT_LOCATION
80+ allows the low bit to be used as a virtual bit, we know
81+ that the address itself must be 2-byte aligned. */
82+ if (TARGET_PTRMEMFUNC_VBIT_LOCATION == ptrmemfunc_vbit_in_pfn)
83+ align = 2 * BITS_PER_UNIT;
84+ else
85+ align = BITS_PER_UNIT;
86+ }
87+ else
88+ align = DECL_ALIGN (exp);
89+ }
90 else if (CONSTANT_CLASS_P (exp))
91 {
92 align = TYPE_ALIGN (TREE_TYPE (exp));
93
94=== modified file 'gcc/fold-const.c'
95--- old/gcc/fold-const.c 2011-05-23 20:37:18 +0000
96+++ new/gcc/fold-const.c 2011-07-04 09:52:27 +0000
97@@ -9232,15 +9232,10 @@
98 0 <= N < M as is common. In general, the precise value of P is unknown.
99 M is chosen as large as possible such that constant N can be determined.
100
101- Returns M and sets *RESIDUE to N.
102-
103- If ALLOW_FUNC_ALIGN is true, do take functions' DECL_ALIGN_UNIT into
104- account. This is not always possible due to PR 35705.
105- */
106+ Returns M and sets *RESIDUE to N. */
107
108 static unsigned HOST_WIDE_INT
109-get_pointer_modulus_and_residue (tree expr, unsigned HOST_WIDE_INT *residue,
110- bool allow_func_align)
111+get_pointer_modulus_and_residue (tree expr, unsigned HOST_WIDE_INT *residue)
112 {
113 enum tree_code code;
114
115@@ -9270,9 +9265,8 @@
116 }
117 }
118
119- if (DECL_P (expr)
120- && (allow_func_align || TREE_CODE (expr) != FUNCTION_DECL))
121- return DECL_ALIGN_UNIT (expr);
122+ if (DECL_P (expr))
123+ return get_object_alignment (expr, ~0U) / BITS_PER_UNIT;
124 }
125 else if (code == POINTER_PLUS_EXPR)
126 {
127@@ -9282,8 +9276,7 @@
128
129 op0 = TREE_OPERAND (expr, 0);
130 STRIP_NOPS (op0);
131- modulus = get_pointer_modulus_and_residue (op0, residue,
132- allow_func_align);
133+ modulus = get_pointer_modulus_and_residue (op0, residue);
134
135 op1 = TREE_OPERAND (expr, 1);
136 STRIP_NOPS (op1);
137@@ -11163,8 +11156,7 @@
138 unsigned HOST_WIDE_INT modulus, residue;
139 unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg1);
140
141- modulus = get_pointer_modulus_and_residue (arg0, &residue,
142- integer_onep (arg1));
143+ modulus = get_pointer_modulus_and_residue (arg0, &residue);
144
145 /* This works because modulus is a power of 2. If this weren't the
146 case, we'd have to replace it by its greatest power-of-2
147
148=== added file 'gcc/testsuite/gcc.dg/torture/pr49169.c'
149--- old/gcc/testsuite/gcc.dg/torture/pr49169.c 1970-01-01 00:00:00 +0000
150+++ new/gcc/testsuite/gcc.dg/torture/pr49169.c 2011-06-29 09:46:06 +0000
151@@ -0,0 +1,15 @@
152+/* { dg-do compile { target { arm*-*-* || mips*-*-* } } } */
153+
154+#include <stdlib.h>
155+#include <stdint.h>
156+
157+int
158+main (void)
159+{
160+ void *p = main;
161+ if ((intptr_t) p & 1)
162+ abort ();
163+ return 0;
164+}
165+
166+/* { dg-final { scan-assembler "abort" } } */
167
168=== modified file 'gcc/tree-ssa-ccp.c'
169--- old/gcc/tree-ssa-ccp.c 2011-05-05 15:42:22 +0000
170+++ new/gcc/tree-ssa-ccp.c 2011-06-29 09:46:06 +0000
171@@ -522,10 +522,6 @@
172 val = bit_value_binop (PLUS_EXPR, TREE_TYPE (expr),
173 TREE_OPERAND (base, 0), TREE_OPERAND (base, 1));
174 else if (base
175- /* ??? While function decls have DECL_ALIGN their addresses
176- may encode extra information in the lower bits on some
177- targets (PR47239). Simply punt for function decls for now. */
178- && TREE_CODE (base) != FUNCTION_DECL
179 && ((align = get_object_alignment (base, BIGGEST_ALIGNMENT))
180 > BITS_PER_UNIT))
181 {
182
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106769.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106769.patch
new file mode 100644
index 000000000..37e3036b2
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106769.patch
@@ -0,0 +1,1294 @@
12011-07-11 Ira Rosen <ira.rosen@linaro.org>
2
3 Backport from FSF:
4 2011-06-16 Ira Rosen <ira.rosen@linaro.org>
5
6 gcc/
7 * tree-vectorizer.h (vect_recog_func_ptr): Change the first
8 argument to be a VEC of statements.
9 * tree-vect-loop.c (vect_determine_vectorization_factor): Remove the
10 assert that pattern statements have to have their vector type set.
11 * tree-vect-patterns.c (vect_recog_widen_sum_pattern):
12 Change the first argument to be a VEC of statements. Update
13 documentation.
14 (vect_recog_dot_prod_pattern, vect_recog_pow_pattern): Likewise.
15 (vect_handle_widen_mult_by_const): New function.
16 (vect_recog_widen_mult_pattern): Change the first argument to be a
17 VEC of statements. Update documentation. Check that the constant is
18 INTEGER_CST. Support multiplication by a constant that fits an
19 intermediate type - call vect_handle_widen_mult_by_const.
20 (vect_pattern_recog_1): Update vect_recog_func_ptr and its
21 call. Handle additional pattern statements if necessary.
22
23 gcc/testsuite/
24 * gcc.dg/vect/vect-widen-mult-half-u8.c: New test.
25
26 and
27 2011-06-30 Ira Rosen <ira.rosen@linaro.org>
28
29 gcc/
30 * tree-vect-loop.c (vect_determine_vectorization_factor): Handle
31 both pattern and original statements if necessary.
32 (vect_transform_loop): Likewise.
33 * tree-vect-patterns.c (vect_pattern_recog): Update documentation.
34 * tree-vect-stmts.c (vect_mark_relevant): Add new argument.
35 Mark the pattern statement only if the original statement doesn't
36 have its own uses.
37 (process_use): Call vect_mark_relevant with additional parameter.
38 (vect_mark_stmts_to_be_vectorized): Likewise.
39 (vect_get_vec_def_for_operand): Use vectorized pattern statement.
40 (vect_analyze_stmt): Handle both pattern and original statements
41 if necessary.
42 (vect_transform_stmt): Don't store vectorized pattern statement
43 in the original statement.
44 (vect_is_simple_use_1): Use related pattern statement only if the
45 original statement is irrelevant.
46 * tree-vect-slp.c (vect_get_and_check_slp_defs): Likewise.
47
48 gcc/testsuite/
49 * gcc.dg/vect/slp-widen-mult-half.c: New test.
50 * gcc.dg/vect/vect-widen-mult-half.c: New test.
51
52=== added file 'gcc/testsuite/gcc.dg/vect/slp-widen-mult-half.c'
53--- old/gcc/testsuite/gcc.dg/vect/slp-widen-mult-half.c 1970-01-01 00:00:00 +0000
54+++ new/gcc/testsuite/gcc.dg/vect/slp-widen-mult-half.c 2011-07-06 12:04:10 +0000
55@@ -0,0 +1,52 @@
56+/* { dg-require-effective-target vect_int } */
57+
58+#include "tree-vect.h"
59+#include <stdlib.h>
60+
61+#define N 32
62+#define COEF 32470
63+#define COEF2 324700
64+
65+unsigned char in[N];
66+int out[N];
67+int out2[N];
68+
69+__attribute__ ((noinline)) void
70+foo ()
71+{
72+ int i;
73+
74+ for (i = 0; i < N/2; i++)
75+ {
76+ out[2*i] = in[2*i] * COEF;
77+ out2[2*i] = in[2*i] + COEF2;
78+ out[2*i+1] = in[2*i+1] * COEF;
79+ out2[2*i+1] = in[2*i+1] + COEF2;
80+ }
81+}
82+
83+int main (void)
84+{
85+ int i;
86+
87+ for (i = 0; i < N; i++)
88+ {
89+ in[i] = i;
90+ __asm__ volatile ("");
91+ }
92+
93+ foo ();
94+
95+ for (i = 0; i < N; i++)
96+ if (out[i] != in[i] * COEF || out2[i] != in[i] + COEF2)
97+ abort ();
98+
99+ return 0;
100+}
101+
102+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_widen_mult_hi_to_si } } } */
103+/* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 2 "vect" { target vect_widen_mult_hi_to_si } } } */
104+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
105+/* { dg-final { scan-tree-dump-times "pattern recognized" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
106+/* { dg-final { cleanup-tree-dump "vect" } } */
107+
108
109=== added file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-half-u8.c'
110--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-half-u8.c 1970-01-01 00:00:00 +0000
111+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-half-u8.c 2011-07-06 12:04:10 +0000
112@@ -0,0 +1,59 @@
113+/* { dg-require-effective-target vect_int } */
114+
115+#include "tree-vect.h"
116+#include <stdlib.h>
117+
118+#define N 32
119+#define COEF 32470
120+
121+unsigned char in[N];
122+int out[N];
123+
124+__attribute__ ((noinline)) void
125+foo ()
126+{
127+ int i;
128+
129+ for (i = 0; i < N; i++)
130+ out[i] = in[i] * COEF;
131+}
132+
133+__attribute__ ((noinline)) void
134+bar ()
135+{
136+ int i;
137+
138+ for (i = 0; i < N; i++)
139+ out[i] = COEF * in[i];
140+}
141+
142+int main (void)
143+{
144+ int i;
145+
146+ for (i = 0; i < N; i++)
147+ {
148+ in[i] = i;
149+ __asm__ volatile ("");
150+ }
151+
152+ foo ();
153+
154+ for (i = 0; i < N; i++)
155+ if (out[i] != in[i] * COEF)
156+ abort ();
157+
158+ bar ();
159+
160+ for (i = 0; i < N; i++)
161+ if (out[i] != in[i] * COEF)
162+ abort ();
163+
164+ return 0;
165+}
166+
167+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" { target vect_widen_mult_hi_to_si } } } */
168+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
169+/* { dg-final { scan-tree-dump-times "pattern recognized" 2 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
170+/* { dg-final { cleanup-tree-dump "vect" } } */
171+
172
173=== added file 'gcc/testsuite/gcc.dg/vect/vect-widen-mult-half.c'
174--- old/gcc/testsuite/gcc.dg/vect/vect-widen-mult-half.c 1970-01-01 00:00:00 +0000
175+++ new/gcc/testsuite/gcc.dg/vect/vect-widen-mult-half.c 2011-07-06 12:04:10 +0000
176@@ -0,0 +1,49 @@
177+/* { dg-require-effective-target vect_int } */
178+
179+#include "tree-vect.h"
180+#include <stdlib.h>
181+
182+#define N 32
183+#define COEF 32470
184+#define COEF2 324700
185+
186+unsigned char in[N];
187+int out[N];
188+int out2[N];
189+
190+__attribute__ ((noinline)) void
191+foo (int a)
192+{
193+ int i;
194+
195+ for (i = 0; i < N; i++)
196+ {
197+ out[i] = in[i] * COEF;
198+ out2[i] = in[i] + a;
199+ }
200+}
201+
202+int main (void)
203+{
204+ int i;
205+
206+ for (i = 0; i < N; i++)
207+ {
208+ in[i] = i;
209+ __asm__ volatile ("");
210+ }
211+
212+ foo (COEF2);
213+
214+ for (i = 0; i < N; i++)
215+ if (out[i] != in[i] * COEF || out2[i] != in[i] + COEF2)
216+ abort ();
217+
218+ return 0;
219+}
220+
221+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_widen_mult_hi_to_si } } } */
222+/* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 1 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
223+/* { dg-final { scan-tree-dump-times "pattern recognized" 1 "vect" { target vect_widen_mult_hi_to_si_pattern } } } */
224+/* { dg-final { cleanup-tree-dump "vect" } } */
225+
226
227=== modified file 'gcc/tree-vect-loop.c'
228--- old/gcc/tree-vect-loop.c 2011-07-04 11:13:51 +0000
229+++ new/gcc/tree-vect-loop.c 2011-07-11 11:02:55 +0000
230@@ -181,6 +181,8 @@
231 stmt_vec_info stmt_info;
232 int i;
233 HOST_WIDE_INT dummy;
234+ gimple stmt, pattern_stmt = NULL;
235+ bool analyze_pattern_stmt = false;
236
237 if (vect_print_dump_info (REPORT_DETAILS))
238 fprintf (vect_dump, "=== vect_determine_vectorization_factor ===");
239@@ -241,12 +243,20 @@
240 }
241 }
242
243- for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
244+ for (si = gsi_start_bb (bb); !gsi_end_p (si) || analyze_pattern_stmt;)
245 {
246- tree vf_vectype;
247- gimple stmt = gsi_stmt (si), pattern_stmt;
248- stmt_info = vinfo_for_stmt (stmt);
249-
250+ tree vf_vectype;
251+
252+ if (analyze_pattern_stmt)
253+ {
254+ stmt = pattern_stmt;
255+ analyze_pattern_stmt = false;
256+ }
257+ else
258+ stmt = gsi_stmt (si);
259+
260+ stmt_info = vinfo_for_stmt (stmt);
261+
262 if (vect_print_dump_info (REPORT_DETAILS))
263 {
264 fprintf (vect_dump, "==> examining statement: ");
265@@ -276,10 +286,17 @@
266 {
267 if (vect_print_dump_info (REPORT_DETAILS))
268 fprintf (vect_dump, "skip.");
269+ gsi_next (&si);
270 continue;
271 }
272 }
273
274+ else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
275+ && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
276+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
277+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
278+ analyze_pattern_stmt = true;
279+
280 if (gimple_get_lhs (stmt) == NULL_TREE)
281 {
282 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
283@@ -311,9 +328,7 @@
284 }
285 else
286 {
287- gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)
288- && !is_pattern_stmt_p (stmt_info));
289-
290+ gcc_assert (!STMT_VINFO_DATA_REF (stmt_info));
291 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
292 if (vect_print_dump_info (REPORT_DETAILS))
293 {
294@@ -385,6 +400,9 @@
295 if (!vectorization_factor
296 || (nunits > vectorization_factor))
297 vectorization_factor = nunits;
298+
299+ if (!analyze_pattern_stmt)
300+ gsi_next (&si);
301 }
302 }
303
304@@ -4740,6 +4758,8 @@
305 tree cond_expr = NULL_TREE;
306 gimple_seq cond_expr_stmt_list = NULL;
307 bool do_peeling_for_loop_bound;
308+ gimple stmt, pattern_stmt;
309+ bool transform_pattern_stmt = false;
310
311 if (vect_print_dump_info (REPORT_DETAILS))
312 fprintf (vect_dump, "=== vec_transform_loop ===");
313@@ -4827,11 +4847,19 @@
314 }
315 }
316
317- for (si = gsi_start_bb (bb); !gsi_end_p (si);)
318+ pattern_stmt = NULL;
319+ for (si = gsi_start_bb (bb); !gsi_end_p (si) || transform_pattern_stmt;)
320 {
321- gimple stmt = gsi_stmt (si), pattern_stmt;
322 bool is_store;
323
324+ if (transform_pattern_stmt)
325+ {
326+ stmt = pattern_stmt;
327+ transform_pattern_stmt = false;
328+ }
329+ else
330+ stmt = gsi_stmt (si);
331+
332 if (vect_print_dump_info (REPORT_DETAILS))
333 {
334 fprintf (vect_dump, "------>vectorizing statement: ");
335@@ -4869,6 +4897,11 @@
336 continue;
337 }
338 }
339+ else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
340+ && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
341+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
342+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
343+ transform_pattern_stmt = true;
344
345 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
346 nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (
347@@ -4897,8 +4930,9 @@
348 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
349 if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info))
350 {
351- gsi_next (&si);
352- continue;
353+ if (!transform_pattern_stmt)
354+ gsi_next (&si);
355+ continue;
356 }
357 }
358
359@@ -4917,7 +4951,7 @@
360 the chain. */
361 vect_remove_stores (DR_GROUP_FIRST_DR (stmt_info));
362 gsi_remove (&si, true);
363- continue;
364+ continue;
365 }
366 else
367 {
368@@ -4927,7 +4961,9 @@
369 continue;
370 }
371 }
372- gsi_next (&si);
373+
374+ if (!transform_pattern_stmt)
375+ gsi_next (&si);
376 } /* stmts in BB */
377 } /* BBs in loop */
378
379
380=== modified file 'gcc/tree-vect-patterns.c'
381--- old/gcc/tree-vect-patterns.c 2011-06-22 12:10:44 +0000
382+++ new/gcc/tree-vect-patterns.c 2011-07-06 12:04:10 +0000
383@@ -39,10 +39,13 @@
384 #include "diagnostic-core.h"
385
386 /* Pattern recognition functions */
387-static gimple vect_recog_widen_sum_pattern (gimple *, tree *, tree *);
388-static gimple vect_recog_widen_mult_pattern (gimple *, tree *, tree *);
389-static gimple vect_recog_dot_prod_pattern (gimple *, tree *, tree *);
390-static gimple vect_recog_pow_pattern (gimple *, tree *, tree *);
391+static gimple vect_recog_widen_sum_pattern (VEC (gimple, heap) **, tree *,
392+ tree *);
393+static gimple vect_recog_widen_mult_pattern (VEC (gimple, heap) **, tree *,
394+ tree *);
395+static gimple vect_recog_dot_prod_pattern (VEC (gimple, heap) **, tree *,
396+ tree *);
397+static gimple vect_recog_pow_pattern (VEC (gimple, heap) **, tree *, tree *);
398 static vect_recog_func_ptr vect_vect_recog_func_ptrs[NUM_PATTERNS] = {
399 vect_recog_widen_mult_pattern,
400 vect_recog_widen_sum_pattern,
401@@ -142,9 +145,9 @@
402
403 Input:
404
405- * LAST_STMT: A stmt from which the pattern search begins. In the example,
406- when this function is called with S7, the pattern {S3,S4,S5,S6,S7} will be
407- detected.
408+ * STMTS: Contains a stmt from which the pattern search begins. In the
409+ example, when this function is called with S7, the pattern {S3,S4,S5,S6,S7}
410+ will be detected.
