summaryrefslogtreecommitdiffstats
path: root/meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch
diff options
context:
space:
mode:
Diffstat (limited to 'meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch')
-rw-r--r--meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch341
1 files changed, 341 insertions, 0 deletions
diff --git a/meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch b/meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch
new file mode 100644
index 00000000..2a3d8632
--- /dev/null
+++ b/meta-microblaze/recipes-core/newlib/files/0009-Patch-MicroBlaze-Removing-the-Assembly-implementatio.patch
@@ -0,0 +1,341 @@
1From 5a7955b68f5066b00413e751d1de967181c88b94 Mon Sep 17 00:00:00 2001
2From: Mahesh Bodapati <mbodapat@xilinx.com>
3Date: Tue, 9 Nov 2021 22:53:44 +0530
4Subject: [PATCH 09/11] [Patch,MicroBlaze] : Removing the Assembly
5 implementation of 64bit string function.
6
7Revisit in next release and fix it
8
9(cherry picked from commit 6b8e5c7a773de4609f9c855aa714eca5a3f8b4ab)
10---
11 newlib/libc/machine/microblaze/mb_endian.h | 4 +
12 newlib/libc/machine/microblaze/strcmp.c | 95 ++++++++--------------
13 newlib/libc/machine/microblaze/strcpy.c | 82 ++++++-------------
14 newlib/libc/machine/microblaze/strlen.c | 59 +++++---------
15 4 files changed, 82 insertions(+), 158 deletions(-)
16
17diff --git a/newlib/libc/machine/microblaze/mb_endian.h b/newlib/libc/machine/microblaze/mb_endian.h
18index fb217ec85..17772c88f 100644
19--- a/newlib/libc/machine/microblaze/mb_endian.h
20+++ b/newlib/libc/machine/microblaze/mb_endian.h
21@@ -8,8 +8,12 @@
22 #ifdef __LITTLE_ENDIAN__
23 #define LOAD4BYTES(rD,rA,rB) "\tlwr\t" rD ", " rA ", " rB "\n"
24 #define STORE4BYTES(rD,rA,rB) "\tswr\t" rD ", " rA ", " rB "\n"
25+#define LOAD8BYTES(rD,rA,rB) "\tllr\t" rD ", " rA ", " rB "\n"
26+#define STORE8BYTES(rD,rA,rB) "\tslr\t" rD ", " rA ", " rB "\n"
27 #else
28 #define LOAD4BYTES(rD,rA,rB) "\tlw\t" rD ", " rA ", " rB "\n"
29 #define STORE4BYTES(rD,rA,rB) "\tsw\t" rD ", " rA ", " rB "\n"
30+#define LOAD8BYTES(rD,rA,rB) "\tll\t" rD ", " rA ", " rB "\n"
31+#define STORE8BYTES(rD,rA,rB) "\tsl\t" rD ", " rA ", " rB "\n"
32 #endif
33 #endif
34diff --git a/newlib/libc/machine/microblaze/strcmp.c b/newlib/libc/machine/microblaze/strcmp.c
35index 2cfef7388..e34c64a0c 100644
36--- a/newlib/libc/machine/microblaze/strcmp.c
37+++ b/newlib/libc/machine/microblaze/strcmp.c
38@@ -129,70 +129,42 @@ strcmp (const char *s1,
39 return (*(unsigned char *) s1) - (*(unsigned char *) s2);
40 #endif /* not PREFER_SIZE_OVER_SPEED */
41
42+#elif __arch64__
43+ unsigned int *a1;
44+ unsigned int *a2;
45+
46+ /* If s1 or s2 are unaligned, then compare bytes. */
47+ if (!UNALIGNED (s1, s2))
48+ {
49+ /* If s1 and s2 are word-aligned, compare them a word at a time. */
50+ a1 = (unsigned int*)s1;
51+ a2 = (unsigned int*)s2;
52+ while (*a1 == *a2)
53+ {
54+ /* To get here, *a1 == *a2, thus if we find a null in *a1,
55+ then the strings must be equal, so return zero. */
56+ if (DETECTNULL (*a1))
57+ return 0;
58+
59+ a1++;
60+ a2++;
61+ }
62+
63+ /* A difference was detected in last few bytes of s1, so search bytewise */
64+ s1 = (char*)a1;
65+ s2 = (char*)a2;
66+ }
67+
68+ while (*s1 != '\0' && *s1 == *s2)
69+ {
70+ s1++;
71+ s2++;
72+ }
73+ return (*(unsigned char *) s1) - (*(unsigned char *) s2);
74 #else
75
76 #include "mb_endian.h"
77
78-#ifdef __arch64__
79- asm volatile (" \n\
80- orl r9, r0, r0 /* Index register */ \n\
81-check_alignment: \n\
82- andli r3, r5, 3 \n\
83- andli r4, r6, 3 \n\
