diff options
authorAlejandro Enedino Hernandez Samaniego <>2019-08-27 17:34:28 (GMT)
committerAnuj Mittal <>2019-08-29 03:34:35 (GMT)
commit0bc64c2ad39fbf2bcd4673c817f5f52a5b9be4be (patch)
parent61dae7326b510960130aa3c93026d5f0abd1773b (diff)
downloadmeta-intel-0bc64c2ad39fbf2bcd4673c817f5f52a5b9be4be.tar.gz Improve handling the newly introduced skylake tune.
The skylake tune introduced on 04510bfa Currently disables qemu-usermode by default, due to the lack of support from QEMU to several of the Intel instruction set extensions introduced within the few latest generations of CPU architectures (e.g. QEMU does not support avx2) While there is a good reason to perform the removal of qemu-usermode from MACHINE_FEATURES, there are several components within the build system that rely on it for proper compilation and behavior, for example anything that uses gobject data introspection or even the components like the chromium web browser require to run a QEMU for the target architecture to build successfully. There is no reason why we can't have those components built without sacrificing the cpu (most) optimizations. The process I followed on meta-chromebook to enable an optimized build and whats being upstreamed by this patch is that by doing some reverse engineering, I was able to figure out which instruction set extensions are not compatible with QEMU Skylake-Client, by performing a bit gcc magic from inside QEMU (target) to get the available optimizations for the native architecture (which is actually our target in this case). These are all (not surprisingly) the avx2 extensions, listed as follows: -mno-avx -mno-avx2 -mno-avx512f -mno-avx512er -mno-avx512cd -mno-avx512pf -mno-avx512dq -mno-avx512bw -mno-avx512vl -mno-avx512ifma -mno-avx512vbmi -mno-avx512vbmi2 -mno-avx512vnni -mno-avx512bitalg Specifically disabling these manually (for now), allows us to build an optimized system for the skylake/skylake based architectures (e.g. KabyLake) while keeping the capability of using qemu-usermode, as a side note GCC does show more unavailable optimization tunes, (hence why there might be some warnings), but getting rid of these specifically seems enough to make it run happily in qemu-usermode. Basically the MACHINE_FEATURES variable is able to dictate how we will tune the build for our device, if qemu-usermode is present, TUNE_CCARGS will expand as follows: TUNE_CCARGS=" -m64 -march=skylake -mtune=skylake -mno-avx -mno-avx2 -mno-avx512f -mno-avx512er -mno-avx512cd -mno-avx512pf -mno-avx512dq -mno-avx512bw -mno-avx512vl -mno-avx512ifma -mno-avx512vbmi -mno-avx512vbmi2 -mno-avx512vnni -mno-avx512bitalg -mfpmath=sse" Whats this means is that the build will be tuned for skylake architectures, enabling all possible extensions, (MOVBE, MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, AES, PCLMUL, FSGSBASE, RDRND, FMA, BMI, BMI2, F16C, RDSEED, ADCX, PREFETCHW, CLFLUSHOPT, XSAVEC and XSAVES) minus the ones QEMU is unable to run, which will result in an illegal instruction error, notice the tune is kept as skylake. Whilst, if qemu-usermode is not found on MACHINE_FEATURES, TUNE_CCARGS will expand to: TUNE_CCARGS=" -m64 -march=skylake -mtune=generic -mavx2 -mfpmath=sse" Which is exactly what its set to with the current tune, so this patch should be harmless, and only extend current functionality, we will still keep qemu-usermode disabled by default (hence enabling avx2) but the user would have the capability to decide how the skylake tune will be handled. As the GCC manual states, we should really try to avoid using mtune=generic when possible, and this patch allows us to do so [1]. This patch also addresses a problem on which the current skylake tune includes instead of to get the list of AVAILTUNES and PACKAGE_EXTRA_ARCHS. Right now, AVAILTUNES are set as follows: AVAILTUNES=" x86 x86-64 x86-64-x32 i586 i686 core2-32 core2-64 core2-64-x32 skylake-64" Where the proper set should be (after this patch): AVAILTUNES=" x86 x86-64 x86-64-x32 i586 i686 core2-32 core2-64 core2-64-x32 corei7-32 corei7-64 corei7-64-x32 skylake-64" When (if), QEMU gains support for the AVX2 instruction set extensions these can be easily removed to provide full support for qemu-usermode. [1] Signed-off-by: Alejandro Enedino Hernandez Samaniego <> Signed-off-by: Anuj Mittal <>
1 files changed, 23 insertions, 3 deletions
diff --git a/conf/machine/include/ b/conf/machine/include/
index 78645de..4882600 100644
--- a/conf/machine/include/
+++ b/conf/machine/include/
@@ -8,11 +8,11 @@
8DEFAULTTUNE ?= "skylake-64" 8DEFAULTTUNE ?= "skylake-64"
9 9
10# Include the previous tune to pull in PACKAGE_EXTRA_ARCHS 10# Include the previous tune to pull in PACKAGE_EXTRA_ARCHS
11require conf/machine/include/ 11require conf/machine/include/
12 12
13# Extra tune features 13# Extra tune features
14TUNEVALID[skylake] = "Enable skylake specific processor optimizations" 14TUNEVALID[skylake] = "Enable skylake specific processor optimizations"
15TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'skylake', ' -march=skylake -mtune=generic -mfpmath=sse -mavx2', '', d)}" 15TUNE_CCARGS .= "${@bb.utils.contains('TUNE_FEATURES', 'skylake', ' -march=skylake ${SKYLAKE_TUNE} -mfpmath=sse', '', d)}"
16 16
17# Extra tune selections 17# Extra tune selections
18 18
@@ -23,6 +23,26 @@ TUNE_PKGARCH_tune-skylake-64 = "skylake-64"
23PACKAGE_EXTRA_ARCHS_tune-skylake-64 = "${PACKAGE_EXTRA_ARCHS_tune-core2-64} skylake-64" 23PACKAGE_EXTRA_ARCHS_tune-skylake-64 = "${PACKAGE_EXTRA_ARCHS_tune-core2-64} skylake-64"
24QEMU_EXTRAOPTIONS_skylake-64 = " -cpu Skylake-Client" 24QEMU_EXTRAOPTIONS_skylake-64 = " -cpu Skylake-Client"
25 25
26# Disable QEMU user 26
27# Disable QEMU usermode by default (get avx2)
27MACHINE_FEATURES_remove = "qemu-usermode" 28MACHINE_FEATURES_remove = "qemu-usermode"
28 29
30# If qemu-usermode is enabled, we have to disable avx2 ISA extensions, but we can keep mtune as skylake vs generic
31SKYLAKE_TUNE .= "${@bb.utils.contains('MACHINE_FEATURES', 'qemu-usermode', ' -mtune=skylake ${QEMU_UNAVAILABLE_ISA}', '-mtune=generic -mavx2', d)}"
34-mno-avx \
35-mno-avx2 \
36-mno-avx512f \
37-mno-avx512er \
38-mno-avx512cd \
39-mno-avx512pf \
40-mno-avx512dq \
41-mno-avx512bw \
42-mno-avx512vl \
43-mno-avx512ifma \
44-mno-avx512vbmi \
45-mno-avx512vbmi2 \
46-mno-avx512vnni \
47-mno-avx512bitalg \