411
412 Output:
413
414@@ -165,12 +168,13 @@
415 inner-loop nested in an outer-loop that us being vectorized). */
416
417 static gimple
418-vect_recog_dot_prod_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
419+vect_recog_dot_prod_pattern (VEC (gimple, heap) **stmts, tree *type_in,
420+ tree *type_out)
421 {
422- gimple stmt;
423+ gimple stmt, last_stmt = VEC_index (gimple, *stmts, 0);
424 tree oprnd0, oprnd1;
425 tree oprnd00, oprnd01;
426- stmt_vec_info stmt_vinfo = vinfo_for_stmt (*last_stmt);
427+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
428 tree type, half_type;
429 gimple pattern_stmt;
430 tree prod_type;
431@@ -178,10 +182,10 @@
432 struct loop *loop = LOOP_VINFO_LOOP (loop_info);
433 tree var, rhs;
434
435- if (!is_gimple_assign (*last_stmt))
436+ if (!is_gimple_assign (last_stmt))
437 return NULL;
438
439- type = gimple_expr_type (*last_stmt);
440+ type = gimple_expr_type (last_stmt);
441
442 /* Look for the following pattern
443 DX = (TYPE1) X;
444@@ -207,7 +211,7 @@
445 /* Starting from LAST_STMT, follow the defs of its uses in search
446 of the above pattern. */
447
448- if (gimple_assign_rhs_code (*last_stmt) != PLUS_EXPR)
449+ if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
450 return NULL;
451
452 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
453@@ -228,12 +232,12 @@
454
455 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
456 return NULL;
457- oprnd0 = gimple_assign_rhs1 (*last_stmt);
458- oprnd1 = gimple_assign_rhs2 (*last_stmt);
459+ oprnd0 = gimple_assign_rhs1 (last_stmt);
460+ oprnd1 = gimple_assign_rhs2 (last_stmt);
461 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
462 || !types_compatible_p (TREE_TYPE (oprnd1), type))
463 return NULL;
464- stmt = *last_stmt;
465+ stmt = last_stmt;
466
467 if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt, true))
468 {
469@@ -319,11 +323,79 @@
470
471 /* We don't allow changing the order of the computation in the inner-loop
472 when doing outer-loop vectorization. */
473- gcc_assert (!nested_in_vect_loop_p (loop, *last_stmt));
474+ gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
475
476 return pattern_stmt;
477 }
478
479+/* Handle two cases of multiplication by a constant. The first one is when
480+ the constant, CONST_OPRND, fits the type (HALF_TYPE) of the second
481+ operand (OPRND). In that case, we can peform widen-mult from HALF_TYPE to
482+ TYPE.
483+
484+ Otherwise, if the type of the result (TYPE) is at least 4 times bigger than
485+ HALF_TYPE, and CONST_OPRND fits an intermediate type (2 times smaller than
486+ TYPE), we can perform widen-mult from the intermediate type to TYPE and
487+ replace a_T = (TYPE) a_t; with a_it - (interm_type) a_t; */
488+
489+static bool
490+vect_handle_widen_mult_by_const (tree const_oprnd, tree *oprnd,
491+ VEC (gimple, heap) **stmts, tree type,
492+ tree *half_type, gimple def_stmt)
493+{
494+ tree new_type, new_oprnd, tmp;
495+ gimple new_stmt;
496+
497+ if (int_fits_type_p (const_oprnd, *half_type))
498+ {
499+ /* CONST_OPRND is a constant of HALF_TYPE. */
500+ *oprnd = gimple_assign_rhs1 (def_stmt);
501+ return true;
502+ }
503+
504+ if (TYPE_PRECISION (type) < (TYPE_PRECISION (*half_type) * 4)
505+ || !vinfo_for_stmt (def_stmt))
506+ return false;
507+
508+ /* TYPE is 4 times bigger than HALF_TYPE, try widen-mult for
509+ a type 2 times bigger than HALF_TYPE. */
510+ new_type = build_nonstandard_integer_type (TYPE_PRECISION (type) / 2,
511+ TYPE_UNSIGNED (type));
512+ if (!int_fits_type_p (const_oprnd, new_type))
513+ return false;
514+
515+ /* Use NEW_TYPE for widen_mult. */
516+ if (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)))
517+ {
518+ new_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt));
519+ /* Check if the already created pattern stmt is what we need. */
520+ if (!is_gimple_assign (new_stmt)
521+ || gimple_assign_rhs_code (new_stmt) != NOP_EXPR
522+ || TREE_TYPE (gimple_assign_lhs (new_stmt)) != new_type)
523+ return false;
524+
525+ *oprnd = gimple_assign_lhs (new_stmt);
526+ }
527+ else
528+ {
529+ /* Create a_T = (NEW_TYPE) a_t; */
530+ *oprnd = gimple_assign_rhs1 (def_stmt);
531+ tmp = create_tmp_var (new_type, NULL);
532+ add_referenced_var (tmp);
533+ new_oprnd = make_ssa_name (tmp, NULL);
534+ new_stmt = gimple_build_assign_with_ops (NOP_EXPR, new_oprnd, *oprnd,
535+ NULL_TREE);
536+ SSA_NAME_DEF_STMT (new_oprnd) = new_stmt;
537+ STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)) = new_stmt;
538+ VEC_safe_push (gimple, heap, *stmts, def_stmt);
539+ *oprnd = new_oprnd;
540+ }
541+
542+ *half_type = new_type;
543+ return true;
544+}
545+
546+
547 /* Function vect_recog_widen_mult_pattern
548
549 Try to find the following pattern:
550@@ -361,28 +433,47 @@
551 S3 a_T = (TYPE) a_t;
552 S5 prod_T = a_T * CONST;
553
554- Input:
555-
556- * LAST_STMT: A stmt from which the pattern search begins. In the example,
557- when this function is called with S5, the pattern {S3,S4,S5,(S6)} is
558- detected.
559-
560- Output:
561-
562- * TYPE_IN: The type of the input arguments to the pattern.
563-
564- * TYPE_OUT: The type of the output of this pattern.
565-
566- * Return value: A new stmt that will be used to replace the sequence of
567- stmts that constitute the pattern. In this case it will be:
568- WIDEN_MULT <a_t, b_t>
569- */
570+ A special case of multiplication by constants is when 'TYPE' is 4 times
571+ bigger than 'type', but CONST fits an intermediate type 2 times smaller
572+ than 'TYPE'. In that case we create an additional pattern stmt for S3
573+ to create a variable of the intermediate type, and perform widen-mult
574+ on the intermediate type as well:
575+
576+ type a_t;
577+ interm_type a_it;
578+ TYPE a_T, prod_T, prod_T';
579+
580+ S1 a_t = ;
581+ S3 a_T = (TYPE) a_t;
582+ '--> a_it = (interm_type) a_t;
583+ S5 prod_T = a_T * CONST;
584+ '--> prod_T' = a_it w* CONST;
585+
586+ Input/Output:
587+
588+ * STMTS: Contains a stmt from which the pattern search begins. In the
589+ example, when this function is called with S5, the pattern {S3,S4,S5,(S6)}
590+ is detected. In case of unsigned widen-mult, the original stmt (S5) is
591+ replaced with S6 in STMTS. In case of multiplication by a constant
592+ of an intermediate type (the last case above), STMTS also contains S3
593+ (inserted before S5).
594+
595+ Output:
596+
597+ * TYPE_IN: The type of the input arguments to the pattern.
598+
599+ * TYPE_OUT: The type of the output of this pattern.
600+
601+ * Return value: A new stmt that will be used to replace the sequence of
602+ stmts that constitute the pattern. In this case it will be:
603+ WIDEN_MULT <a_t, b_t>
604+*/
605
606 static gimple
607-vect_recog_widen_mult_pattern (gimple *last_stmt,
608- tree *type_in,
609- tree *type_out)
610+vect_recog_widen_mult_pattern (VEC (gimple, heap) **stmts,
611+ tree *type_in, tree *type_out)
612 {
613+ gimple last_stmt = VEC_pop (gimple, *stmts);
614 gimple def_stmt0, def_stmt1;
615 tree oprnd0, oprnd1;
616 tree type, half_type0, half_type1;
617@@ -395,27 +486,27 @@
618 VEC (tree, heap) *dummy_vec;
619 bool op0_ok, op1_ok;
620
621- if (!is_gimple_assign (*last_stmt))
622+ if (!is_gimple_assign (last_stmt))
623 return NULL;
624
625- type = gimple_expr_type (*last_stmt);
626+ type = gimple_expr_type (last_stmt);
627
628 /* Starting from LAST_STMT, follow the defs of its uses in search
629 of the above pattern. */
630
631- if (gimple_assign_rhs_code (*last_stmt) != MULT_EXPR)
632+ if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
633 return NULL;
634
635- oprnd0 = gimple_assign_rhs1 (*last_stmt);
636- oprnd1 = gimple_assign_rhs2 (*last_stmt);
637+ oprnd0 = gimple_assign_rhs1 (last_stmt);
638+ oprnd1 = gimple_assign_rhs2 (last_stmt);
639 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
640 || !types_compatible_p (TREE_TYPE (oprnd1), type))
641 return NULL;
642
643 /* Check argument 0. */
644- op0_ok = widened_name_p (oprnd0, *last_stmt, &half_type0, &def_stmt0, false);
645+ op0_ok = widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0, false);
646 /* Check argument 1. */
647- op1_ok = widened_name_p (oprnd1, *last_stmt, &half_type1, &def_stmt1, false);
648+ op1_ok = widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1, false);
649
650 /* In case of multiplication by a constant one of the operands may not match
651 the pattern, but not both. */
652@@ -429,29 +520,21 @@
653 }
654 else if (!op0_ok)
655 {
656- if (CONSTANT_CLASS_P (oprnd0)
657- && TREE_CODE (half_type1) == INTEGER_TYPE
658- && tree_int_cst_lt (oprnd0, TYPE_MAXVAL (half_type1))
659- && tree_int_cst_lt (TYPE_MINVAL (half_type1), oprnd0))
660- {
661- /* OPRND0 is a constant of HALF_TYPE1. */
662- half_type0 = half_type1;
663- oprnd1 = gimple_assign_rhs1 (def_stmt1);
664- }
665+ if (TREE_CODE (oprnd0) == INTEGER_CST
666+ && TREE_CODE (half_type1) == INTEGER_TYPE
667+ && vect_handle_widen_mult_by_const (oprnd0, &oprnd1, stmts, type,
668+ &half_type1, def_stmt1))
669+ half_type0 = half_type1;
670 else
671 return NULL;
672 }
673 else if (!op1_ok)
674 {
675- if (CONSTANT_CLASS_P (oprnd1)
676+ if (TREE_CODE (oprnd1) == INTEGER_CST
677 && TREE_CODE (half_type0) == INTEGER_TYPE
678- && tree_int_cst_lt (oprnd1, TYPE_MAXVAL (half_type0))
679- && tree_int_cst_lt (TYPE_MINVAL (half_type0), oprnd1))
680- {
681- /* OPRND1 is a constant of HALF_TYPE0. */
682- half_type1 = half_type0;
683- oprnd0 = gimple_assign_rhs1 (def_stmt0);
684- }
685+ && vect_handle_widen_mult_by_const (oprnd1, &oprnd0, stmts, type,
686+ &half_type0, def_stmt0))
687+ half_type1 = half_type0;
688 else
689 return NULL;
690 }
691@@ -461,7 +544,7 @@
692 Use unsigned TYPE as the type for WIDEN_MULT_EXPR. */
693 if (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (half_type0))
694 {
695- tree lhs = gimple_assign_lhs (*last_stmt), use_lhs;
696+ tree lhs = gimple_assign_lhs (last_stmt), use_lhs;
697 imm_use_iterator imm_iter;
698 use_operand_p use_p;
699 int nuses = 0;
700@@ -491,7 +574,7 @@
701 return NULL;
702
703 type = use_type;
704- *last_stmt = use_stmt;
705+ last_stmt = use_stmt;
706 }
707
708 if (!types_compatible_p (half_type0, half_type1))
709@@ -506,7 +589,7 @@
710 vectype_out = get_vectype_for_scalar_type (type);
711 if (!vectype
712 || !vectype_out
713- || !supportable_widening_operation (WIDEN_MULT_EXPR, *last_stmt,
714+ || !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt,
715 vectype_out, vectype,
716 &dummy, &dummy, &dummy_code,
717 &dummy_code, &dummy_int, &dummy_vec))
718@@ -524,6 +607,7 @@
719 if (vect_print_dump_info (REPORT_DETAILS))
720 print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
721
722+ VEC_safe_push (gimple, heap, *stmts, last_stmt);
723 return pattern_stmt;
724 }
725
726@@ -555,16 +639,17 @@
727 */
728
729 static gimple
730-vect_recog_pow_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
731+vect_recog_pow_pattern (VEC (gimple, heap) **stmts, tree *type_in, tree *type_out)
732 {
733+ gimple last_stmt = VEC_index (gimple, *stmts, 0);
734 tree fn, base, exp = NULL;
735 gimple stmt;
736 tree var;
737
738- if (!is_gimple_call (*last_stmt) || gimple_call_lhs (*last_stmt) == NULL)
739+ if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
740 return NULL;
741
742- fn = gimple_call_fndecl (*last_stmt);
743+ fn = gimple_call_fndecl (last_stmt);
744 if (fn == NULL_TREE || DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)
745 return NULL;
746
747@@ -574,8 +659,8 @@
748 case BUILT_IN_POWI:
749 case BUILT_IN_POWF:
750 case BUILT_IN_POW:
751- base = gimple_call_arg (*last_stmt, 0);
752- exp = gimple_call_arg (*last_stmt, 1);
753+ base = gimple_call_arg (last_stmt, 0);
754+ exp = gimple_call_arg (last_stmt, 1);
755 if (TREE_CODE (exp) != REAL_CST
756 && TREE_CODE (exp) != INTEGER_CST)
757 return NULL;
758@@ -667,21 +752,23 @@
759 inner-loop nested in an outer-loop that us being vectorized). */
760
761 static gimple
762-vect_recog_widen_sum_pattern (gimple *last_stmt, tree *type_in, tree *type_out)
763+vect_recog_widen_sum_pattern (VEC (gimple, heap) **stmts, tree *type_in,
764+ tree *type_out)
765 {
766+ gimple last_stmt = VEC_index (gimple, *stmts, 0);
767 gimple stmt;
768 tree oprnd0, oprnd1;
769- stmt_vec_info stmt_vinfo = vinfo_for_stmt (*last_stmt);
770+ stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
771 tree type, half_type;
772 gimple pattern_stmt;
773 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
774 struct loop *loop = LOOP_VINFO_LOOP (loop_info);
775 tree var;
776
777- if (!is_gimple_assign (*last_stmt))
778+ if (!is_gimple_assign (last_stmt))
779 return NULL;
780
781- type = gimple_expr_type (*last_stmt);
782+ type = gimple_expr_type (last_stmt);
783
784 /* Look for the following pattern
785 DX = (TYPE) X;
786@@ -693,25 +780,25 @@
787 /* Starting from LAST_STMT, follow the defs of its uses in search
788 of the above pattern. */
789
790- if (gimple_assign_rhs_code (*last_stmt) != PLUS_EXPR)
791+ if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
792 return NULL;
793
794 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
795 return NULL;
796
797- oprnd0 = gimple_assign_rhs1 (*last_stmt);
798- oprnd1 = gimple_assign_rhs2 (*last_stmt);
799+ oprnd0 = gimple_assign_rhs1 (last_stmt);
800+ oprnd1 = gimple_assign_rhs2 (last_stmt);
801 if (!types_compatible_p (TREE_TYPE (oprnd0), type)
802 || !types_compatible_p (TREE_TYPE (oprnd1), type))
803 return NULL;
804
805- /* So far so good. Since *last_stmt was detected as a (summation) reduction,
806+ /* So far so good. Since last_stmt was detected as a (summation) reduction,
807 we know that oprnd1 is the reduction variable (defined by a loop-header
808 phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
809 Left to check that oprnd0 is defined by a cast from type 'type' to type
810 'TYPE'. */
811
812- if (!widened_name_p (oprnd0, *last_stmt, &half_type, &stmt, true))
813+ if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt, true))
814 return NULL;
815
816 oprnd0 = gimple_assign_rhs1 (stmt);
817@@ -732,8 +819,9 @@
818
819 /* We don't allow changing the order of the computation in the inner-loop
820 when doing outer-loop vectorization. */
821- gcc_assert (!nested_in_vect_loop_p (loop, *last_stmt));
822+ gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
823
824+ VEC_safe_push (gimple, heap, *stmts, last_stmt);
825 return pattern_stmt;
826 }
827
828@@ -762,7 +850,7 @@
829
830 static void
831 vect_pattern_recog_1 (
832- gimple (* vect_recog_func) (gimple *, tree *, tree *),
833+ gimple (* vect_recog_func) (VEC (gimple, heap) **, tree *, tree *),
834 gimple_stmt_iterator si)
835 {
836 gimple stmt = gsi_stmt (si), pattern_stmt;
837@@ -774,12 +862,14 @@
838 enum tree_code code;
839 int i;
840 gimple next;
841+ VEC (gimple, heap) *stmts_to_replace = VEC_alloc (gimple, heap, 1);
842
843- pattern_stmt = (* vect_recog_func) (&stmt, &type_in, &type_out);
844+ VEC_quick_push (gimple, stmts_to_replace, stmt);
845+ pattern_stmt = (* vect_recog_func) (&stmts_to_replace, &type_in, &type_out);
846 if (!pattern_stmt)
847 return;
848
849- si = gsi_for_stmt (stmt);
850+ stmt = VEC_last (gimple, stmts_to_replace);
851 stmt_info = vinfo_for_stmt (stmt);
852 loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
853
854@@ -849,6 +939,35 @@
855 FOR_EACH_VEC_ELT (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i, next)
856 if (next == stmt)
857 VEC_ordered_remove (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i);
858+
859+ /* In case of widen-mult by a constant, it is possible that an additional
860+ pattern stmt is created and inserted in STMTS_TO_REPLACE. We create a
861+ stmt_info for it, and mark the relevant statements. */
862+ for (i = 0; VEC_iterate (gimple, stmts_to_replace, i, stmt)
863+ && (unsigned) i < (VEC_length (gimple, stmts_to_replace) - 1);
864+ i++)
865+ {
866+ stmt_info = vinfo_for_stmt (stmt);
867+ pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
868+ if (vect_print_dump_info (REPORT_DETAILS))
869+ {
870+ fprintf (vect_dump, "additional pattern stmt: ");
871+ print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
872+ }
873+
874+ set_vinfo_for_stmt (pattern_stmt,
875+ new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL));
876+ gimple_set_bb (pattern_stmt, gimple_bb (stmt));
877+ pattern_stmt_info = vinfo_for_stmt (pattern_stmt);
878+
879+ STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt;
880+ STMT_VINFO_DEF_TYPE (pattern_stmt_info)
881+ = STMT_VINFO_DEF_TYPE (stmt_info);
882+ STMT_VINFO_VECTYPE (pattern_stmt_info) = STMT_VINFO_VECTYPE (stmt_info);
883+ STMT_VINFO_IN_PATTERN_P (stmt_info) = true;
884+ }
885+
886+ VEC_free (gimple, heap, stmts_to_replace);
887 }
888
889
890@@ -896,10 +1015,8 @@
891
892 If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
893 (because they are marked as irrelevant). It will vectorize S6, and record
894- a pointer to the new vector stmt VS6 both from S6 (as usual), and also
895- from S4. We do that so that when we get to vectorizing stmts that use the
896- def of S4 (like S5 that uses a_0), we'll know where to take the relevant
897- vector-def from. S4 will be skipped, and S5 will be vectorized as usual:
898+ a pointer to the new vector stmt VS6 from S6 (as usual).
899+ S4 will be skipped, and S5 will be vectorized as usual:
900
901 in_pattern_p related_stmt vec_stmt
902 S1: a_i = .... - - -
903@@ -915,7 +1032,21 @@
904 elsewhere), and we'll end up with:
905
906 VS6: va_new = ....
907- VS5: ... = ..vuse(va_new).. */
908+ VS5: ... = ..vuse(va_new)..