84- beanei r3, try_align_args \n\
85- beanei r4, regular_strcmp /* At this point we don't have a choice */ \n\
86-cmp_loop: \n"
87- LOAD4BYTES("r3", "r5", "r9")
88- LOAD4BYTES("r4", "r6", "r9")
89-" \n\
90- pcmplbf r7, r3, r0 /* See if there is Null byte */ \n\
91- beanei r7, end_cmp_loop /* IF yes (r7 > 0) use byte compares in end_cmp_loop */ \n\
92- cmplu r7, r4, r3 /* ELSE compare whole word */ \n\
93- beanei r7, end_cmp \n\
94- addlik r9, r9, 4 /* delay slot */ \n\
95- breaid cmp_loop \n\
96- nop /* delay slot */ \n\
97-end_cmp_loop: \n\
98- lbu r3, r5, r9 /* byte compare loop */ \n\
99- lbu r4, r6, r9 \n\
100- cmplu r7, r4, r3 /* Compare bytes */ \n\
101- beanei r7, end_cmp_early \n\
102- addlik r9, r9, 1 /* delay slot */ \n\
103- beaneid r3, end_cmp_loop /* If reached null on one string, terminate */ \n\
104- nop \n\
105-end_cmp_early: \n\
106- orl r3, r0, r7 /* delay slot */ \n\
107- rtsd r15, 8 \n\
108- nop \n\
109-try_align_args: \n\
110- xorl r7, r4, r3 \n\
111- beanei r7, regular_strcmp /* cannot align args */ \n\
112- rsublik r10, r3, 4 /* Number of initial bytes to align */ \n\
113-align_loop: \n\
114- lbu r3, r5, r9 \n\
115- lbu r4, r6, r9 \n\
116- cmplu r7, r4, r3 \n\
117- beanei r7, end_cmp \n\
118- beaeqi r3, end_cmp \n\
119- addlik r10, r10, -1 \n\
120- addlik r9, r9, 1 \n\
121- beaeqid r10, cmp_loop \n\
122- nop \n\
123- breai align_loop \n\
124-regular_strcmp: \n\
125- lbu r3, r5, r9 \n\
126- lbu r4, r6, r9 \n\
127- cmplu r7, r4, r3 \n\
128- beanei r7, end_cmp \n\
129- beaeqi r3, end_cmp \n\
130- addlik r9, r9, 1 \n\
131- breaid regular_strcmp \n\
132- nop \n\
133-end_cmp: \n\
134- orl r3, r0, r7 \n\
135- rtsd r15, 8 \n\
136- nop /* Return strcmp result */");
137-#else
138 asm volatile (" \n\
139 or r9, r0, r0 /* Index register */\n\
140 check_alignment: \n\
141@@ -241,12 +213,11 @@ regular_strcmp:
142 bnei r7, end_cmp \n\
143 beqi r3, end_cmp \n\
144 brid regular_strcmp \n\
145- addik r9, r9, 1
146+ addik r9, r9, 1 \n\
147 end_cmp: \n\
148 rtsd r15, 8 \n\
149 or r3, r0, r7 /* Return strcmp result */");
150
151-#endif
152 #endif /* ! HAVE_HW_PCMP */
153 }
154
155diff --git a/newlib/libc/machine/microblaze/strcpy.c b/newlib/libc/machine/microblaze/strcpy.c
156index 6dbc60d77..ddb69227e 100644
157--- a/newlib/libc/machine/microblaze/strcpy.c
158+++ b/newlib/libc/machine/microblaze/strcpy.c
159@@ -121,67 +121,36 @@ strcpy (char *__restrict dst0,
160 ;
161 return dst0;
162 #endif /* not PREFER_SIZE_OVER_SPEED */
163+#elif __arch64__
164+ char *dst = dst0;
165+ const char *src = src0;
166+ long *aligned_dst;
167+ const long *aligned_src;
168
169-#else
170+ /* If SRC or DEST is unaligned, then copy bytes. */
171+ if (!UNALIGNED (src, dst))
172+ {
173+ aligned_dst = (long*)dst;
174+ aligned_src = (long*)src;
175
176-#include "mb_endian.h"
177-#ifdef __arch64__
178+ /* SRC and DEST are both "long int" aligned, try to do "long int"
179+ sized copies. */
180+ while (!DETECTNULL(*aligned_src))
181+ {
182+ *aligned_dst++ = *aligned_src++;
183+ }
184
185- asm volatile (" \n\
186- orl r9, r0, r0 /* Index register */ \n\
187-check_alignment: \n\
188- andli r3, r5, 3 \n\
189- andli r4, r6, 3 \n\
190- beanei r3, try_align_args \n\
191- beanei r4, regular_strcpy /* At this point we dont have a choice */ \n\
192-cpy_loop: \n"
193- LOAD4BYTES("r3", "r6", "r9")
194-" \n\
195- pcmplbf r4, r0, r3 \n\
196- beanei r4, cpy_bytes /* If r4 != 0, then null present within string */\n"
197- STORE4BYTES("r3", "r5", "r9")
198-" \n\
199- addlik r9, r9, 4 \n\
200- breaid cpy_loop \n\
201- nop \n\
202-cpy_bytes: \n\
