Re: [dpdk-dev] [PATCH v2] build: check AVX512 rather than compiler version

On Tue, Jul 20, 2021 at 11:19:48AM +0100, Bruce Richardson wrote:

On Mon, Jul 19, 2021 at 11:34:33PM +0100, Liang Ma wrote:

quoted

From: Liang Ma <redacted>

GCC 6.3.0 has a known bug which related to _mm512_extracti64x4_epi64.
Please reference https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82887

Some DPDK PMD avx512 version heavily use _mm512_extracti64x4_epi6,
which cause building failure with debug buildtype.

Therefore, it's helpful to check if compiler work with
_mm512_extracti64x4_epi6.

This patch check the compiler compile result against the test code
snippet. If the checking is failed then disable avx512.

Bugzilla ID: 717
Fixes: e6a6a138919f (net/i40e: add AVX512 vector path)
Fixes: 808a17b3c1e6 (net/ice: add Rx AVX512 offload path)
Fixes: 4b64ccb328c9 (net/iavf: fix VLAN extraction in AVX512 path)
Cc: stable@dpdk.org

Reported-by: Liang Ma <redacted>
Signed-off-by: Liang Ma <redacted>
---

Looks generally ok, but some comments below.

quoted

 config/x86/meson.build | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/config/x86/meson.build b/config/x86/meson.build
index b9348c44de..77370a91f7 100644
--- a/config/x86/meson.build
+++ b/config/x86/meson.build

@@ -10,6 +10,19 @@ if not is_windows
     endif
 endif
 
+#check if compiler is working with _mm512_extracti64x4_epi64
+#Ref https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82887
+if not is_windows

Don't think this is needed here. The reason for it in the previous check
for avx512 is because that check is done by a shell script which won't work
on windows. Since we use meson functions for this check, it will work
anywhere.

quoted

+    code = '''#include <immintrin.h>
+    void test(__m512i zmm){
+        __m256i ymm = _mm512_extracti64x4_epi64(zmm, 0);}'''
+    result = cc.compiles(code, args : '-mavx512f', name : 'avx512 checking')
+    if result == false and cc.has_argument('-mno-avx512f')

Rather than checking for -mno-avx512f here, the whole block should probably
be in an avx512 block itself. If the compiler doesn't have the "-mavx512f"
flag, there may be problems with the "cc.compiles" command (or maybe it
just counts as an error case?).

I'd suggest changing the "is_windows" condition to a cc.has_argument() one
for avx512.

agree, I will fix that on v3.

quoted

+        machine_args += '-mno-avx512f'
+        warning('Broken _mm512_extracti64x4_epi64, disabling AVX512 support')
+    endif
+endif
+
 # we require SSE4.2 for DPDK
 if cc.get_define('__SSE4_2__', args: machine_args) == ''
     message('SSE 4.2 not enabled by default, explicitly enabling')
-- 
2.17.1