909+
910+ In case of more than one pattern statements, e.g., widen-mult with
911+ intermediate type:
912+
913+ S1 a_t = ;
914+ S2 a_T = (TYPE) a_t;
915+ '--> S3: a_it = (interm_type) a_t;
916+ S4 prod_T = a_T * CONST;
917+ '--> S5: prod_T' = a_it w* CONST;
918+
919+ there may be other users of a_T outside the pattern. In that case S2 will
920+ be marked as relevant (as well as S3), and both S2 and S3 will be analyzed
921+ and vectorized. The vector stmt VS2 will be recorded in S2, and VS3 will
922+ be recorded in S3. */
923
924 void
925 vect_pattern_recog (loop_vec_info loop_vinfo)
926@@ -925,7 +1056,7 @@
927 unsigned int nbbs = loop->num_nodes;
928 gimple_stmt_iterator si;
929 unsigned int i, j;
930- gimple (* vect_recog_func_ptr) (gimple *, tree *, tree *);
931+ gimple (* vect_recog_func_ptr) (VEC (gimple, heap) **, tree *, tree *);
932
933 if (vect_print_dump_info (REPORT_DETAILS))
934 fprintf (vect_dump, "=== vect_pattern_recog ===");
935
936=== modified file 'gcc/tree-vect-slp.c'
937--- old/gcc/tree-vect-slp.c 2011-06-19 10:59:13 +0000
938+++ new/gcc/tree-vect-slp.c 2011-07-06 12:04:10 +0000
939@@ -152,7 +152,9 @@
940 if (loop && def_stmt && gimple_bb (def_stmt)
941 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
942 && vinfo_for_stmt (def_stmt)
943- && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt)))
944+ && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (def_stmt))
945+ && !STMT_VINFO_RELEVANT (vinfo_for_stmt (def_stmt))
946+ && !STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt)))
947 {
948 if (!*first_stmt_dt0)
949 *pattern0 = true;
950
951=== modified file 'gcc/tree-vect-stmts.c'
952--- old/gcc/tree-vect-stmts.c 2011-06-22 06:21:13 +0000
953+++ new/gcc/tree-vect-stmts.c 2011-07-06 12:04:10 +0000
954@@ -126,33 +126,72 @@
955
956 static void
957 vect_mark_relevant (VEC(gimple,heap) **worklist, gimple stmt,
958- enum vect_relevant relevant, bool live_p)
959+ enum vect_relevant relevant, bool live_p,
960+ bool used_in_pattern)
961 {
962 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
963 enum vect_relevant save_relevant = STMT_VINFO_RELEVANT (stmt_info);
964 bool save_live_p = STMT_VINFO_LIVE_P (stmt_info);
965+ gimple pattern_stmt;
966
967 if (vect_print_dump_info (REPORT_DETAILS))
968 fprintf (vect_dump, "mark relevant %d, live %d.", relevant, live_p);
969
970+ /* If this stmt is an original stmt in a pattern, we might need to mark its
971+ related pattern stmt instead of the original stmt. However, such stmts
972+ may have their own uses that are not in any pattern, in such cases the
973+ stmt itself should be marked. */
974 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
975 {
976- gimple pattern_stmt;
977-
978- /* This is the last stmt in a sequence that was detected as a
979- pattern that can potentially be vectorized. Don't mark the stmt
980- as relevant/live because it's not going to be vectorized.
981- Instead mark the pattern-stmt that replaces it. */
982-
983- pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
984-
985- if (vect_print_dump_info (REPORT_DETAILS))
986- fprintf (vect_dump, "last stmt in pattern. don't mark relevant/live.");
987- stmt_info = vinfo_for_stmt (pattern_stmt);
988- gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt);
989- save_relevant = STMT_VINFO_RELEVANT (stmt_info);
990- save_live_p = STMT_VINFO_LIVE_P (stmt_info);
991- stmt = pattern_stmt;
992+ bool found = false;
993+ if (!used_in_pattern)
994+ {
995+ imm_use_iterator imm_iter;
996+ use_operand_p use_p;
997+ gimple use_stmt;
998+ tree lhs;
999+
1000+ if (is_gimple_assign (stmt))
1001+ lhs = gimple_assign_lhs (stmt);
1002+ else
1003+ lhs = gimple_call_lhs (stmt);
1004+
1005+ /* This use is out of pattern use, if LHS has other uses that are
1006+ pattern uses, we should mark the stmt itself, and not the pattern
1007+ stmt. */
1008+ FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs)
1009+ {
1010+ if (is_gimple_debug (USE_STMT (use_p)))
1011+ continue;
1012+ use_stmt = USE_STMT (use_p);
1013+
1014+ if (vinfo_for_stmt (use_stmt)
1015+ && STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt)))
1016+ {
1017+ found = true;
1018+ break;
1019+ }
1020+ }
1021+ }
1022+
1023+ if (!found)
1024+ {
1025+ /* This is the last stmt in a sequence that was detected as a
1026+ pattern that can potentially be vectorized. Don't mark the stmt
1027+ as relevant/live because it's not going to be vectorized.
1028+ Instead mark the pattern-stmt that replaces it. */
1029+
1030+ pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
1031+
1032+ if (vect_print_dump_info (REPORT_DETAILS))
1033+ fprintf (vect_dump, "last stmt in pattern. don't mark"
1034+ " relevant/live.");
1035+ stmt_info = vinfo_for_stmt (pattern_stmt);
1036+ gcc_assert (STMT_VINFO_RELATED_STMT (stmt_info) == stmt);
1037+ save_relevant = STMT_VINFO_RELEVANT (stmt_info);
1038+ save_live_p = STMT_VINFO_LIVE_P (stmt_info);
1039+ stmt = pattern_stmt;
1040+ }
1041 }
1042
1043 STMT_VINFO_LIVE_P (stmt_info) |= live_p;
1044@@ -437,7 +476,8 @@
1045 }
1046 }
1047
1048- vect_mark_relevant (worklist, def_stmt, relevant, live_p);
1049+ vect_mark_relevant (worklist, def_stmt, relevant, live_p,
1050+ is_pattern_stmt_p (stmt_vinfo));
1051 return true;
1052 }
1053
1054@@ -494,7 +534,7 @@
1055 }
1056
1057 if (vect_stmt_relevant_p (phi, loop_vinfo, &relevant, &live_p))
1058- vect_mark_relevant (&worklist, phi, relevant, live_p);
1059+ vect_mark_relevant (&worklist, phi, relevant, live_p, false);
1060 }
1061 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
1062 {
1063@@ -506,7 +546,7 @@
1064 }
1065
1066 if (vect_stmt_relevant_p (stmt, loop_vinfo, &relevant, &live_p))
1067- vect_mark_relevant (&worklist, stmt, relevant, live_p);
1068+ vect_mark_relevant (&worklist, stmt, relevant, live_p, false);
1069 }
1070 }
1071
1072@@ -613,42 +653,55 @@
1073 if (is_gimple_assign (stmt))
1074 {
1075 tree rhs = gimple_assign_rhs1 (stmt);
1076- if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1077- == GIMPLE_SINGLE_RHS)
1078+ unsigned int op_num;
1079+ tree op;
1080+ switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
1081 {
1082- unsigned int op_num = TREE_OPERAND_LENGTH (gimple_assign_rhs1
1083- (stmt));
1084- for (i = 0; i < op_num; i++)
1085- {
1086- tree op = TREE_OPERAND (rhs, i);
1087- if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1088- &worklist))
1089+ case GIMPLE_SINGLE_RHS:
1090+ op_num = TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt));
1091+ for (i = 0; i < op_num; i++)
1092 {
1093- VEC_free (gimple, heap, worklist);
1094- return false;
1095+ op = TREE_OPERAND (rhs, i);
1096+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1097+ &worklist))
1098+ {
1099+ VEC_free (gimple, heap, worklist);
1100+ return false;
1101+ }
1102 }
1103- }
1104- }
1105- else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1106- == GIMPLE_BINARY_RHS)
1107- {
1108- tree op = gimple_assign_rhs1 (stmt);
1109- if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1110- &worklist))
1111- {
1112- VEC_free (gimple, heap, worklist);
1113- return false;
1114- }
1115- op = gimple_assign_rhs2 (stmt);
1116- if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1117- &worklist))
1118- {
1119- VEC_free (gimple, heap, worklist);
1120- return false;
1121- }
1122- }
1123- else
1124- return false;
1125+ break;
1126+
1127+ case GIMPLE_BINARY_RHS:
1128+ op = gimple_assign_rhs1 (stmt);
1129+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1130+ &worklist))
1131+ {
1132+ VEC_free (gimple, heap, worklist);
1133+ return false;
1134+ }
1135+ op = gimple_assign_rhs2 (stmt);
1136+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1137+ &worklist))
1138+ {
1139+ VEC_free (gimple, heap, worklist);
1140+ return false;
1141+ }
1142+ break;
1143+
1144+ case GIMPLE_UNARY_RHS:
1145+ op = gimple_assign_rhs1 (stmt);
1146+ if (!process_use (stmt, op, loop_vinfo, live_p, relevant,
1147+ &worklist))
1148+ {
1149+ VEC_free (gimple, heap, worklist);
1150+ return false;
1151+ }
1152+
1153+ break;
1154+
1155+ default:
1156+ return false;
1157+ }
1158 }
1159 else if (is_gimple_call (stmt))
1160 {
1161@@ -1210,7 +1263,14 @@
1162
1163 /* Get the def from the vectorized stmt. */
1164 def_stmt_info = vinfo_for_stmt (def_stmt);
1165+
1166 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1167+ /* Get vectorized pattern statement. */
1168+ if (!vec_stmt
1169+ && STMT_VINFO_IN_PATTERN_P (def_stmt_info)
1170+ && !STMT_VINFO_RELEVANT (def_stmt_info))
1171+ vec_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (
1172+ STMT_VINFO_RELATED_STMT (def_stmt_info)));
1173 gcc_assert (vec_stmt);
1174 if (gimple_code (vec_stmt) == GIMPLE_PHI)
1175 vec_oprnd = PHI_RESULT (vec_stmt);
1176@@ -4886,6 +4946,7 @@
1177 enum vect_relevant relevance = STMT_VINFO_RELEVANT (stmt_info);
1178 bool ok;
1179 tree scalar_type, vectype;
1180+ gimple pattern_stmt;
1181
1182 if (vect_print_dump_info (REPORT_DETAILS))
1183 {
1184@@ -4907,16 +4968,22 @@
1185 - any LABEL_EXPRs in the loop
1186 - computations that are used only for array indexing or loop control.
1187 In basic blocks we only analyze statements that are a part of some SLP
1188- instance, therefore, all the statements are relevant. */
1189-
1190+ instance, therefore, all the statements are relevant.
1191+
1192+ Pattern statement need to be analyzed instead of the original statement
1193+ if the original statement is not relevant. Otherwise, we analyze both
1194+ statements. */
1195+
1196+ pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
1197 if (!STMT_VINFO_RELEVANT_P (stmt_info)
1198 && !STMT_VINFO_LIVE_P (stmt_info))
1199 {
1200- gimple pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
1201 if (STMT_VINFO_IN_PATTERN_P (stmt_info)
1202+ && pattern_stmt
1203 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
1204 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
1205 {
1206+ /* Analyze PATTERN_STMT instead of the original stmt. */
1207 stmt = pattern_stmt;
1208 stmt_info = vinfo_for_stmt (pattern_stmt);
1209 if (vect_print_dump_info (REPORT_DETAILS))
1210@@ -4933,6 +5000,21 @@
1211 return true;
1212 }
1213 }
1214+ else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
1215+ && pattern_stmt
1216+ && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
1217+ || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
1218+ {
1219+ /* Analyze PATTERN_STMT too. */
1220+ if (vect_print_dump_info (REPORT_DETAILS))
1221+ {
1222+ fprintf (vect_dump, "==> examining pattern statement: ");
1223+ print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
1224+ }
1225+
1226+ if (!vect_analyze_stmt (pattern_stmt, need_to_vectorize, node))
1227+ return false;
1228+ }
1229
1230 switch (STMT_VINFO_DEF_TYPE (stmt_info))
1231 {
1232@@ -5066,7 +5148,6 @@
1233 bool is_store = false;
1234 gimple vec_stmt = NULL;
1235 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1236- gimple orig_stmt_in_pattern, orig_scalar_stmt = stmt;
1237 bool done;
1238
1239 switch (STMT_VINFO_TYPE (stmt_info))
1240@@ -5205,25 +5286,7 @@
1241 }
1242
1243 if (vec_stmt)
1244- {
1245- STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
1246- orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
1247- if (orig_stmt_in_pattern)
1248- {
1249- stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
1250- /* STMT was inserted by the vectorizer to replace a computation idiom.
1251- ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
1252- computed this idiom. We need to record a pointer to VEC_STMT in
1253- the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
1254- documentation of vect_pattern_recog. */
1255- if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
1256- {
1257- gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo)
1258- == orig_scalar_stmt);
1259- STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
1260- }
1261- }
1262- }
1263+ STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
1264
1265 return is_store;
1266 }
1267@@ -5601,8 +5664,12 @@
1268 || *dt == vect_nested_cycle)
1269 {
1270 stmt_vec_info stmt_info = vinfo_for_stmt (*def_stmt);
1271- if (STMT_VINFO_IN_PATTERN_P (stmt_info))
1272+
1273+ if (STMT_VINFO_IN_PATTERN_P (stmt_info)
1274+ && !STMT_VINFO_RELEVANT (stmt_info)
1275+ && !STMT_VINFO_LIVE_P (stmt_info))
1276 stmt_info = vinfo_for_stmt (STMT_VINFO_RELATED_STMT (stmt_info));
1277+
1278 *vectype = STMT_VINFO_VECTYPE (stmt_info);
1279 gcc_assert (*vectype != NULL_TREE);
1280 }
1281
1282=== modified file 'gcc/tree-vectorizer.h'
1283--- old/gcc/tree-vectorizer.h 2011-07-04 11:13:51 +0000
1284+++ new/gcc/tree-vectorizer.h 2011-07-11 11:02:55 +0000
1285@@ -890,7 +890,7 @@
1286 /* Pattern recognition functions.
1287 Additional pattern recognition functions can (and will) be added
1288 in the future. */
1289-typedef gimple (* vect_recog_func_ptr) (gimple *, tree *, tree *);
1290+typedef gimple (* vect_recog_func_ptr) (VEC (gimple, heap) **, tree *, tree *);
1291 #define NUM_PATTERNS 4
1292 void vect_pattern_recog (loop_vec_info);
1293
1294
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106770.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106770.patch
new file mode 100644
index 000000000..82ae3a132
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106770.patch
@@ -0,0 +1,138 @@
12011-07-11 Revital Eres <revital.eres@linaro.org>
2
3 Backport from mainline -r175090.
4 gcc/
5 * ddg.c (add_intra_loop_mem_dep): New function.
6 (build_intra_loop_deps): Call it.
7
8 gcc/testsuite
9 * gcc.dg/sms-9.c: New file.
10
11=== modified file 'gcc/ddg.c'
12--- old/gcc/ddg.c 2011-05-13 16:03:40 +0000
13+++ new/gcc/ddg.c 2011-07-04 11:00:06 +0000
14@@ -390,6 +390,33 @@
15 &PATTERN (insn2));
16 }
17
18+/* Given two nodes, analyze their RTL insns and add intra-loop mem deps
19+ to ddg G. */
20+static void
21+add_intra_loop_mem_dep (ddg_ptr g, ddg_node_ptr from, ddg_node_ptr to)
22+{
23+
24+ if ((from->cuid == to->cuid)
25+ || !insns_may_alias_p (from->insn, to->insn))
26+ /* Do not create edge if memory references have disjoint alias sets
27+ or 'to' and 'from' are the same instruction. */
28+ return;
29+
30+ if (mem_write_insn_p (from->insn))
31+ {
32+ if (mem_read_insn_p (to->insn))
33+ create_ddg_dep_no_link (g, from, to,
34+ DEBUG_INSN_P (to->insn)
35+ ? ANTI_DEP : TRUE_DEP, MEM_DEP, 0);
36+ else
37+ create_ddg_dep_no_link (g, from, to,
38+ DEBUG_INSN_P (to->insn)
39+ ? ANTI_DEP : OUTPUT_DEP, MEM_DEP, 0);
40+ }
41+ else if (!mem_read_insn_p (to->insn))
42+ create_ddg_dep_no_link (g, from, to, ANTI_DEP, MEM_DEP, 0);
43+}
44+
45 /* Given two nodes, analyze their RTL insns and add inter-loop mem deps
46 to ddg G. */
47 static void
48@@ -477,10 +504,22 @@
49 if (DEBUG_INSN_P (j_node->insn))
50 continue;
51 if (mem_access_insn_p (j_node->insn))
52- /* Don't bother calculating inter-loop dep if an intra-loop dep
53- already exists. */
54+ {
55+ /* Don't bother calculating inter-loop dep if an intra-loop dep
56+ already exists. */
57 if (! TEST_BIT (dest_node->successors, j))
58 add_inter_loop_mem_dep (g, dest_node, j_node);
59+ /* If -fmodulo-sched-allow-regmoves
60+ is set certain anti-dep edges are not created.
61+ It might be that these anti-dep edges are on the
62+ path from one memory instruction to another such that
63+ removing these edges could cause a violation of the
64+ memory dependencies. Thus we add intra edges between
65+ every two memory instructions in this case. */
66+ if (flag_modulo_sched_allow_regmoves
67+ && !TEST_BIT (dest_node->predecessors, j))
68+ add_intra_loop_mem_dep (g, j_node, dest_node);
69+ }
70 }
71 }
72 }
73
74=== added file 'gcc/testsuite/gcc.dg/sms-9.c'
75--- old/gcc/testsuite/gcc.dg/sms-9.c 1970-01-01 00:00:00 +0000
76+++ new/gcc/testsuite/gcc.dg/sms-9.c 2011-07-04 11:00:06 +0000
77@@ -0,0 +1,60 @@
78+/* { dg-do run } */
79+/* { dg-options "-O2 -fmodulo-sched -fno-auto-inc-dec -O2 -fmodulo-sched-allow-regmoves" } */
80+
81+#include <stdlib.h>
82+#include <stdarg.h>
83+
84+struct df_ref_info
85+{
86+ unsigned int *begin;
87+ unsigned int *count;
88+};
89+
90+extern void *memset (void *s, int c, __SIZE_TYPE__ n);
91+
92+
93+__attribute__ ((noinline))
94+ int
95+ df_reorganize_refs_by_reg_by_insn (struct df_ref_info *ref_info,
96+ int num, unsigned int start)
97+{
98+ unsigned int m = num;
99+ unsigned int offset = 77;
100+ unsigned int r;
101+
102+ for (r = start; r < m; r++)
103+ {
104+ ref_info->begin[r] = offset;
105+ offset += ref_info->count[r];
106+ ref_info->count[r] = 0;
107+ }
108+
109+ return offset;
110+}
111+
112+int
113+main ()
114+{
115+ struct df_ref_info temp;
116+ int num = 100;
117+ unsigned int start = 5;
118+ int i, offset;
119+
120+ temp.begin = malloc (100 * sizeof (unsigned int));
121+ temp.count = malloc (100 * sizeof (unsigned int));
122+
123+ memset (temp.begin, 0, sizeof (unsigned int) * num);
124+ memset (temp.count, 0, sizeof (unsigned int) * num);
125+
126+ for (i = 0; i < num; i++)
127+ temp.count[i] = i + 1;
128+
129+ offset = df_reorganize_refs_by_reg_by_insn (&temp, num, start);
130+
131+ if (offset != 5112)
132+ abort ();
133+
134+ free (temp.begin);
135+ free (temp.count);
136+ return 0;
137+}
138
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106771.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106771.patch
new file mode 100644
index 000000000..70c8638cd
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106771.patch
@@ -0,0 +1,211 @@
12011-07-11 Revital Eres <revital.eres@linaro.org>
2
3 Backport from mainline -r175091
4 gcc/
5 * modulo-sched.c (struct ps_insn): Remove row_rest_count
6 field.