203- lbu r3, r6, r9 \n\
204- sb r3, r5, r9 \n\
205- addlik r4, r4, -1 \n\
206- addlik r9, r9, 1 /* delay slot */\n\
207- beaneid r4, cpy_bytes \n\
208- nop \n\
209-cpy_null: \n\
210- orl r3, r0, r5 /* Return strcpy result */\n\
211- rtsd r15, 8 \n\
212- nop \n\
213-try_align_args: \n\
214- xorl r7, r4, r3 \n\
215- beanei r7, regular_strcpy /* cannot align args */\n\
216- rsublik r10, r3, 4 /* Number of initial bytes to align */\n\
217-align_loop: \n\
218- lbu r3, r6, r9 \n\
219- sb r3, r5, r9 \n\
220- addlik r10, r10, -1 \n\
221- beaeqid r3, end_cpy /* Break if we have seen null character */\n\
222- nop \n\
223- addlik r9, r9, 1 \n\
224- beaneid r10, align_loop \n\
225- nop \n\
226- breai cpy_loop \n\
227-regular_strcpy: \n\
228- lbu r3, r6, r9 \n\
229- sb r3, r5, r9 \n\
230- addlik r9, r9, 1 \n\
231- beaneid r3, regular_strcpy \n\
232- nop \n\
233-end_cpy: \n\
234- orl r3, r0, r5 \n\
235- rtsd r15, 8 \n\
236- nop /* Return strcpy result */");
237+ dst = (char*)aligned_dst;
238+ src = (char*)aligned_src;
239+ }
240
241-#else
242+ while (*dst++ = *src++)
243+ ;
244+ return dst0;
245+
246+#else
247
248+#include "mb_endian.h"
249 asm volatile (" \n\
250 or r9, r0, r0 /* Index register */ \n\
251 check_alignment: \n\
252@@ -227,7 +196,6 @@ regular_strcpy: \n\
253 end_cpy: \n\
254 rtsd r15, 8 \n\
255 or r3, r0, r5 /* Return strcpy result */");
256-#endif
257 #endif /* ! HAVE_HW_PCMP */
258 }
259
260diff --git a/newlib/libc/machine/microblaze/strlen.c b/newlib/libc/machine/microblaze/strlen.c
261index b6f2d3c13..940753996 100644
262--- a/newlib/libc/machine/microblaze/strlen.c
263+++ b/newlib/libc/machine/microblaze/strlen.c
264@@ -112,47 +112,29 @@ strlen (const char *str)
265 return str - start;
266 #endif /* not PREFER_SIZE_OVER_SPEED */
267
268-#else
269-
270-#include "mb_endian.h"
271+#elif __arch64__
272+ const char *start = str;
273+ unsigned long *aligned_addr;
274
275-#ifdef __arch64__
276- asm volatile (" \n\
277- orl r9, r0, r0 /* Index register */ \n\
278-check_alignment: \n\
279- andli r3, r5, 3 \n\
280- beanei r3, align_arg \n\
281-len_loop: \n"
282- LOAD4BYTES("r3", "r5", "r9")
283-" \n\
284- pcmplbf r4, r3, r0 \n\
285- beanei r4, end_len \n\
286- addlik r9, r9, 4 \n\
287- breaid len_loop \n\
288- nop \n\
289-end_len: \n\
290- lbu r3, r5, r9 \n\
291- beaeqi r3, done_len \n\
292- addlik r9, r9, 1 \n\
293- breaid end_len \n\
294- nop \n\
295-done_len: \n\
296- orl r3, r0, r9 /* Return len */ \n\
297- rtsd r15, 8 \n\
298- nop \n\
299-align_arg: \n\
300- rsublik r10, r3, 4 \n\
301-align_loop: \n\
302- lbu r3, r5, r9 \n\
303- addlik r10, r10, -1 \n\
304- beaeqid r3, done_len \n\
305- nop \n\
306- addlik r9, r9, 1 \n\
307- beaneid r10, align_loop \n\
308- nop \n\
309- breai len_loop");
310+ if (!UNALIGNED (str))
311+ {
312+ /* If the string is word-aligned, we can check for the presence of
313+ a null in each word-sized block. */
314+ aligned_addr = (unsigned long*)str;
315+ while (!DETECTNULL (*aligned_addr))
316+ aligned_addr++;
317
318+ /* Once a null is detected, we check each byte in that block for a
319+ precise position of the null. */
320+ str = (char*)aligned_addr;
321+ }
322+
323+ while (*str)
324+ str++;
325+ return str - start;
326 #else
327+
328+#include "mb_endian.h"
329 asm volatile (" \n\
330 or r9, r0, r0 /* Index register */ \n\
331 check_alignment: \n\
332@@ -183,6 +165,5 @@ align_loop: \n\
333 addik r9, r9, 1 \n\
334 bri len_loop");
335
336-#endif
337 #endif /* ! HAVE_HW_PCMP */
338 }
339--
3402.34.1
341
generated by cgit v1.2.3-54-g00ecf (git 2.39.0) at 2026-04-22 21:40:44 +0000