* add windows ssse3,sse4_1,sse4_2 detection for non avx path
* remove special WESTMERE case
* move windows conditional redefinition to separate header
* fix minor issues
* Fix getauxval comment (API 18 not 20)
getauxval is available in Android starting with API 18, not 20.
The comment about __ANDROID_API__ appears to have been copied from the
NDK's cpufeatures, which always uses dlopen/dlsym and doesn't assume it
can directly call getauxval, even if __ANDROID_API__ is new enough.
With this project, though, when __ANDROID_API__ is 18 or up, the
CMakeLists.txt file would detect that getauxval is available and define
HAVE_STRONG_GETAUXVAL.
* Broaden Qualcomm Krait idiv workaround
Some Qualcomm Krait CPUs have IDIV support but the kernel doesn't
report it. Previously, this code looked for two CPUs:
- 0x510006F2 (0x51/'Q', variant 0, part 0x06f, rev 2)
- 0x510006F3 (0x51/'Q', variant 0, part 0x06f, rev 3)
This check misses my 2013 Nexus 7 device, which has this CPU ID:
- 0x511006f0 (0x51/'Q', variant 1, part 0x06f, rev 0)
My Nexus 7 device doesn't report idiv through AT_HWCAP or through
/proc/cpuinfo (AT_HWCAP is 0x1b0d7).
Expand the check to anything with:
- implementer 0x51
- architecture 7
- part 0x4d or 0x6f
Part 0x4d appears to be a dual-core Krait (e.g. see
https://crbug.com/341598#c43).
This new matching behavior is a subset of what the upstream kernel
does (patch[1] contributed by CodeAurora), and also closely matches the
behavior of pytorch/cpuinfo.
[1] 120ecfafab
This patch reads the most important cpuid leaves and stores the data in
a struct. A followup patch will inline micro architecture detection and
brand string inside X86Info so we don't have to call `ReadLeaves` multiple
times. This wil allow further simplification of `HasSecondFMA` and help
fix https://github.com/google/cpu_features/issues/200.
Since buffers are a few tens of bytes there is no need for optimized memfunctions. For compile time sizes, the compiler will generate optimal code already.
This commit helps with platform code separation (fixes#3). It should also help with the build as we can simply include all `impl_*.c` files regardless of OS / arch.
Note: this patch contains breaking changes in `include/cpu_features_macros.h`
- `CPU_FEATURES_OS_LINUX_OR_ANDROID` does not exist anymore
- `CPU_FEATURES_OS_FREEBSD`, `CPU_FEATURES_OS_ANDROID` and `CPU_FEATURES_OS_LINUX` are now mutually exclusive (i.e. `CPU_FEATURES_OS_ANDROID` does not imply `CPU_FEATURES_OS_LINUX`)
- `CPU_FEATURES_OS_DARWIN` has been renamed into `CPU_FEATURES_OS_MACOS` to be able to target non-Mac Apple products (IOS, TV, WATCH). They are now targetable with `CPU_FEATURES_OS_IPHONE`. This matches Apple naming convention described in [this stackoverflow](https://stackoverflow.com/a/49560690).
The build fails with following message when -Werror
and -Werror=return-type are enabled.
In function ‘GetCacheTypeString’:
error: control reaches end of non-void function [-Werror=return-type]
Simple fix is to return explicitly communicate to
the compiler that certain block is not reachable.