7 (struct partial_schedule): Add rows_length field.
8 (verify_partial_schedule): Check rows_length.
9 (ps_insert_empty_row): Handle rows_length.
10 (create_partial_schedule): Likewise.
11 (free_partial_schedule): Likewise.
12 (reset_partial_schedule): Likewise.
13 (create_ps_insn): Remove rest_count argument.
14 (remove_node_from_ps): Update rows_length.
15 (add_node_to_ps): Update rows_length and call create_ps_insn without
16 passing row_rest_count.
17 (rotate_partial_schedule): Update rows_length.
18
19=== modified file 'gcc/modulo-sched.c'
20--- old/gcc/modulo-sched.c 2011-05-13 16:03:40 +0000
21+++ new/gcc/modulo-sched.c 2011-07-04 12:01:34 +0000
22@@ -134,8 +134,6 @@
23 ps_insn_ptr next_in_row,
24 prev_in_row;
25
26- /* The number of nodes in the same row that come after this node. */
27- int row_rest_count;
28 };
29
30 /* Holds the partial schedule as an array of II rows. Each entry of the
31@@ -149,6 +147,12 @@
32 /* rows[i] points to linked list of insns scheduled in row i (0<=i<ii). */
33 ps_insn_ptr *rows;
34
35+ /* rows_length[i] holds the number of instructions in the row.
36+ It is used only (as an optimization) to back off quickly from
37+ trying to schedule a node in a full row; that is, to avoid running
38+ through futile DFA state transitions. */
39+ int *rows_length;
40+
41 /* The earliest absolute cycle of an insn in the partial schedule. */
42 int min_cycle;
43
44@@ -1907,6 +1911,7 @@
45 int ii = ps->ii;
46 int new_ii = ii + 1;
47 int row;
48+ int *rows_length_new;
49
50 verify_partial_schedule (ps, sched_nodes);
51
52@@ -1921,9 +1926,11 @@
53 rotate_partial_schedule (ps, PS_MIN_CYCLE (ps));
54
55 rows_new = (ps_insn_ptr *) xcalloc (new_ii, sizeof (ps_insn_ptr));
56+ rows_length_new = (int *) xcalloc (new_ii, sizeof (int));
57 for (row = 0; row < split_row; row++)
58 {
59 rows_new[row] = ps->rows[row];
60+ rows_length_new[row] = ps->rows_length[row];
61 ps->rows[row] = NULL;
62 for (crr_insn = rows_new[row];
63 crr_insn; crr_insn = crr_insn->next_in_row)
64@@ -1944,6 +1951,7 @@
65 for (row = split_row; row < ii; row++)
66 {
67 rows_new[row + 1] = ps->rows[row];
68+ rows_length_new[row + 1] = ps->rows_length[row];
69 ps->rows[row] = NULL;
70 for (crr_insn = rows_new[row + 1];
71 crr_insn; crr_insn = crr_insn->next_in_row)
72@@ -1965,6 +1973,8 @@
73 + (SMODULO (ps->max_cycle, ii) >= split_row ? 1 : 0);
74 free (ps->rows);
75 ps->rows = rows_new;
76+ free (ps->rows_length);
77+ ps->rows_length = rows_length_new;
78 ps->ii = new_ii;
79 gcc_assert (ps->min_cycle >= 0);
80
81@@ -2040,16 +2050,23 @@
82 ps_insn_ptr crr_insn;
83
84 for (row = 0; row < ps->ii; row++)
85- for (crr_insn = ps->rows[row]; crr_insn; crr_insn = crr_insn->next_in_row)
86- {
87- ddg_node_ptr u = crr_insn->node;
88-
89- gcc_assert (TEST_BIT (sched_nodes, u->cuid));
90- /* ??? Test also that all nodes of sched_nodes are in ps, perhaps by
91- popcount (sched_nodes) == number of insns in ps. */
92- gcc_assert (SCHED_TIME (u) >= ps->min_cycle);
93- gcc_assert (SCHED_TIME (u) <= ps->max_cycle);
94- }
95+ {
96+ int length = 0;
97+
98+ for (crr_insn = ps->rows[row]; crr_insn; crr_insn = crr_insn->next_in_row)
99+ {
100+ ddg_node_ptr u = crr_insn->node;
101+
102+ length++;
103+ gcc_assert (TEST_BIT (sched_nodes, u->cuid));
104+ /* ??? Test also that all nodes of sched_nodes are in ps, perhaps by
105+ popcount (sched_nodes) == number of insns in ps. */
106+ gcc_assert (SCHED_TIME (u) >= ps->min_cycle);
107+ gcc_assert (SCHED_TIME (u) <= ps->max_cycle);
108+ }
109+
110+ gcc_assert (ps->rows_length[row] == length);
111+ }
112 }
113
114
115@@ -2455,6 +2472,7 @@
116 {
117 partial_schedule_ptr ps = XNEW (struct partial_schedule);
118 ps->rows = (ps_insn_ptr *) xcalloc (ii, sizeof (ps_insn_ptr));
119+ ps->rows_length = (int *) xcalloc (ii, sizeof (int));
120 ps->ii = ii;
121 ps->history = history;
122 ps->min_cycle = INT_MAX;
123@@ -2493,6 +2511,7 @@
124 return;
125 free_ps_insns (ps);
126 free (ps->rows);
127+ free (ps->rows_length);
128 free (ps);
129 }
130
131@@ -2510,6 +2529,8 @@
132 ps->rows = (ps_insn_ptr *) xrealloc (ps->rows, new_ii
133 * sizeof (ps_insn_ptr));
134 memset (ps->rows, 0, new_ii * sizeof (ps_insn_ptr));
135+ ps->rows_length = (int *) xrealloc (ps->rows_length, new_ii * sizeof (int));
136+ memset (ps->rows_length, 0, new_ii * sizeof (int));
137 ps->ii = new_ii;
138 ps->min_cycle = INT_MAX;
139 ps->max_cycle = INT_MIN;
140@@ -2538,14 +2559,13 @@
141
142 /* Creates an object of PS_INSN and initializes it to the given parameters. */
143 static ps_insn_ptr
144-create_ps_insn (ddg_node_ptr node, int rest_count, int cycle)
145+create_ps_insn (ddg_node_ptr node, int cycle)
146 {
147 ps_insn_ptr ps_i = XNEW (struct ps_insn);
148
149 ps_i->node = node;
150 ps_i->next_in_row = NULL;
151 ps_i->prev_in_row = NULL;
152- ps_i->row_rest_count = rest_count;
153 ps_i->cycle = cycle;
154
155 return ps_i;
156@@ -2578,6 +2598,8 @@
157 if (ps_i->next_in_row)
158 ps_i->next_in_row->prev_in_row = ps_i->prev_in_row;
159 }
160+
161+ ps->rows_length[row] -= 1;
162 free (ps_i);
163 return true;
164 }
165@@ -2734,17 +2756,12 @@
166 sbitmap must_precede, sbitmap must_follow)
167 {
168 ps_insn_ptr ps_i;
169- int rest_count = 1;
170 int row = SMODULO (cycle, ps->ii);
171
172- if (ps->rows[row]
173- && ps->rows[row]->row_rest_count >= issue_rate)
174+ if (ps->rows_length[row] >= issue_rate)
175 return NULL;
176
177- if (ps->rows[row])
178- rest_count += ps->rows[row]->row_rest_count;
179-
180- ps_i = create_ps_insn (node, rest_count, cycle);
181+ ps_i = create_ps_insn (node, cycle);
182
183 /* Finds and inserts PS_I according to MUST_FOLLOW and
184 MUST_PRECEDE. */
185@@ -2754,6 +2771,7 @@
186 return NULL;
187 }
188
189+ ps->rows_length[row] += 1;
190 return ps_i;
191 }
192
193@@ -2909,11 +2927,16 @@
194 for (i = 0; i < backward_rotates; i++)
195 {
196 ps_insn_ptr first_row = ps->rows[0];
197+ int first_row_length = ps->rows_length[0];
198
199 for (row = 0; row < last_row; row++)
200- ps->rows[row] = ps->rows[row+1];
201+ {
202+ ps->rows[row] = ps->rows[row + 1];
203+ ps->rows_length[row] = ps->rows_length[row + 1];
204+ }
205
206 ps->rows[last_row] = first_row;
207+ ps->rows_length[last_row] = first_row_length;
208 }
209
210 ps->max_cycle -= start_cycle;
211
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106772.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106772.patch
new file mode 100644
index 000000000..d918f9c58
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106772.patch
@@ -0,0 +1,350 @@
12011-07-11 Ramana Radhakrishnan <ramana.radhakrishnan@linaro.org>
2
3 gcc/
4 2011-06-22 Dmitry Plotnikov <dplotnikov@ispras.ru>
5 Dmitry Melnik <dm@ispras.ru>
6
7 * config/arm/arm.c (neon_immediate_valid_for_shift): New function.
8 (neon_output_shift_immediate): Ditto.
9 * config/arm/arm-protos.h (neon_immediate_valid_for_shift): New
10 prototype.
11 (neon_output_shift_immediate): Ditto.
12 * config/arm/neon.md (vashl<mode>3): Modified constraint.
13 (vashr<mode>3_imm): New insn pattern.
14 (vlshr<mode>3_imm): Ditto.
15 (vashr<mode>3): Modified constraint.
16 (vlshr<mode>3): Ditto.
17 * config/arm/predicates.md (imm_for_neon_lshift_operand): New
18 predicate.
19 (imm_for_neon_rshift_operand): Ditto.
20 (imm_lshift_or_reg_neon): Ditto.
21 (imm_rshift_or_reg_neon): Ditto.
22
23 * optabs.c (init_optabs): Init optab codes for vashl, vashr, vlshr.
24
25=== modified file 'gcc/config/arm/arm-protos.h'
26--- old/gcc/config/arm/arm-protos.h 2011-06-14 16:00:30 +0000
27+++ new/gcc/config/arm/arm-protos.h 2011-07-04 14:03:49 +0000
28@@ -64,8 +64,12 @@
29 extern int neon_immediate_valid_for_move (rtx, enum machine_mode, rtx *, int *);
30 extern int neon_immediate_valid_for_logic (rtx, enum machine_mode, int, rtx *,
31 int *);
32+extern int neon_immediate_valid_for_shift (rtx, enum machine_mode, rtx *,
33+ int *, bool);
34 extern char *neon_output_logic_immediate (const char *, rtx *,
35 enum machine_mode, int, int);
36+extern char *neon_output_shift_immediate (const char *, char, rtx *,
37+ enum machine_mode, int, bool);
38 extern void neon_pairwise_reduce (rtx, rtx, enum machine_mode,
39 rtx (*) (rtx, rtx, rtx));
40 extern rtx neon_make_constant (rtx);
41
42=== modified file 'gcc/config/arm/arm.c'
43--- old/gcc/config/arm/arm.c 2011-06-29 09:13:17 +0000
44+++ new/gcc/config/arm/arm.c 2011-07-04 14:03:49 +0000
45@@ -8863,6 +8863,66 @@
46 return 1;
47 }
48
49+/* Return TRUE if rtx OP is legal for use in a VSHR or VSHL instruction. If
50+ the immediate is valid, write a constant suitable for using as an operand
51+ to VSHR/VSHL to *MODCONST and the corresponding element width to
52+ *ELEMENTWIDTH. ISLEFTSHIFT is for determine left or right shift,
53+ because they have different limitations. */
54+
55+int
56+neon_immediate_valid_for_shift (rtx op, enum machine_mode mode,
57+ rtx *modconst, int *elementwidth,
58+ bool isleftshift)
59+{
60+ unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode));
61+ unsigned int n_elts = CONST_VECTOR_NUNITS (op), i;
62+ unsigned HOST_WIDE_INT last_elt = 0;
63+ unsigned HOST_WIDE_INT maxshift;
64+
65+ /* Split vector constant out into a byte vector. */
66+ for (i = 0; i < n_elts; i++)
67+ {
68+ rtx el = CONST_VECTOR_ELT (op, i);
69+ unsigned HOST_WIDE_INT elpart;
70+
71+ if (GET_CODE (el) == CONST_INT)
72+ elpart = INTVAL (el);
73+ else if (GET_CODE (el) == CONST_DOUBLE)
74+ return 0;
75+ else
76+ gcc_unreachable ();
77+
78+ if (i != 0 && elpart != last_elt)
79+ return 0;
80+
81+ last_elt = elpart;
82+ }
83+
84+ /* Shift less than element size. */
85+ maxshift = innersize * 8;
86+
87+ if (isleftshift)
88+ {
89+ /* Left shift immediate value can be from 0 to <size>-1. */
90+ if (last_elt >= maxshift)
91+ return 0;
92+ }
93+ else
94+ {
95+ /* Right shift immediate value can be from 1 to <size>. */
96+ if (last_elt == 0 || last_elt > maxshift)
97+ return 0;
98+ }
99+
100+ if (elementwidth)
101+ *elementwidth = innersize * 8;
102+
103+ if (modconst)
104+ *modconst = CONST_VECTOR_ELT (op, 0);
105+
106+ return 1;
107+}
108+
109 /* Return a string suitable for output of Neon immediate logic operation
110 MNEM. */
111
112@@ -8885,6 +8945,28 @@
113 return templ;
114 }
115
116+/* Return a string suitable for output of Neon immediate shift operation
117+ (VSHR or VSHL) MNEM. */
118+
119+char *
120+neon_output_shift_immediate (const char *mnem, char sign, rtx *op2,
121+ enum machine_mode mode, int quad,
122+ bool isleftshift)
123+{
124+ int width, is_valid;
125+ static char templ[40];
126+
127+ is_valid = neon_immediate_valid_for_shift (*op2, mode, op2, &width, isleftshift);
128+ gcc_assert (is_valid != 0);
129+
130+ if (quad)
131+ sprintf (templ, "%s.%c%d\t%%q0, %%q1, %%2", mnem, sign, width);
132+ else
133+ sprintf (templ, "%s.%c%d\t%%P0, %%P1, %%2", mnem, sign, width);
134+
135+ return templ;
136+}
137+
138 /* Output a sequence of pairwise operations to implement a reduction.
139 NOTE: We do "too much work" here, because pairwise operations work on two
140 registers-worth of operands in one go. Unfortunately we can't exploit those
141
142=== modified file 'gcc/config/arm/neon.md'
143--- old/gcc/config/arm/neon.md 2011-07-01 09:19:21 +0000
144+++ new/gcc/config/arm/neon.md 2011-07-04 14:03:49 +0000
145@@ -956,15 +956,57 @@
146 ; SImode elements.
147
148 (define_insn "vashl<mode>3"
149- [(set (match_operand:VDQIW 0 "s_register_operand" "=w")
150- (ashift:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
151- (match_operand:VDQIW 2 "s_register_operand" "w")))]
152- "TARGET_NEON"
153- "vshl.<V_s_elem>\t%<V_reg>0, %<V_reg>1, %<V_reg>2"
154- [(set (attr "neon_type")
155- (if_then_else (ne (symbol_ref "<Is_d_reg>") (const_int 0))
156- (const_string "neon_vshl_ddd")
157- (const_string "neon_shift_3")))]
158+ [(set (match_operand:VDQIW 0 "s_register_operand" "=w,w")
159+ (ashift:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w,w")
160+ (match_operand:VDQIW 2 "imm_lshift_or_reg_neon" "w,Dn")))]
161+ "TARGET_NEON"
162+ {
163+ switch (which_alternative)
164+ {
165+ case 0: return "vshl.<V_s_elem>\t%<V_reg>0, %<V_reg>1, %<V_reg>2";
166+ case 1: return neon_output_shift_immediate ("vshl", 'i', &operands[2],
167+ <MODE>mode,
168+ VALID_NEON_QREG_MODE (<MODE>mode),
169+ true);
170+ default: gcc_unreachable ();
171+ }
172+ }
173+ [(set (attr "neon_type")
174+ (if_then_else (ne (symbol_ref "<Is_d_reg>") (const_int 0))
175+ (const_string "neon_vshl_ddd")
176+ (const_string "neon_shift_3")))]
177+)
178+
179+(define_insn "vashr<mode>3_imm"
180+ [(set (match_operand:VDQIW 0 "s_register_operand" "=w")
181+ (ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
182+ (match_operand:VDQIW 2 "imm_for_neon_rshift_operand" "Dn")))]
183+ "TARGET_NEON"
184+ {
185+ return neon_output_shift_immediate ("vshr", 's', &operands[2],
186+ <MODE>mode, VALID_NEON_QREG_MODE (<MODE>mode),
187+ false);
188+ }
189+ [(set (attr "neon_type")
190+ (if_then_else (ne (symbol_ref "<Is_d_reg>") (const_int 0))
191+ (const_string "neon_vshl_ddd")
192+ (const_string "neon_shift_3")))]
193+)
194+
195+(define_insn "vlshr<mode>3_imm"
196+ [(set (match_operand:VDQIW 0 "s_register_operand" "=w")
197+ (lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "w")
198+ (match_operand:VDQIW 2 "imm_for_neon_rshift_operand" "Dn")))]
199+ "TARGET_NEON"
200+ {
201+ return neon_output_shift_immediate ("vshr", 'u', &operands[2],
202+ <MODE>mode, VALID_NEON_QREG_MODE (<MODE>mode),
203+ false);
204+ }
205+ [(set (attr "neon_type")
206+ (if_then_else (ne (symbol_ref "<Is_d_reg>") (const_int 0))
207+ (const_string "neon_vshl_ddd")
208+ (const_string "neon_shift_3")))]
209 )
210
211 ; Used for implementing logical shift-right, which is a left-shift by a negative
212@@ -1004,28 +1046,34 @@
213 (define_expand "vashr<mode>3"
214 [(set (match_operand:VDQIW 0 "s_register_operand" "")
215 (ashiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "")
216- (match_operand:VDQIW 2 "s_register_operand" "")))]
217+ (match_operand:VDQIW 2 "imm_rshift_or_reg_neon" "")))]
218 "TARGET_NEON"
219 {
220 rtx neg = gen_reg_rtx (<MODE>mode);
221-
222- emit_insn (gen_neg<mode>2 (neg, operands[2]));
223- emit_insn (gen_ashl<mode>3_signed (operands[0], operands[1], neg));
224-
225+ if (REG_P (operands[2]))
226+ {
227+ emit_insn (gen_neg<mode>2 (neg, operands[2]));
228+ emit_insn (gen_ashl<mode>3_signed (operands[0], operands[1], neg));
229+ }
230+ else
231+ emit_insn (gen_vashr<mode>3_imm (operands[0], operands[1], operands[2]));
232 DONE;
233 })
234
235 (define_expand "vlshr<mode>3"
236 [(set (match_operand:VDQIW 0 "s_register_operand" "")
237 (lshiftrt:VDQIW (match_operand:VDQIW 1 "s_register_operand" "")
238- (match_operand:VDQIW 2 "s_register_operand" "")))]
239+ (match_operand:VDQIW 2 "imm_rshift_or_reg_neon" "")))]
240 "TARGET_NEON"
241 {
242 rtx neg = gen_reg_rtx (<MODE>mode);
243-
244- emit_insn (gen_neg<mode>2 (neg, operands[2]));
245- emit_insn (gen_ashl<mode>3_unsigned (operands[0], operands[1], neg));
246-
247+ if (REG_P (operands[2]))
248+ {
249+ emit_insn (gen_neg<mode>2 (neg, operands[2]));
250+ emit_insn (gen_ashl<mode>3_unsigned (operands[0], operands[1], neg));
251+ }
252+ else
253+ emit_insn (gen_vlshr<mode>3_imm (operands[0], operands[1], operands[2]));
254 DONE;
255 })
256
257
258=== modified file 'gcc/config/arm/predicates.md'
259--- old/gcc/config/arm/predicates.md 2011-06-22 15:50:23 +0000
260+++ new/gcc/config/arm/predicates.md 2011-07-04 14:03:49 +0000
261@@ -585,6 +585,26 @@
262 return neon_immediate_valid_for_move (op, mode, NULL, NULL);
263 })
264
265+(define_predicate "imm_for_neon_lshift_operand"
266+ (match_code "const_vector")
267+{
268+ return neon_immediate_valid_for_shift (op, mode, NULL, NULL, true);
269+})
270+
271+(define_predicate "imm_for_neon_rshift_operand"
272+ (match_code "const_vector")
273+{
274+ return neon_immediate_valid_for_shift (op, mode, NULL, NULL, false);
275+})
276+
277+(define_predicate "imm_lshift_or_reg_neon"
278+ (ior (match_operand 0 "s_register_operand")
279+ (match_operand 0 "imm_for_neon_lshift_operand")))
280+
281+(define_predicate "imm_rshift_or_reg_neon"
282+ (ior (match_operand 0 "s_register_operand")
283+ (match_operand 0 "imm_for_neon_rshift_operand")))
284+
285 (define_predicate "imm_for_neon_logic_operand"
286 (match_code "const_vector")
287 {
288
289=== modified file 'gcc/optabs.c'
290--- old/gcc/optabs.c 2011-03-04 10:27:10 +0000
291+++ new/gcc/optabs.c 2011-07-04 14:03:49 +0000
292@@ -6171,6 +6171,9 @@
293 init_optab (usashl_optab, US_ASHIFT);
294 init_optab (ashr_optab, ASHIFTRT);
295 init_optab (lshr_optab, LSHIFTRT);
296+ init_optabv (vashl_optab, ASHIFT);
297+ init_optabv (vashr_optab, ASHIFTRT);
298+ init_optabv (vlshr_optab, LSHIFTRT);
299 init_optab (rotl_optab, ROTATE);
300 init_optab (rotr_optab, ROTATERT);
301 init_optab (smin_optab, SMIN);
302
303=== added file 'gcc/testsuite/gcc.target/arm/neon-vlshr-imm-1.c'
304--- old/gcc/testsuite/gcc.target/arm/neon-vlshr-imm-1.c 1970-01-01 00:00:00 +0000
305+++ new/gcc/testsuite/gcc.target/arm/neon-vlshr-imm-1.c 2011-07-04 14:03:49 +0000
306@@ -0,0 +1,11 @@
307+/* { dg-do compile } */
308+/* { dg-require-effective-target arm_neon_ok } */
309+/* { dg-options "-O2 -mfpu=neon -mfloat-abi=softfp -ftree-vectorize" } */
310+/* { dg-final { scan-assembler "vshr\.u32.*#3" } } */
311+
312+/* Verify that VSHR immediate is used. */
313+void f1(int n, unsigned int x[], unsigned int y[]) {
314+ int i;
315+ for (i = 0; i < n; ++i)
316+ y[i] = x[i] >> 3;
317+}
318
319=== added file 'gcc/testsuite/gcc.target/arm/neon-vshl-imm-1.c'
320--- old/gcc/testsuite/gcc.target/arm/neon-vshl-imm-1.c 1970-01-01 00:00:00 +0000
321+++ new/gcc/testsuite/gcc.target/arm/neon-vshl-imm-1.c 2011-07-04 14:03:49 +0000
322@@ -0,0 +1,11 @@
323+/* { dg-do compile } */
324+/* { dg-require-effective-target arm_neon_ok } */
325+/* { dg-options "-O2 -mfpu=neon -mfloat-abi=softfp -ftree-vectorize" } */
326+/* { dg-final { scan-assembler "vshl\.i32.*#3" } } */
327+
328+/* Verify that VSHR immediate is used. */
329+void f1(int n, int x[], int y[]) {
330+ int i;
331+ for (i = 0; i < n; ++i)
332+ y[i] = x[i] << 3;
333+}
334
335=== added file 'gcc/testsuite/gcc.target/arm/neon-vshr-imm-1.c'
336--- old/gcc/testsuite/gcc.target/arm/neon-vshr-imm-1.c 1970-01-01 00:00:00 +0000
337+++ new/gcc/testsuite/gcc.target/arm/neon-vshr-imm-1.c 2011-07-04 14:03:49 +0000
338@@ -0,0 +1,11 @@
339+/* { dg-do compile } */
340+/* { dg-require-effective-target arm_neon_ok } */
341+/* { dg-options "-O2 -mfpu=neon -mfloat-abi=softfp -ftree-vectorize" } */
342+/* { dg-final { scan-assembler "vshr\.s32.*#3" } } */
343+
344+/* Verify that VSHR immediate is used. */
345+void f1(int n, int x[], int y[]) {
346+ int i;
347+ for (i = 0; i < n; ++i)
348+ y[i] = x[i] >> 3;
349+}
350
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106773.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106773.patch
new file mode 100644
index 000000000..de3f29e19
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106773.patch
@@ -0,0 +1,119 @@
12011-07-13 Richard Sandiford <richard.sandiford@linaro.org>
2
3 Backport from mainline:
4 gcc/
5 2011-07-07 Richard Sandiford <richard.sandiford@linaro.org>
6
7 * reload1.c (choose_reload_regs): Use mode sizes to check whether
8 an old reload register completely defines the required value.
9
10 gcc/testsuite/
11 2011-07-07 Richard Sandiford <richard.sandiford@linaro.org>
12
13 * gcc.target/arm/neon-modes-3.c: New test.
14
15=== modified file 'gcc/reload1.c'
16--- old/gcc/reload1.c 2011-07-01 09:19:21 +0000
17+++ new/gcc/reload1.c 2011-07-11 10:06:50 +0000
18@@ -6451,6 +6451,8 @@
19
20 if (regno >= 0
21 && reg_last_reload_reg[regno] != 0
22+ && (GET_MODE_SIZE (GET_MODE (reg_last_reload_reg[regno]))
23+ >= GET_MODE_SIZE (mode) + byte)
24 #ifdef CANNOT_CHANGE_MODE_CLASS
25 /* Verify that the register it's in can be used in
26 mode MODE. */
27@@ -6462,24 +6464,12 @@
28 {
29 enum reg_class rclass = rld[r].rclass, last_class;
30 rtx last_reg = reg_last_reload_reg[regno];
31- enum machine_mode need_mode;
32
33 i = REGNO (last_reg);
34 i += subreg_regno_offset (i, GET_MODE (last_reg), byte, mode);
35 last_class = REGNO_REG_CLASS (i);
36
37- if (byte == 0)
38- need_mode = mode;
39- else
40- need_mode
41- = smallest_mode_for_size
42- (GET_MODE_BITSIZE (mode) + byte * BITS_PER_UNIT,
43- GET_MODE_CLASS (mode) == MODE_PARTIAL_INT
44- ? MODE_INT : GET_MODE_CLASS (mode));
45-
46- if ((GET_MODE_SIZE (GET_MODE (last_reg))
47- >= GET_MODE_SIZE (need_mode))
48- && reg_reloaded_contents[i] == regno
49+ if (reg_reloaded_contents[i] == regno
50 && TEST_HARD_REG_BIT (reg_reloaded_valid, i)
51 && HARD_REGNO_MODE_OK (i, rld[r].mode)
52 && (TEST_HARD_REG_BIT (reg_class_contents[(int) rclass], i)
53
54=== added file 'gcc/testsuite/gcc.target/arm/neon-modes-3.c'
55--- old/gcc/testsuite/gcc.target/arm/neon-modes-3.c 1970-01-01 00:00:00 +0000
56+++ new/gcc/testsuite/gcc.target/arm/neon-modes-3.c 2011-07-11 10:06:50 +0000
57@@ -0,0 +1,61 @@
58+/* { dg-do compile } */
59+/* { dg-require-effective-target arm_neon_ok } */
60+/* { dg-options "-O" } */
61+/* { dg-add-options arm_neon } */
62+
63+#include <arm_neon.h>
64+
65+void f1 (volatile float32x4_t *dest, volatile float32x4x4_t *src, int n)
66+{
67+ float32x4x4_t a5, a6, a7, a8, a9;
68+ int i;
69+
70+ a5 = *src;
71+ a6 = *src;
72+ a7 = *src;
73+ a8 = *src;
74+ a9 = *src;
75+ while (n--)
76+ {
77+ for (i = 0; i < 8; i++)
78+ {
79+ float32x4x4_t a0, a1, a2, a3, a4;
80+
81+ a0 = *src;
82+ a1 = *src;
83+ a2 = *src;
84+ a3 = *src;
85+ a4 = *src;
86+ *src = a0;
87+ *dest = a0.val[0];
88+ *dest = a0.val[3];
89+ *src = a1;
90+ *dest = a1.val[0];
91+ *dest = a1.val[3];
92+ *src = a2;
93+ *dest = a2.val[0];
94+ *dest = a2.val[3];
95+ *src = a3;
96+ *dest = a3.val[0];
97+ *dest = a3.val[3];
98+ *src = a4;
99+ *dest = a4.val[0];
100+ *dest = a4.val[3];
101+ }
102+ *src = a5;
103+ *dest = a5.val[0];
104+ *dest = a5.val[3];
105+ *src = a6;
106+ *dest = a6.val[0];
107+ *dest = a6.val[3];
108+ *src = a7;
109+ *dest = a7.val[0];
110+ *dest = a7.val[3];
111+ *src = a8;
112+ *dest = a8.val[0];
113+ *dest = a8.val[3];
114+ *src = a9;
115+ *dest = a9.val[0];
116+ *dest = a9.val[3];
117+ }
118+}
119
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106775.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106775.patch
new file mode 100644
index 000000000..0b05c3824
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106775.patch
@@ -0,0 +1,67 @@
12011-07-15 Michael Hope <michael.hope@linaro.org>
2
3 gcc/
4 Backport from mainline:
5
6 2011-04-05 Eric Botcazou <ebotcazou@adacore.com>
7
8 * ifcvt.c (cond_exec_process_insns): Disallow converting a block
9 that contains the prologue.
10
11 gcc/testsuite/
12 Backport from mainline:
13
14 2011-04-01 Bernd Schmidt <bernds@codesourcery.com>
15
16 * gcc.c-torture/compile/20110401-1.c: New test.
17
18=== modified file 'gcc/ifcvt.c'
19--- old/gcc/ifcvt.c 2010-12-14 00:23:40 +0000
20+++ new/gcc/ifcvt.c 2011-07-11 04:02:28 +0000
21@@ -1,5 +1,6 @@
22 /* If-conversion support.
23- Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2010
24+ Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2010,
25+ 2011
26 Free Software Foundation, Inc.
27
28 This file is part of GCC.
29@@ -304,6 +305,10 @@
30
31 for (insn = start; ; insn = NEXT_INSN (insn))
32 {
33+ /* dwarf2out can't cope with conditional prologues. */
34+ if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END)
35+ return FALSE;
36+
37 if (NOTE_P (insn) || DEBUG_INSN_P (insn))
38 goto insn_done;
39
40
41=== added file 'gcc/testsuite/gcc.c-torture/compile/20110401-1.c'
42--- old/gcc/testsuite/gcc.c-torture/compile/20110401-1.c 1970-01-01 00:00:00 +0000
43+++ new/gcc/testsuite/gcc.c-torture/compile/20110401-1.c 2011-07-11 04:02:28 +0000
44@@ -0,0 +1,22 @@
45+void asn1_length_der (unsigned long int len, unsigned char *ans, int *ans_len)
46+{
47+ int k;
48+ unsigned char temp[4];
49+ if (len < 128) {
50+ if (ans != ((void *) 0))
51+ ans[0] = (unsigned char) len;
52+ *ans_len = 1;
53+ } else {
54+ k = 0;
55+ while (len) {
56+ temp[k++] = len & 0xFF;
57+ len = len >> 8;
58+ }
59+ *ans_len = k + 1;
60+ if (ans != ((void *) 0)) {
61+ ans[0] = ((unsigned char) k & 0x7F) + 128;
62+ while (k--)
63+ ans[*ans_len - 1 - k] = temp[k];
64+ }
65+ }
66+}
67
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106776.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106776.patch
new file mode 100644
index 000000000..3d4d5c504
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106776.patch
@@ -0,0 +1,46 @@
12011-07-15 Michael Hope <michael.hope@linaro.org>
2
3 gcc/
4 Backport from mainline:
5 2011-03-22 Eric Botcazou <ebotcazou@adacore.com>
6
7 * combine.c (simplify_set): Try harder to find the best CC mode when
8 simplifying a nested COMPARE on the RHS.
9
10=== modified file 'gcc/combine.c'
11--- old/gcc/combine.c 2011-05-27 14:31:18 +0000
12+++ new/gcc/combine.c 2011-07-11 03:52:31 +0000
13@@ -6287,10 +6287,18 @@
14 enum rtx_code new_code;
15 rtx op0, op1, tmp;
16 int other_changed = 0;
17+ rtx inner_compare = NULL_RTX;
18 enum machine_mode compare_mode = GET_MODE (dest);
19
20 if (GET_CODE (src) == COMPARE)
21- op0 = XEXP (src, 0), op1 = XEXP (src, 1);
22+ {
23+ op0 = XEXP (src, 0), op1 = XEXP (src, 1);
24+ if (GET_CODE (op0) == COMPARE && op1 == const0_rtx)
25+ {
26+ inner_compare = op0;
27+ op0 = XEXP (inner_compare, 0), op1 = XEXP (inner_compare, 1);
28+ }
29+ }
30 else
31 op0 = src, op1 = CONST0_RTX (GET_MODE (src));
32
33@@ -6332,6 +6340,12 @@
34 need to use a different CC mode here. */
35 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
36 compare_mode = GET_MODE (op0);
37+ else if (inner_compare
38+ && GET_MODE_CLASS (GET_MODE (inner_compare)) == MODE_CC
39+ && new_code == old_code
40+ && op0 == XEXP (inner_compare, 0)
41+ && op1 == XEXP (inner_compare, 1))
42+ compare_mode = GET_MODE (inner_compare);
43 else
44 compare_mode = SELECT_CC_MODE (new_code, op0, op1);
45
46
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106777.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106777.patch
new file mode 100644
index 000000000..68b682b3c
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106777.patch
@@ -0,0 +1,192 @@
12011-07-15 Michael Hope <michael.hope@linaro.org>
2
3 gcc/
4 Backport from mainline:
5 2011-06-29 Nathan Sidwell <nathan@codesourcery.com>
6
7 * config/arm/unwind-arm.c (enum __cxa_type_match_result): New.
8 (cxa_type_match): Correct declaration.
9 (__gnu_unwind_pr_common): Reconstruct
10 additional indirection when __cxa_type_match returns
11 succeeded_with_ptr_to_base.
12
13 libstdc++-v3/
14 Backport from mainline:
15
16 2011-06-29 Nathan Sidwell <nathan@codesourcery.com>
17
18 * libsupc++/eh_arm.c (__cxa_type_match): Construct address of
19 thrown object here. Return succeded_with_ptr_to_base for all
20 pointer cases.
21
22=== modified file 'gcc/config/arm/unwind-arm.c'
23--- old/gcc/config/arm/unwind-arm.c 2011-03-22 10:59:10 +0000
24+++ new/gcc/config/arm/unwind-arm.c 2011-07-11 03:35:44 +0000
25@@ -32,13 +32,18 @@
26 typedef unsigned char bool;
27
28 typedef struct _ZSt9type_info type_info; /* This names C++ type_info type */
29+enum __cxa_type_match_result
30+ {
31+ ctm_failed = 0,
32+ ctm_succeeded = 1,
33+ ctm_succeeded_with_ptr_to_base = 2
34+ };
35
36 void __attribute__((weak)) __cxa_call_unexpected(_Unwind_Control_Block *ucbp);
37 bool __attribute__((weak)) __cxa_begin_cleanup(_Unwind_Control_Block *ucbp);
38-bool __attribute__((weak)) __cxa_type_match(_Unwind_Control_Block *ucbp,
39- const type_info *rttip,
40- bool is_reference,
41- void **matched_object);
42+enum __cxa_type_match_result __attribute__((weak)) __cxa_type_match
43+ (_Unwind_Control_Block *ucbp, const type_info *rttip,
44+ bool is_reference, void **matched_object);
45
46 _Unwind_Ptr __attribute__((weak))
47 __gnu_Unwind_Find_exidx (_Unwind_Ptr, int *);
48@@ -1107,6 +1112,7 @@
49 _uw rtti;
50 bool is_reference = (data[0] & uint32_highbit) != 0;
51 void *matched;
52+ enum __cxa_type_match_result match_type;
53
54 /* Check for no-throw areas. */
55 if (data[1] == (_uw) -2)
56@@ -1118,17 +1124,31 @@
57 {
58 /* Match a catch specification. */
59 rtti = _Unwind_decode_target2 ((_uw) &data[1]);
60- if (!__cxa_type_match (ucbp, (type_info *) rtti,
61- is_reference,
62- &matched))
63- matched = (void *)0;
64+ match_type = __cxa_type_match (ucbp,
65+ (type_info *) rtti,
66+ is_reference,
67+ &matched);
68 }
69+ else
70+ match_type = ctm_succeeded;
71
72- if (matched)
73+ if (match_type)
74 {
75 ucbp->barrier_cache.sp =
76 _Unwind_GetGR (context, R_SP);
77- ucbp->barrier_cache.bitpattern[0] = (_uw) matched;
78+ // ctm_succeeded_with_ptr_to_base really
79+ // means _c_t_m indirected the pointer
80+ // object. We have to reconstruct the
81+ // additional pointer layer by using a temporary.
82+ if (match_type == ctm_succeeded_with_ptr_to_base)
83+ {
84+ ucbp->barrier_cache.bitpattern[2]
85+ = (_uw) matched;
86+ ucbp->barrier_cache.bitpattern[0]
87+ = (_uw) &ucbp->barrier_cache.bitpattern[2];
88+ }
89+ else
90+ ucbp->barrier_cache.bitpattern[0] = (_uw) matched;
91 ucbp->barrier_cache.bitpattern[1] = (_uw) data;
92 return _URC_HANDLER_FOUND;
93 }
94
95=== modified file 'libstdc++-v3/libsupc++/eh_arm.cc'
96--- old/libstdc++-v3/libsupc++/eh_arm.cc 2011-01-03 20:52:22 +0000
97+++ new/libstdc++-v3/libsupc++/eh_arm.cc 2011-07-11 03:35:44 +0000
98@@ -30,10 +30,11 @@
99 using namespace __cxxabiv1;
100
101
102-// Given the thrown type THROW_TYPE, pointer to a variable containing a
103-// pointer to the exception object THROWN_PTR_P and a type CATCH_TYPE to
104-// compare against, return whether or not there is a match and if so,
105-// update *THROWN_PTR_P.
106+// Given the thrown type THROW_TYPE, exception object UE_HEADER and a
107+// type CATCH_TYPE to compare against, return whether or not there is
108+// a match and if so, update *THROWN_PTR_P to point to either the
109+// type-matched object, or in the case of a pointer type, the object
110+// pointed to by the pointer.
111
112 extern "C" __cxa_type_match_result
113 __cxa_type_match(_Unwind_Exception* ue_header,
114@@ -41,51 +42,51 @@
115 bool is_reference __attribute__((__unused__)),
116 void** thrown_ptr_p)
117 {
118- bool forced_unwind = __is_gxx_forced_unwind_class(ue_header->exception_class);
119- bool foreign_exception = !forced_unwind && !__is_gxx_exception_class(ue_header->exception_class);
120- bool dependent_exception =
121- __is_dependent_exception(ue_header->exception_class);
122+ bool forced_unwind
123+ = __is_gxx_forced_unwind_class(ue_header->exception_class);
124+ bool foreign_exception
125+ = !forced_unwind && !__is_gxx_exception_class(ue_header->exception_class);
126+ bool dependent_exception
127+ = __is_dependent_exception(ue_header->exception_class);
128 __cxa_exception* xh = __get_exception_header_from_ue(ue_header);
129 __cxa_dependent_exception *dx = __get_dependent_exception_from_ue(ue_header);
130 const std::type_info* throw_type;
131+ void *thrown_ptr = 0;
132
133 if (forced_unwind)
134 throw_type = &typeid(abi::__forced_unwind);
135 else if (foreign_exception)
136 throw_type = &typeid(abi::__foreign_exception);
137- else if (dependent_exception)
138- throw_type = __get_exception_header_from_obj
139- (dx->primaryException)->exceptionType;
140 else
141- throw_type = xh->exceptionType;
142-
143- void* thrown_ptr = *thrown_ptr_p;
144+ {
145+ if (dependent_exception)
146+ xh = __get_exception_header_from_obj (dx->primaryException);
147+ throw_type = xh->exceptionType;
148+ // We used to require the caller set the target of thrown_ptr_p,
149+ // but that's incorrect -- the EHABI makes no such requirement
150+ // -- and not all callers will set it. Fortunately callers that
151+ // do initialize will always pass us the value we calculate
152+ // here, so there's no backwards compatibility problem.
153+ thrown_ptr = __get_object_from_ue (ue_header);
154+ }
155+
156+ __cxa_type_match_result result = ctm_succeeded;
157
158 // Pointer types need to adjust the actual pointer, not
159 // the pointer to pointer that is the exception object.
160 // This also has the effect of passing pointer types
161 // "by value" through the __cxa_begin_catch return value.
162 if (throw_type->__is_pointer_p())
163- thrown_ptr = *(void**) thrown_ptr;
164+ {
165+ thrown_ptr = *(void**) thrown_ptr;
166+ // We need to indicate the indirection to our caller.
167+ result = ctm_succeeded_with_ptr_to_base;
168+ }
169
170 if (catch_type->__do_catch(throw_type, &thrown_ptr, 1))
171 {
172 *thrown_ptr_p = thrown_ptr;
173-
174- if (typeid(*catch_type) == typeid (typeid(void*)))
175- {
176- const __pointer_type_info *catch_pointer_type =
177- static_cast<const __pointer_type_info *> (catch_type);
178- const __pointer_type_info *throw_pointer_type =
179- static_cast<const __pointer_type_info *> (throw_type);
180-
181- if (typeid (*catch_pointer_type->__pointee) != typeid (void)
182- && (*catch_pointer_type->__pointee !=
183- *throw_pointer_type->__pointee))
184- return ctm_succeeded_with_ptr_to_base;
185- }
186-
187- return ctm_succeeded;
188+ return result;
189 }
190
191 return ctm_failed;
192
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106778.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106778.patch
new file mode 100644
index 000000000..b42b425dc
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106778.patch
@@ -0,0 +1,225 @@
12011-07-15 Michael Hope <michael.hope@linaro.org>
2
3 Backport from mainline r174540
4 LP: #807573
5
6 gcc/
7 2011-06-01 Richard Sandiford <rdsandiford@googlemail.com>
8
9 PR rtl-optimization/48830
10 PR rtl-optimization/48808
11 PR rtl-optimization/48792
12 * reload.c (push_reload): Check contains_reg_of_mode.
13 * reload1.c (strip_paradoxical_subreg): New function.
14 (gen_reload_chain_without_interm_reg_p): Use it to handle
15 paradoxical subregs.
16 (emit_output_reload_insns, gen_reload): Likewise.
17
18 gcc/testsuite/
19 2011-06-01 Eric Botcazou <ebotcazou@adacore.com>
20 Hans-Peter Nilsson <hp@axis.com>
21
22 PR rtl-optimization/48830
23 * gcc.target/sparc/ultrasp12.c: New test.
24
25=== modified file 'gcc/reload.c'
26--- old/gcc/reload.c 2011-07-01 09:19:21 +0000
27+++ new/gcc/reload.c 2011-07-13 02:09:08 +0000
28@@ -1017,6 +1017,7 @@
29 #ifdef CANNOT_CHANGE_MODE_CLASS
30 && !CANNOT_CHANGE_MODE_CLASS (GET_MODE (SUBREG_REG (in)), inmode, rclass)
31 #endif
32+ && contains_reg_of_mode[(int) rclass][(int) GET_MODE (SUBREG_REG (in))]
33 && (CONSTANT_P (SUBREG_REG (in))
34 || GET_CODE (SUBREG_REG (in)) == PLUS
35 || strict_low
36@@ -1123,6 +1124,7 @@
37 #ifdef CANNOT_CHANGE_MODE_CLASS
38 && !CANNOT_CHANGE_MODE_CLASS (GET_MODE (SUBREG_REG (out)), outmode, rclass)
39 #endif
40+ && contains_reg_of_mode[(int) rclass][(int) GET_MODE (SUBREG_REG (out))]
41 && (CONSTANT_P (SUBREG_REG (out))
42 || strict_low
43 || (((REG_P (SUBREG_REG (out))
44
45=== modified file 'gcc/reload1.c'
46--- old/gcc/reload1.c 2011-07-11 10:06:50 +0000
47+++ new/gcc/reload1.c 2011-07-14 22:14:45 +0000
48@@ -4476,6 +4476,43 @@
49 }
50 }
51 }
52+
53+/* *OP_PTR and *OTHER_PTR are two operands to a conceptual reload.
54+ If *OP_PTR is a paradoxical subreg, try to remove that subreg
55+ and apply the corresponding narrowing subreg to *OTHER_PTR.
56+ Return true if the operands were changed, false otherwise. */
57+
58+static bool
59+strip_paradoxical_subreg (rtx *op_ptr, rtx *other_ptr)
60+{
61+ rtx op, inner, other, tem;
62+
63+ op = *op_ptr;
64+ if (GET_CODE (op) != SUBREG)
65+ return false;
66+
67+ inner = SUBREG_REG (op);
68+ if (GET_MODE_SIZE (GET_MODE (op)) <= GET_MODE_SIZE (GET_MODE (inner)))
69+ return false;
70+
71+ other = *other_ptr;
72+ tem = gen_lowpart_common (GET_MODE (inner), other);
73+ if (!tem)
74+ return false;
75+
76+ /* If the lowpart operation turned a hard register into a subreg,
77+ rather than simplifying it to another hard register, then the
78+ mode change cannot be properly represented. For example, OTHER
79+ might be valid in its current mode, but not in the new one. */
80+ if (GET_CODE (tem) == SUBREG
81+ && REG_P (other)
82+ && HARD_REGISTER_P (other))
83+ return false;
84+
85+ *op_ptr = inner;
86+ *other_ptr = tem;
87+ return true;
88+}
89
90 /* A subroutine of reload_as_needed. If INSN has a REG_EH_REGION note,
91 examine all of the reload insns between PREV and NEXT exclusive, and
92@@ -5556,7 +5593,7 @@
93 chain reloads or do need an intermediate hard registers. */
94 bool result = true;
95 int regno, n, code;
96- rtx out, in, tem, insn;
97+ rtx out, in, insn;
98 rtx last = get_last_insn ();
99
100 /* Make r2 a component of r1. */
101@@ -5575,11 +5612,7 @@
102
103 /* If IN is a paradoxical SUBREG, remove it and try to put the
104 opposite SUBREG on OUT. Likewise for a paradoxical SUBREG on OUT. */
105- if (GET_CODE (in) == SUBREG
106- && (GET_MODE_SIZE (GET_MODE (in))
107- > GET_MODE_SIZE (GET_MODE (SUBREG_REG (in))))
108- && (tem = gen_lowpart_common (GET_MODE (SUBREG_REG (in)), out)) != 0)
109- in = SUBREG_REG (in), out = tem;
110+ strip_paradoxical_subreg (&in, &out);
111
112 if (GET_CODE (in) == PLUS
113 && (REG_P (XEXP (in, 0))
114@@ -7571,7 +7604,6 @@
115 if (tertiary_icode != CODE_FOR_nothing)
116 {
117 rtx third_reloadreg = rld[tertiary_reload].reg_rtx;
118- rtx tem;
119
120 /* Copy primary reload reg to secondary reload reg.
121 (Note that these have been swapped above, then
122@@ -7580,13 +7612,7 @@
123 /* If REAL_OLD is a paradoxical SUBREG, remove it
124 and try to put the opposite SUBREG on
125 RELOADREG. */
126- if (GET_CODE (real_old) == SUBREG
127- && (GET_MODE_SIZE (GET_MODE (real_old))
128- > GET_MODE_SIZE (GET_MODE (SUBREG_REG (real_old))))
129- && 0 != (tem = gen_lowpart_common
130- (GET_MODE (SUBREG_REG (real_old)),
131- reloadreg)))
132- real_old = SUBREG_REG (real_old), reloadreg = tem;
133+ strip_paradoxical_subreg (&real_old, &reloadreg);
134
135 gen_reload (reloadreg, second_reloadreg,
136 rl->opnum, rl->when_needed);
137@@ -8402,16 +8428,8 @@
138
139 /* If IN is a paradoxical SUBREG, remove it and try to put the
140 opposite SUBREG on OUT. Likewise for a paradoxical SUBREG on OUT. */
141- if (GET_CODE (in) == SUBREG
142- && (GET_MODE_SIZE (GET_MODE (in))
143- > GET_MODE_SIZE (GET_MODE (SUBREG_REG (in))))
144- && (tem = gen_lowpart_common (GET_MODE (SUBREG_REG (in)), out)) != 0)
145- in = SUBREG_REG (in), out = tem;
146- else if (GET_CODE (out) == SUBREG
147- && (GET_MODE_SIZE (GET_MODE (out))
148- > GET_MODE_SIZE (GET_MODE (SUBREG_REG (out))))
149- && (tem = gen_lowpart_common (GET_MODE (SUBREG_REG (out)), in)) != 0)
150- out = SUBREG_REG (out), in = tem;
151+ if (!strip_paradoxical_subreg (&in, &out))
152+ strip_paradoxical_subreg (&out, &in);
153
154 /* How to do this reload can get quite tricky. Normally, we are being
155 asked to reload a simple operand, such as a MEM, a constant, or a pseudo
156
157=== added file 'gcc/testsuite/gcc.target/sparc/ultrasp12.c'
158--- old/gcc/testsuite/gcc.target/sparc/ultrasp12.c 1970-01-01 00:00:00 +0000
159+++ new/gcc/testsuite/gcc.target/sparc/ultrasp12.c 2011-07-13 02:09:08 +0000
160@@ -0,0 +1,64 @@
161+/* PR rtl-optimization/48830 */
162+/* Testcase by Hans-Peter Nilsson <hp@gcc.gnu.org> */
163+
164+/* { dg-do compile } */
165+/* { dg-require-effective-target lp64 } */
166+/* { dg-options "-O2 -mcpu=ultrasparc -mvis" } */
167+
168+typedef unsigned char uint8_t;
169+typedef unsigned int uint32_t;
170+typedef unsigned long int uint64_t;
171+typedef unsigned long int uintmax_t;
172+typedef unsigned char rc_vec_t __attribute__((__vector_size__(8)));
173+typedef short rc_svec_type_ __attribute__((__vector_size__(8)));
174+typedef unsigned char rc_vec4_type_ __attribute__((__vector_size__(4)));
175+
176+void
177+rc_stat_xsum_acc(const uint8_t *__restrict src1, int src1_dim,
178+ const uint8_t *__restrict src2, int src2_dim,
179+ int len, int height, uintmax_t sum[5])
180+{
181+ uint32_t s1 = 0;
182+ uint32_t s2 = 0;
183+ uintmax_t s11 = 0;
184+ uintmax_t s22 = 0;
185+ uintmax_t s12 = 0;
186+ int full = len / ((1024) < (1024) ? (1024) : (1024));
187+ int rem = len % ((1024) < (1024) ? (1024) : (1024));
188+ int rem1 = rem / 1;
189+ int y;
190+ unsigned int rc_gsr_scale_ __attribute__ ((__unused__)) = 7; unsigned int rc_gsr_align_ __attribute__ ((__unused__)) = 4; unsigned int rc_gsr_set_ __attribute__ ((__unused__)) = 0; register unsigned int rc_gsr_fakedep_ __attribute__ ((__unused__)) = 0; unsigned int rc_gsr_ldinit_ __attribute__ ((__unused__)) = 0;
191+ for (y = 0; y < height; y++) {
192+ rc_vec_t a1, a2, a11, a22, a12;
193+ int i1 = (y)*(src1_dim);
194+ int i2 = (y)*(src2_dim);
195+ int x;
196+ ((a1) = ((rc_vec_t) {0}));
197+ ((a2) = ((rc_vec_t) {0}));
198+ ((a11) = ((rc_vec_t) {0}));
199+ ((a22) = ((rc_vec_t) {0}));
200+ ((a12) = ((rc_vec_t) {0}));
201+ for (x = 0; x < full; x++) {
202+ int k;
203+ for (k = 0; k < ((1024) < (1024) ? (1024) : (1024)) /
204+ 1; k++)
205+ {
206+ do { rc_vec_t v1, v2; ((v1) = *(const rc_vec_t*)(&(src1)[i1])); ((v2) = *(const rc_vec_t*)(&(src2)[i2])); ((a1) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(__builtin_vis_pdist (v1, ((rc_vec_t) {0}), (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a1)).i)))).v)); ((a2) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(__builtin_vis_pdist (v2, ((rc_vec_t) {0}), (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a2)).i)))).v)); do { rc_vec_t s1_ = (v1); rc_vec_t s2_ = (v1); rc_vec_t accvin_ = (a11); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a11) = accvout_; } while (0); do { rc_vec_t s1_ = (v2); rc_vec_t s2_ = (v2); rc_vec_t accvin_ = (a22); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a22) = accvout_; } while (0); do { rc_vec_t s1_ = (v1); rc_vec_t s2_ = (v2); rc_vec_t accvin_ = (a12); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a12) = accvout_; } while (0); (i1) += 8; (i2) += 8; } while (0);
207+
208+ }
209+ do { uint32_t t1, t2, t11, t22, t12; ((t1) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a1)).i)); ((t2) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a2)).i)); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a11); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t11) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a22); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t22) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a12); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t12) = maclo_ + machi_ * 256; } while (0); ((a1) = ((rc_vec_t) {0})); ((a2) = ((rc_vec_t) {0})); ((a11) = ((rc_vec_t) {0})); ((a22) = ((rc_vec_t) {0})); ((a12) = ((rc_vec_t) {0})); (s1) += t1; (s2) += t2; (s11) += t11; (s22) += t22; (s12) += t12; } while (0);
210+ }
211+ for (x = 0; x < rem1; x++) {
212+ do { rc_vec_t v1, v2; ((v1) = *(const rc_vec_t*)(&(src1)[i1])); ((v2) = *(const rc_vec_t*)(&(src2)[i2])); ((a1) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(__builtin_vis_pdist (v1, ((rc_vec_t) {0}), (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a1)).i)))).v)); ((a2) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(__builtin_vis_pdist (v2, ((rc_vec_t) {0}), (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a2)).i)))).v)); do { rc_vec_t s1_ = (v1); rc_vec_t s2_ = (v1); rc_vec_t accvin_ = (a11); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a11) = accvout_; } while (0); do { rc_vec_t s1_ = (v2); rc_vec_t s2_ = (v2); rc_vec_t accvin_ = (a22); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a22) = accvout_; } while (0); do { rc_vec_t s1_ = (v1); rc_vec_t s2_ = (v2); rc_vec_t accvin_ = (a12); rc_vec_t s1lo7_, s1msb_, accvout_; uint32_t maclo_, machi_; rc_svec_type_ masklow_ = (rc_svec_type_){(255), (255), (255), (255)}; rc_svec_type_ s1msbhi_, s1msblo_, s1lo7hi_, s1lo7lo_; rc_svec_type_ s1msbdiv2hi_, s1msbdiv2lo_; rc_vec4_type_ s1lo7hi4_, s1lo7lo4_, s1msbhi4_, s1msblo4_; rc_vec4_type_ s1msbdiv2hi4_, s1msbdiv2lo4_, s2hi4_, s2lo4_; rc_vec4_type_ accvhi4_, accvlo4_; rc_svec_type_ mulhilo7_, mullolo7_, mulhimsbdiv2_, mullomsbdiv2_; rc_svec_type_ mulhi_, mullo_, mulhihi_, mullohi_; rc_svec_type_ mulhilo_, mullolo_; rc_vec4_type_ zero4_ = (((union { rc_vec4_type_ v; uint64_t i; })(uint64_t)(0)).v); rc_vec_t msb_ = (rc_vec_t){(0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80), (0x80)}; ((s1msb_) = (s1_) & (msb_)); ((s1lo7_) = (s1_) & (~msb_)); do { if (rc_gsr_ldinit_) { extern void rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(void); rc_mixing_GSR_setting_with_RC_VEC_LDINIT_(); } if (!__builtin_constant_p(rc_gsr_align_) || !__builtin_constant_p(2) || !rc_gsr_set_ || (unsigned) (rc_gsr_align_) != rc_gsr_align_ || (unsigned) (2) != rc_gsr_scale_) { rc_gsr_set_ = 1; rc_gsr_align_ = (rc_gsr_align_); rc_gsr_scale_ = (2); unsigned int val_ = (rc_gsr_scale_ << 3) | rc_gsr_align_; if (__builtin_constant_p (val_)) { __asm__("wr %%g0,%[gsrval],%%gsr\n" ";# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "i" (val_), "1" (rc_gsr_fakedep_)); } else { __asm__("wr %[gsrval],0,%%gsr" "\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_), [fakegsr] "=rm" (rc_gsr_fakedep_) : "0" (s1msb_), [gsrval] "r" (val_), "1" (rc_gsr_fakedep_)); } } else { __asm__("\n;# dep %[depvec] on fake GSR %[fakegsr]" : [depvec] "=brm" (s1msb_) : "0" (s1msb_), [fakegsr] "g" (rc_gsr_fakedep_)); } } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1msb_); (s1msbhi4_) = hl_.hilo_.hi_; (s1msblo4_) = hl_.hilo_.lo_; } while (0); s1msbhi_ = __builtin_vis_fexpand(s1msbhi4_); s1msblo_ = __builtin_vis_fexpand(s1msblo4_); s1msbdiv2hi4_ = __builtin_vis_fpack16(s1msbhi_); s1msbdiv2lo4_ = __builtin_vis_fpack16(s1msblo_); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s2_); (s2hi4_) = hl_.hilo_.hi_; (s2lo4_) = hl_.hilo_.lo_; } while (0); do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (s1lo7_); (s1lo7hi4_) = hl_.hilo_.hi_; (s1lo7lo4_) = hl_.hilo_.lo_; } while (0); s1msbdiv2hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2hi4_, zero4_); s1msbdiv2lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1msbdiv2lo4_, zero4_); s1lo7hi_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7hi4_, zero4_); s1lo7lo_ = (rc_svec_type_)__builtin_vis_fpmerge(s1lo7lo4_, zero4_); mulhilo7_ = __builtin_vis_fmul8x16(s2hi4_, s1lo7hi_); mullolo7_ = __builtin_vis_fmul8x16(s2lo4_, s1lo7lo_); mulhimsbdiv2_ = __builtin_vis_fmul8x16(s2hi4_, s1msbdiv2hi_); mullomsbdiv2_ = __builtin_vis_fmul8x16(s2lo4_, s1msbdiv2lo_); mulhi_ = mulhilo7_ + mulhimsbdiv2_ + mulhimsbdiv2_; mullo_ = mullolo7_ + mullomsbdiv2_ + mullomsbdiv2_; mulhihi_ = mulhi_ & ~masklow_; mulhilo_ = mulhi_ & masklow_; mullohi_ = mullo_ & ~masklow_; mullolo_ = mullo_ & masklow_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (accvin_); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); maclo_ = __builtin_vis_pdist ((rc_vec_t)mullolo_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i)); maclo_ = __builtin_vis_pdist ((rc_vec_t)mulhilo_, ((rc_vec_t) {0}), maclo_); machi_ = __builtin_vis_pdist ((rc_vec_t)mullohi_, ((rc_vec_t) {0}), (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i)); machi_ = __builtin_vis_pdist ((rc_vec_t)mulhihi_, ((rc_vec_t) {0}), machi_); do { typedef union { struct { rc_vec4_type_ hi_, lo_; } hilo_; rc_vec_t v_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) {{((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)machi_)).v)), ((((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)((uint32_t)maclo_)).v))}}; (accvout_) = hl_.v_; } while (0); __asm__("\n;# dep fake GSR %[fakegsr] on %[xdep]" : [fakegsr] "=brm" (rc_gsr_fakedep_) : [xdep] "brm" (accvout_), "0" (rc_gsr_fakedep_)); (a12) = accvout_; } while (0); (i1) += 8; (i2) += 8; } while (0);
213+ }
214+ do { uint32_t t1, t2, t11, t22, t12; ((t1) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a1)).i)); ((t2) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a2)).i)); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a11); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t11) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a22); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t22) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a12); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t12) = maclo_ + machi_ * 256; } while (0); ((a1) = ((rc_vec_t) {0})); ((a2) = ((rc_vec_t) {0})); ((a11) = ((rc_vec_t) {0})); ((a22) = ((rc_vec_t) {0})); ((a12) = ((rc_vec_t) {0})); (s1) += t1; (s2) += t2; (s11) += t11; (s22) += t22; (s12) += t12; } while (0);
215+
216+ do { uint32_t t1, t2, t11, t22, t12; ((t1) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a1)).i)); ((t2) = (((union { rc_vec_t v; uint64_t i; })(uint64_t)(a2)).i)); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a11); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t11) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a22); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t22) = maclo_ + machi_ * 256; } while (0); do { rc_vec4_type_ accvhi4_, accvlo4_; uint64_t machi_, maclo_; do { typedef union { rc_vec_t v_; struct { rc_vec4_type_ hi_, lo_; } hilo_; } RC_hl_type_; RC_hl_type_ hl_ = (RC_hl_type_) (a12); (accvhi4_) = hl_.hilo_.hi_; (accvlo4_) = hl_.hilo_.lo_; } while (0); machi_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvhi4_)).i); maclo_ = (((union { rc_vec4_type_ v; uint32_t i; })(uint32_t)(accvlo4_)).i); (t12) = maclo_ + machi_ * 256; } while (0); ((a1) = ((rc_vec_t) {0})); ((a2) = ((rc_vec_t) {0})); ((a11) = ((rc_vec_t) {0})); ((a22) = ((rc_vec_t) {0})); ((a12) = ((rc_vec_t) {0})); (s1) += t1; (s2) += t2; (s11) += t11; (s22) += t22; (s12) += t12; } while (0);
217+ }
218+ sum[0] = s1;
219+ sum[1] = s2;
220+ sum[2] = s11;
221+ sum[3] = s22;
222+ sum[4] = s12;
223+ ;
224+}
225
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106781.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106781.patch
new file mode 100644
index 000000000..a86ddfdec
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106781.patch
@@ -0,0 +1,741 @@
12011-07-21 Richard Sandiford <richard.sandiford@linaro.org>
2
3 gcc/
4 PR middle-end/49736
5 * expr.c (all_zeros_p): Undo bogus part of last change.
6
72011-07-21 Richard Sandiford <richard.sandiford@linaro.org>
8
9 Backport from mainline:
10 gcc/cp/
11 2011-07-13 Richard Sandiford <richard.sandiford@linaro.org>
12
13 * typeck2.c (split_nonconstant_init_1): Pass the initializer directly,
14 rather than a pointer to it. Return true if the whole of the value
15 was initialized by the generated statements. Use
16 complete_ctor_at_level_p instead of count_type_elements.
17
18 gcc/
19 2011-07-13 Richard Sandiford <richard.sandiford@linaro.org>
20
21 * tree.h (categorize_ctor_elements): Remove comment. Fix long line.
22 (count_type_elements): Delete.
23 (complete_ctor_at_level_p): Declare.
24 * expr.c (flexible_array_member_p): New function, split out from...
25 (count_type_elements): ...here. Make static. Replace allow_flexarr
26 parameter with for_ctor_p. When for_ctor_p is true, return the
27 number of elements that should appear in the top-level constructor,
28 otherwise return an estimate of the number of scalars.
29 (categorize_ctor_elements): Replace p_must_clear with p_complete.
30 (categorize_ctor_elements_1): Likewise. Use complete_ctor_at_level_p.
31 (complete_ctor_at_level_p): New function, borrowing union logic
32 from old categorize_ctor_elements_1.
33 (mostly_zeros_p): Return true if the constructor is not complete.
34 (all_zeros_p): Update call to categorize_ctor_elements.
35 * gimplify.c (gimplify_init_constructor): Update call to
36 categorize_ctor_elements. Don't call count_type_elements.
37 Unconditionally prevent clearing for variable-sized types,
38 otherwise rely on categorize_ctor_elements to detect
39 incomplete initializers.
40
41 gcc/testsuite/
42 2011-07-13 Chung-Lin Tang <cltang@codesourcery.com>
43
44 * gcc.target/arm/pr48183.c: New test.
45
46=== modified file 'gcc/cp/typeck2.c'
47--- old/gcc/cp/typeck2.c 2011-05-20 21:29:14 +0000
48+++ new/gcc/cp/typeck2.c 2011-07-13 13:17:31 +0000
49@@ -473,18 +473,20 @@
50
51
52 /* The recursive part of split_nonconstant_init. DEST is an lvalue
53- expression to which INIT should be assigned. INIT is a CONSTRUCTOR. */
54+ expression to which INIT should be assigned. INIT is a CONSTRUCTOR.
55+ Return true if the whole of the value was initialized by the
56+ generated statements. */
57
58-static void
59-split_nonconstant_init_1 (tree dest, tree *initp)
60+static bool
61+split_nonconstant_init_1 (tree dest, tree init)
62 {
63 unsigned HOST_WIDE_INT idx;
64- tree init = *initp;
65 tree field_index, value;
66 tree type = TREE_TYPE (dest);
67 tree inner_type = NULL;
68 bool array_type_p = false;
69- HOST_WIDE_INT num_type_elements, num_initialized_elements;
70+ bool complete_p = true;
71+ HOST_WIDE_INT num_split_elts = 0;
72
73 switch (TREE_CODE (type))
74 {
75@@ -496,7 +498,6 @@
76 case RECORD_TYPE:
77 case UNION_TYPE:
78 case QUAL_UNION_TYPE:
79- num_initialized_elements = 0;
80 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (init), idx,
81 field_index, value)
82 {
83@@ -519,13 +520,14 @@
84 sub = build3 (COMPONENT_REF, inner_type, dest, field_index,
85 NULL_TREE);
86
87- split_nonconstant_init_1 (sub, &value);
88+ if (!split_nonconstant_init_1 (sub, value))
89+ complete_p = false;
90+ num_split_elts++;
91 }
92 else if (!initializer_constant_valid_p (value, inner_type))
93 {
94 tree code;
95 tree sub;
96- HOST_WIDE_INT inner_elements;
97
98 /* FIXME: Ordered removal is O(1) so the whole function is
99 worst-case quadratic. This could be fixed using an aside
100@@ -549,21 +551,9 @@
101 code = build_stmt (input_location, EXPR_STMT, code);
102 add_stmt (code);
103
104- inner_elements = count_type_elements (inner_type, true);
105- if (inner_elements < 0)
106- num_initialized_elements = -1;
107- else if (num_initialized_elements >= 0)
108- num_initialized_elements += inner_elements;
109- continue;
110+ num_split_elts++;
111 }
112 }
113-
114- num_type_elements = count_type_elements (type, true);
115- /* If all elements of the initializer are non-constant and
116- have been split out, we don't need the empty CONSTRUCTOR. */
117- if (num_type_elements > 0
118- && num_type_elements == num_initialized_elements)
119- *initp = NULL;
120 break;
121
122 case VECTOR_TYPE:
123@@ -575,6 +565,7 @@
124 code = build2 (MODIFY_EXPR, type, dest, cons);
125 code = build_stmt (input_location, EXPR_STMT, code);
126 add_stmt (code);
127+ num_split_elts += CONSTRUCTOR_NELTS (init);
128 }
129 break;
130
131@@ -584,6 +575,8 @@
132
133 /* The rest of the initializer is now a constant. */
134 TREE_CONSTANT (init) = 1;
135+ return complete_p && complete_ctor_at_level_p (TREE_TYPE (init),
136+ num_split_elts, inner_type);
137 }
138
139 /* A subroutine of store_init_value. Splits non-constant static
140@@ -599,7 +592,8 @@
141 if (TREE_CODE (init) == CONSTRUCTOR)
142 {
143 code = push_stmt_list ();
144- split_nonconstant_init_1 (dest, &init);
145+ if (split_nonconstant_init_1 (dest, init))
146+ init = NULL_TREE;
147 code = pop_stmt_list (code);
148 DECL_INITIAL (dest) = init;
149 TREE_READONLY (dest) = 0;
150
151=== modified file 'gcc/expr.c'
152--- old/gcc/expr.c 2011-06-02 12:12:00 +0000
153+++ new/gcc/expr.c 2011-07-14 11:52:32 +0000
154@@ -4866,16 +4866,136 @@
155 return NULL_RTX;
156 }
157
158+/* Return true if field F of structure TYPE is a flexible array. */
159+
160+static bool
161+flexible_array_member_p (const_tree f, const_tree type)
162+{
163+ const_tree tf;
164+
165+ tf = TREE_TYPE (f);
166+ return (DECL_CHAIN (f) == NULL
167+ && TREE_CODE (tf) == ARRAY_TYPE
168+ && TYPE_DOMAIN (tf)
169+ && TYPE_MIN_VALUE (TYPE_DOMAIN (tf))
170+ && integer_zerop (TYPE_MIN_VALUE (TYPE_DOMAIN (tf)))
171+ && !TYPE_MAX_VALUE (TYPE_DOMAIN (tf))
172+ && int_size_in_bytes (type) >= 0);
173+}
174+
175+/* If FOR_CTOR_P, return the number of top-level elements that a constructor
176+ must have in order for it to completely initialize a value of type TYPE.
177+ Return -1 if the number isn't known.
178+
179+ If !FOR_CTOR_P, return an estimate of the number of scalars in TYPE. */
180+
181+static HOST_WIDE_INT
182+count_type_elements (const_tree type, bool for_ctor_p)
183+{
184+ switch (TREE_CODE (type))
185+ {
186+ case ARRAY_TYPE:
187+ {
188+ tree nelts;
189+
190+ nelts = array_type_nelts (type);
191+ if (nelts && host_integerp (nelts, 1))
192+ {
193+ unsigned HOST_WIDE_INT n;
194+
195+ n = tree_low_cst (nelts, 1) + 1;
196+ if (n == 0 || for_ctor_p)
197+ return n;
198+ else
199+ return n * count_type_elements (TREE_TYPE (type), false);
200+ }
201+ return for_ctor_p ? -1 : 1;
202+ }
203+
204+ case RECORD_TYPE:
205+ {
206+ unsigned HOST_WIDE_INT n;
207+ tree f;
208+
209+ n = 0;
210+ for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f))
211+ if (TREE_CODE (f) == FIELD_DECL)
212+ {
213+ if (!for_ctor_p)
214+ n += count_type_elements (TREE_TYPE (f), false);
215+ else if (!flexible_array_member_p (f, type))
216+ /* Don't count flexible arrays, which are not supposed
217+ to be initialized. */
218+ n += 1;
219+ }
220+
221+ return n;
222+ }
223+
224+ case UNION_TYPE:
225+ case QUAL_UNION_TYPE:
226+ {
227+ tree f;
228+ HOST_WIDE_INT n, m;
229+
230+ gcc_assert (!for_ctor_p);
231+ /* Estimate the number of scalars in each field and pick the
232+ maximum. Other estimates would do instead; the idea is simply
233+ to make sure that the estimate is not sensitive to the ordering
234+ of the fields. */
235+ n = 1;
236+ for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f))
237+ if (TREE_CODE (f) == FIELD_DECL)
238+ {
239+ m = count_type_elements (TREE_TYPE (f), false);
240+ /* If the field doesn't span the whole union, add an extra
241+ scalar for the rest. */
242+ if (simple_cst_equal (TYPE_SIZE (TREE_TYPE (f)),
243+ TYPE_SIZE (type)) != 1)
244+ m++;
245+ if (n < m)
246+ n = m;
247+ }
248+ return n;
249+ }
250+
251+ case COMPLEX_TYPE:
252+ return 2;
253+
254+ case VECTOR_TYPE:
255+ return TYPE_VECTOR_SUBPARTS (type);
256+
257+ case INTEGER_TYPE:
258+ case REAL_TYPE:
259+ case FIXED_POINT_TYPE:
260+ case ENUMERAL_TYPE:
261+ case BOOLEAN_TYPE:
262+ case POINTER_TYPE:
263+ case OFFSET_TYPE:
264+ case REFERENCE_TYPE:
265+ return 1;
266+
267+ case ERROR_MARK:
268+ return 0;
269+
270+ case VOID_TYPE:
271+ case METHOD_TYPE:
272+ case FUNCTION_TYPE:
273+ case LANG_TYPE:
274+ default:
275+ gcc_unreachable ();
276+ }
277+}
278+
279 /* Helper for categorize_ctor_elements. Identical interface. */
280
281 static bool
282 categorize_ctor_elements_1 (const_tree ctor, HOST_WIDE_INT *p_nz_elts,
283- HOST_WIDE_INT *p_elt_count,
284- bool *p_must_clear)
285+ HOST_WIDE_INT *p_init_elts, bool *p_complete)
286 {
287 unsigned HOST_WIDE_INT idx;
288- HOST_WIDE_INT nz_elts, elt_count;
289- tree value, purpose;
290+ HOST_WIDE_INT nz_elts, init_elts, num_fields;
291+ tree value, purpose, elt_type;
292
293 /* Whether CTOR is a valid constant initializer, in accordance with what
294 initializer_constant_valid_p does. If inferred from the constructor
295@@ -4884,7 +5004,9 @@
296 bool const_p = const_from_elts_p ? true : TREE_STATIC (ctor);
297
298 nz_elts = 0;
299- elt_count = 0;
300+ init_elts = 0;
301+ num_fields = 0;
302+ elt_type = NULL_TREE;
303
304 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (ctor), idx, purpose, value)
305 {
306@@ -4899,6 +5021,8 @@
307 mult = (tree_low_cst (hi_index, 1)
308 - tree_low_cst (lo_index, 1) + 1);
309 }
310+ num_fields += mult;
311+ elt_type = TREE_TYPE (value);
312
313 switch (TREE_CODE (value))
314 {
315@@ -4906,11 +5030,11 @@
316 {
317 HOST_WIDE_INT nz = 0, ic = 0;
318
319- bool const_elt_p
320- = categorize_ctor_elements_1 (value, &nz, &ic, p_must_clear);
321+ bool const_elt_p = categorize_ctor_elements_1 (value, &nz, &ic,
322+ p_complete);
323
324 nz_elts += mult * nz;
325- elt_count += mult * ic;
326+ init_elts += mult * ic;
327
328 if (const_from_elts_p && const_p)
329 const_p = const_elt_p;
330@@ -4922,12 +5046,12 @@
331 case FIXED_CST:
332 if (!initializer_zerop (value))
333 nz_elts += mult;
334- elt_count += mult;
335+ init_elts += mult;
336 break;
337
338 case STRING_CST:
339 nz_elts += mult * TREE_STRING_LENGTH (value);
340- elt_count += mult * TREE_STRING_LENGTH (value);
341+ init_elts += mult * TREE_STRING_LENGTH (value);
342 break;
343
344 case COMPLEX_CST:
345@@ -4935,7 +5059,7 @@
346 nz_elts += mult;
347 if (!initializer_zerop (TREE_IMAGPART (value)))
348 nz_elts += mult;
349- elt_count += mult;
350+ init_elts += mult;
351 break;
352
353 case VECTOR_CST:
354@@ -4945,65 +5069,31 @@
355 {
356 if (!initializer_zerop (TREE_VALUE (v)))
357 nz_elts += mult;
358- elt_count += mult;
359+ init_elts += mult;
360 }
361 }
362 break;
363
364 default:
365 {
366- HOST_WIDE_INT tc = count_type_elements (TREE_TYPE (value), true);
367- if (tc < 1)
368- tc = 1;
369+ HOST_WIDE_INT tc = count_type_elements (elt_type, false);
370 nz_elts += mult * tc;
371- elt_count += mult * tc;
372+ init_elts += mult * tc;
373
374 if (const_from_elts_p && const_p)
375- const_p = initializer_constant_valid_p (value, TREE_TYPE (value))
376+ const_p = initializer_constant_valid_p (value, elt_type)
377 != NULL_TREE;
378 }
379 break;
380 }
381 }
382
383- if (!*p_must_clear
384- && (TREE_CODE (TREE_TYPE (ctor)) == UNION_TYPE
385- || TREE_CODE (TREE_TYPE (ctor)) == QUAL_UNION_TYPE))
386- {
387- tree init_sub_type;
388- bool clear_this = true;
389-
390- if (!VEC_empty (constructor_elt, CONSTRUCTOR_ELTS (ctor)))
391- {
392- /* We don't expect more than one element of the union to be
393- initialized. Not sure what we should do otherwise... */
394- gcc_assert (VEC_length (constructor_elt, CONSTRUCTOR_ELTS (ctor))
395- == 1);
396-
397- init_sub_type = TREE_TYPE (VEC_index (constructor_elt,
398- CONSTRUCTOR_ELTS (ctor),
399- 0)->value);
400-
401- /* ??? We could look at each element of the union, and find the
402- largest element. Which would avoid comparing the size of the
403- initialized element against any tail padding in the union.
404- Doesn't seem worth the effort... */
405- if (simple_cst_equal (TYPE_SIZE (TREE_TYPE (ctor)),
406- TYPE_SIZE (init_sub_type)) == 1)
407- {
408- /* And now we have to find out if the element itself is fully
409- constructed. E.g. for union { struct { int a, b; } s; } u
410- = { .s = { .a = 1 } }. */
411- if (elt_count == count_type_elements (init_sub_type, false))
412- clear_this = false;
413- }
414- }
415-
416- *p_must_clear = clear_this;
417- }
418+ if (*p_complete && !complete_ctor_at_level_p (TREE_TYPE (ctor),
419+ num_fields, elt_type))
420+ *p_complete = false;
421
422 *p_nz_elts += nz_elts;
423- *p_elt_count += elt_count;
424+ *p_init_elts += init_elts;
425
426 return const_p;
427 }
428@@ -5013,111 +5103,50 @@
429 and place it in *P_NZ_ELTS;
430 * how many scalar fields in total are in CTOR,
431 and place it in *P_ELT_COUNT.
432- * if a type is a union, and the initializer from the constructor
433- is not the largest element in the union, then set *p_must_clear.
434+ * whether the constructor is complete -- in the sense that every
435+ meaningful byte is explicitly given a value --
436+ and place it in *P_COMPLETE.
437
438 Return whether or not CTOR is a valid static constant initializer, the same
439 as "initializer_constant_valid_p (CTOR, TREE_TYPE (CTOR)) != 0". */
440
441 bool
442 categorize_ctor_elements (const_tree ctor, HOST_WIDE_INT *p_nz_elts,
443- HOST_WIDE_INT *p_elt_count,
444- bool *p_must_clear)
445+ HOST_WIDE_INT *p_init_elts, bool *p_complete)
446 {
447 *p_nz_elts = 0;
448- *p_elt_count = 0;
449- *p_must_clear = false;
450+ *p_init_elts = 0;
451+ *p_complete = true;
452
453- return
454- categorize_ctor_elements_1 (ctor, p_nz_elts, p_elt_count, p_must_clear);
455+ return categorize_ctor_elements_1 (ctor, p_nz_elts, p_init_elts, p_complete);
456 }
457
458-/* Count the number of scalars in TYPE. Return -1 on overflow or
459- variable-sized. If ALLOW_FLEXARR is true, don't count flexible
460- array member at the end of the structure. */
461+/* TYPE is initialized by a constructor with NUM_ELTS elements, the last
462+ of which had type LAST_TYPE. Each element was itself a complete
463+ initializer, in the sense that every meaningful byte was explicitly
464+ given a value. Return true if the same is true for the constructor
465+ as a whole. */
466
467-HOST_WIDE_INT
468-count_type_elements (const_tree type, bool allow_flexarr)
469+bool
470+complete_ctor_at_level_p (const_tree type, HOST_WIDE_INT num_elts,
471+ const_tree last_type)
472 {
473- const HOST_WIDE_INT max = ~((HOST_WIDE_INT)1 << (HOST_BITS_PER_WIDE_INT-1));
474- switch (TREE_CODE (type))
475+ if (TREE_CODE (type) == UNION_TYPE
476+ || TREE_CODE (type) == QUAL_UNION_TYPE)
477 {
478- case ARRAY_TYPE:
479- {
480- tree telts = array_type_nelts (type);
481- if (telts && host_integerp (telts, 1))
482- {
483- HOST_WIDE_INT n = tree_low_cst (telts, 1) + 1;
484- HOST_WIDE_INT m = count_type_elements (TREE_TYPE (type), false);
485- if (n == 0)
486- return 0;
487- else if (max / n > m)
488- return n * m;
489- }
490- return -1;
491- }
492-
493- case RECORD_TYPE:
494- {
495- HOST_WIDE_INT n = 0, t;
496- tree f;
497-
498- for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f))
499- if (TREE_CODE (f) == FIELD_DECL)
500- {
501- t = count_type_elements (TREE_TYPE (f), false);
502- if (t < 0)
503- {
504- /* Check for structures with flexible array member. */
505- tree tf = TREE_TYPE (f);
506- if (allow_flexarr
507- && DECL_CHAIN (f) == NULL
508- && TREE_CODE (tf) == ARRAY_TYPE
509- && TYPE_DOMAIN (tf)
510- && TYPE_MIN_VALUE (TYPE_DOMAIN (tf))
511- && integer_zerop (TYPE_MIN_VALUE (TYPE_DOMAIN (tf)))
512- && !TYPE_MAX_VALUE (TYPE_DOMAIN (tf))
513- && int_size_in_bytes (type) >= 0)
514- break;
515-
516- return -1;
517- }
518- n += t;
519- }
520-
521- return n;
522- }
523-
524- case UNION_TYPE:
525- case QUAL_UNION_TYPE:
526- return -1;
527-
528- case COMPLEX_TYPE:
529- return 2;
530-
531- case VECTOR_TYPE:
532- return TYPE_VECTOR_SUBPARTS (type);
533-
534- case INTEGER_TYPE:
535- case REAL_TYPE:
536- case FIXED_POINT_TYPE:
537- case ENUMERAL_TYPE:
538- case BOOLEAN_TYPE:
539- case POINTER_TYPE:
540- case OFFSET_TYPE:
541- case REFERENCE_TYPE:
542- return 1;
543-
544- case ERROR_MARK:
545- return 0;
546-
547- case VOID_TYPE:
548- case METHOD_TYPE:
549- case FUNCTION_TYPE:
550- case LANG_TYPE:
551- default:
552- gcc_unreachable ();
553+ if (num_elts == 0)
554+ return false;
555+
556+ gcc_assert (num_elts == 1 && last_type);
557+
558+ /* ??? We could look at each element of the union, and find the
559+ largest element. Which would avoid comparing the size of the
560+ initialized element against any tail padding in the union.
561+ Doesn't seem worth the effort... */
562+ return simple_cst_equal (TYPE_SIZE (type), TYPE_SIZE (last_type)) == 1;
563 }
564+
565+ return count_type_elements (type, true) == num_elts;
566 }
567
568 /* Return 1 if EXP contains mostly (3/4) zeros. */
569@@ -5126,18 +5155,12 @@
570 mostly_zeros_p (const_tree exp)
571 {
572 if (TREE_CODE (exp) == CONSTRUCTOR)
573-
574 {
575- HOST_WIDE_INT nz_elts, count, elts;
576- bool must_clear;
577-
578- categorize_ctor_elements (exp, &nz_elts, &count, &must_clear);
579- if (must_clear)
580- return 1;
581-
582- elts = count_type_elements (TREE_TYPE (exp), false);
583-
584- return nz_elts < elts / 4;
585+ HOST_WIDE_INT nz_elts, init_elts;
586+ bool complete_p;
587+
588+ categorize_ctor_elements (exp, &nz_elts, &init_elts, &complete_p);
589+ return !complete_p || nz_elts < init_elts / 4;
590 }
591
592 return initializer_zerop (exp);
593@@ -5149,12 +5172,11 @@
594 all_zeros_p (const_tree exp)
595 {
596 if (TREE_CODE (exp) == CONSTRUCTOR)
597-
598 {
599- HOST_WIDE_INT nz_elts, count;
600- bool must_clear;
601+ HOST_WIDE_INT nz_elts, init_elts;
602+ bool complete_p;
603
604- categorize_ctor_elements (exp, &nz_elts, &count, &must_clear);
605+ categorize_ctor_elements (exp, &nz_elts, &init_elts, &complete_p);
606 return nz_elts == 0;
607 }
608
609
610=== modified file 'gcc/gimplify.c'
611--- old/gcc/gimplify.c 2011-05-26 10:27:57 +0000
612+++ new/gcc/gimplify.c 2011-07-13 13:17:31 +0000
613@@ -3693,9 +3693,8 @@
614 case ARRAY_TYPE:
615 {
616 struct gimplify_init_ctor_preeval_data preeval_data;
617- HOST_WIDE_INT num_type_elements, num_ctor_elements;
618- HOST_WIDE_INT num_nonzero_elements;
619- bool cleared, valid_const_initializer;
620+ HOST_WIDE_INT num_ctor_elements, num_nonzero_elements;
621+ bool cleared, complete_p, valid_const_initializer;
622
623 /* Aggregate types must lower constructors to initialization of
624 individual elements. The exception is that a CONSTRUCTOR node
625@@ -3712,7 +3711,7 @@
626 can only do so if it known to be a valid constant initializer. */
627 valid_const_initializer
628 = categorize_ctor_elements (ctor, &num_nonzero_elements,
629- &num_ctor_elements, &cleared);
630+ &num_ctor_elements, &complete_p);
631
632 /* If a const aggregate variable is being initialized, then it
633 should never be a lose to promote the variable to be static. */
634@@ -3750,26 +3749,29 @@
635 parts in, then generate code for the non-constant parts. */
636 /* TODO. There's code in cp/typeck.c to do this. */
637
638- num_type_elements = count_type_elements (type, true);
639+ if (int_size_in_bytes (TREE_TYPE (ctor)) < 0)
640+ /* store_constructor will ignore the clearing of variable-sized
641+ objects. Initializers for such objects must explicitly set
642+ every field that needs to be set. */
643+ cleared = false;
644+ else if (!complete_p)
645+ /* If the constructor isn't complete, clear the whole object
646+ beforehand.
647
648- /* If count_type_elements could not determine number of type elements
649- for a constant-sized object, assume clearing is needed.
650- Don't do this for variable-sized objects, as store_constructor
651- will ignore the clearing of variable-sized objects. */
652- if (num_type_elements < 0 && int_size_in_bytes (type) >= 0)
653+ ??? This ought not to be needed. For any element not present
654+ in the initializer, we should simply set them to zero. Except
655+ we'd need to *find* the elements that are not present, and that
656+ requires trickery to avoid quadratic compile-time behavior in
657+ large cases or excessive memory use in small cases. */
658 cleared = true;
659- /* If there are "lots" of zeros, then block clear the object first. */
660- else if (num_type_elements - num_nonzero_elements
661+ else if (num_ctor_elements - num_nonzero_elements
662 > CLEAR_RATIO (optimize_function_for_speed_p (cfun))
663- && num_nonzero_elements < num_type_elements/4)
664- cleared = true;
665- /* ??? This bit ought not be needed. For any element not present
666- in the initializer, we should simply set them to zero. Except
667- we'd need to *find* the elements that are not present, and that
668- requires trickery to avoid quadratic compile-time behavior in
669- large cases or excessive memory use in small cases. */
670- else if (num_ctor_elements < num_type_elements)
671- cleared = true;
672+ && num_nonzero_elements < num_ctor_elements / 4)
673+ /* If there are "lots" of zeros, it's more efficient to clear
674+ the memory and then set the nonzero elements. */
675+ cleared = true;
676+ else
677+ cleared = false;
678
679 /* If there are "lots" of initialized elements, and all of them
680 are valid address constants, then the entire initializer can
681
682=== added file 'gcc/testsuite/gcc.target/arm/pr48183.c'
683--- old/gcc/testsuite/gcc.target/arm/pr48183.c 1970-01-01 00:00:00 +0000
684+++ new/gcc/testsuite/gcc.target/arm/pr48183.c 2011-07-13 13:17:31 +0000
685@@ -0,0 +1,25 @@
686+/* testsuite/gcc.target/arm/pr48183.c */
687+
688+/* { dg-do compile } */
689+/* { dg-require-effective-target arm_neon_ok } */
690+/* { dg-options "-O -g" } */
691+/* { dg-add-options arm_neon } */
692+
693+#include <arm_neon.h>
694+
695+void move_16bit_to_32bit (int32_t *dst, const short *src, unsigned n)
696+{
697+ unsigned i;
698+ int16x4x2_t input;
699+ int32x4x2_t mid;
700+ int32x4x2_t output;
701+
702+ for (i = 0; i < n/2; i += 8) {
703+ input = vld2_s16(src + i);
704+ mid.val[0] = vmovl_s16(input.val[0]);
705+ mid.val[1] = vmovl_s16(input.val[1]);
706+ output.val[0] = vshlq_n_s32(mid.val[0], 8);
707+ output.val[1] = vshlq_n_s32(mid.val[1], 8);
708+ vst2q_s32((int32_t *)dst + i, output);
709+ }
710+}
711
712=== modified file 'gcc/tree.h'
713--- old/gcc/tree.h 2011-07-01 09:19:21 +0000
714+++ new/gcc/tree.h 2011-07-13 13:17:31 +0000
715@@ -4627,21 +4627,10 @@
716
717 extern VEC(tree,gc) *ctor_to_vec (tree);
718
719-/* Examine CTOR to discover:
720- * how many scalar fields are set to nonzero values,
721- and place it in *P_NZ_ELTS;
722- * how many scalar fields in total are in CTOR,
723- and place it in *P_ELT_COUNT.
724- * if a type is a union, and the initializer from the constructor
725- is not the largest element in the union, then set *p_must_clear.
726-
727- Return whether or not CTOR is a valid static constant initializer, the same
728- as "initializer_constant_valid_p (CTOR, TREE_TYPE (CTOR)) != 0". */
729-
730-extern bool categorize_ctor_elements (const_tree, HOST_WIDE_INT *, HOST_WIDE_INT *,
731- bool *);
732-
733-extern HOST_WIDE_INT count_type_elements (const_tree, bool);
734+extern bool categorize_ctor_elements (const_tree, HOST_WIDE_INT *,
735+ HOST_WIDE_INT *, bool *);
736+
737+extern bool complete_ctor_at_level_p (const_tree, HOST_WIDE_INT, const_tree);
738
739 /* integer_zerop (tree x) is nonzero if X is an integer constant of value 0. */
740
741
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106782.patch b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106782.patch
new file mode 100644
index 000000000..5335a9e37
--- /dev/null
+++ b/meta-oe/recipes-devtools/gcc/gcc-4.6/linaro/gcc-4.6-linaro-r106782.patch
@@ -0,0 +1,27 @@
12011-07-21 Richard Sandiford <rdsandiford@googlemail.com>
2
3 gcc/
4 Backport from mainline:
5
6 2011-07-21 Richard Sandiford <richard.sandiford@linaro.org>
7
8 * regcprop.c (maybe_mode_change): Check HARD_REGNO_MODE_OK.
9
10=== modified file 'gcc/regcprop.c'
11--- old/gcc/regcprop.c 2010-12-17 22:51:25 +0000
12+++ new/gcc/regcprop.c 2011-07-21 11:30:53 +0000
13@@ -418,10 +418,9 @@
14
15 offset = ((WORDS_BIG_ENDIAN ? wordoffset : 0)
16 + (BYTES_BIG_ENDIAN ? byteoffset : 0));
17- return gen_rtx_raw_REG (new_mode,
18- regno + subreg_regno_offset (regno, orig_mode,
19- offset,
20- new_mode));
21+ regno += subreg_regno_offset (regno, orig_mode, offset, new_mode);
22+ if (HARD_REGNO_MODE_OK (regno, new_mode))
23+ return gen_rtx_raw_REG (new_mode, regno);
24 }
25 return NULL_RTX;
26 }
27
diff --git a/meta-oe/recipes-devtools/gcc/gcc-4_6-branch-linaro-backports.inc b/meta-oe/recipes-devtools/gcc/gcc-4_6-branch-linaro-backports.inc
index e3f6114e5..86dceabc3 100644
--- a/meta-oe/recipes-devtools/gcc/gcc-4_6-branch-linaro-backports.inc
+++ b/meta-oe/recipes-devtools/gcc/gcc-4_6-branch-linaro-backports.inc
@@ -18,4 +18,22 @@ file://linaro/gcc-4.6-linaro-r106751.patch \
18file://linaro/gcc-4.6-linaro-r106753.patch \ 18file://linaro/gcc-4.6-linaro-r106753.patch \
19file://linaro/gcc-4.6-linaro-r106754.patch \ 19file://linaro/gcc-4.6-linaro-r106754.patch \
20file://linaro/gcc-4.6-linaro-r106755.patch \ 20file://linaro/gcc-4.6-linaro-r106755.patch \
21file://linaro/gcc-4.6-linaro-r106759.patch \
22file://linaro/gcc-4.6-linaro-r106761.patch \
23file://linaro/gcc-4.6-linaro-r106762.patch \
24file://linaro/gcc-4.6-linaro-r106763.patch \
25file://linaro/gcc-4.6-linaro-r106764.patch \
26file://linaro/gcc-4.6-linaro-r106766.patch \
27file://linaro/gcc-4.6-linaro-r106768.patch \
28file://linaro/gcc-4.6-linaro-r106769.patch \
29file://linaro/gcc-4.6-linaro-r106770.patch \
30file://linaro/gcc-4.6-linaro-r106771.patch \
31file://linaro/gcc-4.6-linaro-r106772.patch \
32file://linaro/gcc-4.6-linaro-r106773.patch \
33file://linaro/gcc-4.6-linaro-r106775.patch \
34file://linaro/gcc-4.6-linaro-r106776.patch \
35file://linaro/gcc-4.6-linaro-r106777.patch \
36file://linaro/gcc-4.6-linaro-r106778.patch \
37file://linaro/gcc-4.6-linaro-r106781.patch \
38file://linaro/gcc-4.6-linaro-r106782.patch \
21" 39"
diff --git a/meta-oe/recipes-devtools/gcc/gcc-common-4.6.inc b/meta-oe/recipes-devtools/gcc/gcc-common-4.6.inc
index e5a1fba59..0faf45e93 100644
--- a/meta-oe/recipes-devtools/gcc/gcc-common-4.6.inc
+++ b/meta-oe/recipes-devtools/gcc/gcc-common-4.6.inc
@@ -1,4 +1,4 @@
1# this will prepend this layer to FILESPATH 1# this will prepend this layer to FILESPATH
2FILESEXTRAPATHS := "${THISDIR}/gcc-4.6" 2FILESEXTRAPATHS := "${THISDIR}/gcc-4.6"
3PRINC = "1" 3PRINC = "2"
4ARM_INSTRUCTION_SET = "arm" 4ARM_INSTRUCTION_SET = "arm"
generated by cgit v1.2.3-54-g00ecf (git 2.39.0) at 2024-11-15 05:37:46 +0000