]> git.ipfire.org Git - thirdparty/glibc.git/commitdiff
x86-64: Require BMI2 for AVX2 str(n)casecmp implementations
authorAurelien Jarno <aurelien@aurel32.net>
Mon, 3 Oct 2022 21:16:46 +0000 (23:16 +0200)
committerAurelien Jarno <aurelien@aurel32.net>
Mon, 3 Oct 2022 21:52:50 +0000 (23:52 +0200)
The AVX2 str(n)casecmp implementations use the 'bzhi' instruction, which
belongs to the BMI2 CPU feature.

NB: It also uses the 'tzcnt' BMI1 instruction, but it is executed as BSF
as BSF if the CPU doesn't support TZCNT, and produces the same result
for non-zero input.

Partially fixes: b77b06e0e296 ("x86: Optimize strcmp-avx2.S")
Partially resolves: BZ #29611

Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
(cherry picked from commit 10f79d3670b036925da63dc532b122d27ce65ff8)

sysdeps/x86_64/multiarch/ifunc-impl-list.c
sysdeps/x86_64/multiarch/ifunc-strcasecmp.h

index e97218f624bb9e077bf809dd1331bd638aab5e75..2d8c87a42eda4e5acdd1db285b4eaee97536e143 100644 (file)
@@ -404,13 +404,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, strcasecmp,
              IFUNC_IMPL_ADD (array, i, strcasecmp,
                              (CPU_FEATURE_USABLE (AVX512VL)
-                              && CPU_FEATURE_USABLE (AVX512BW)),
+                              && CPU_FEATURE_USABLE (AVX512BW)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strcasecmp_evex)
              IFUNC_IMPL_ADD (array, i, strcasecmp,
-                             CPU_FEATURE_USABLE (AVX2),
+                             (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strcasecmp_avx2)
              IFUNC_IMPL_ADD (array, i, strcasecmp,
                              (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)
                               && CPU_FEATURE_USABLE (RTM)),
                              __strcasecmp_avx2_rtm)
              IFUNC_IMPL_ADD (array, i, strcasecmp,
@@ -425,13 +428,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, strcasecmp_l,
              IFUNC_IMPL_ADD (array, i, strcasecmp,
                              (CPU_FEATURE_USABLE (AVX512VL)
-                              && CPU_FEATURE_USABLE (AVX512BW)),
+                              && CPU_FEATURE_USABLE (AVX512BW)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strcasecmp_l_evex)
              IFUNC_IMPL_ADD (array, i, strcasecmp,
-                             CPU_FEATURE_USABLE (AVX2),
+                             (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strcasecmp_l_avx2)
              IFUNC_IMPL_ADD (array, i, strcasecmp,
                              (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)
                               && CPU_FEATURE_USABLE (RTM)),
                              __strcasecmp_l_avx2_rtm)
              IFUNC_IMPL_ADD (array, i, strcasecmp_l,
@@ -560,13 +566,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, strncasecmp,
              IFUNC_IMPL_ADD (array, i, strncasecmp,
                              (CPU_FEATURE_USABLE (AVX512VL)
-                              && CPU_FEATURE_USABLE (AVX512BW)),
+                              && CPU_FEATURE_USABLE (AVX512BW)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strncasecmp_evex)
              IFUNC_IMPL_ADD (array, i, strncasecmp,
-                             CPU_FEATURE_USABLE (AVX2),
+                             (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strncasecmp_avx2)
              IFUNC_IMPL_ADD (array, i, strncasecmp,
                              (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)
                               && CPU_FEATURE_USABLE (RTM)),
                              __strncasecmp_avx2_rtm)
              IFUNC_IMPL_ADD (array, i, strncasecmp,
@@ -582,13 +591,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, strncasecmp_l,
              IFUNC_IMPL_ADD (array, i, strncasecmp,
                              (CPU_FEATURE_USABLE (AVX512VL)
-                              && CPU_FEATURE_USABLE (AVX512BW)),
+                              & CPU_FEATURE_USABLE (AVX512BW)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strncasecmp_l_evex)
              IFUNC_IMPL_ADD (array, i, strncasecmp,
-                             CPU_FEATURE_USABLE (AVX2),
+                             (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)),
                              __strncasecmp_l_avx2)
              IFUNC_IMPL_ADD (array, i, strncasecmp,
                              (CPU_FEATURE_USABLE (AVX2)
+                              && CPU_FEATURE_USABLE (BMI2)
                               && CPU_FEATURE_USABLE (RTM)),
                              __strncasecmp_l_avx2_rtm)
              IFUNC_IMPL_ADD (array, i, strncasecmp_l,
index 766539c241bbad69165039ce6f09341261e74c80..e201f6e4d8ffdcf97ddccc345120726313db6e2a 100644 (file)
@@ -32,6 +32,7 @@ IFUNC_SELECTOR (void)
   const struct cpu_features* cpu_features = __get_cpu_features ();
 
   if (CPU_FEATURE_USABLE_P (cpu_features, AVX2)
+      && CPU_FEATURE_USABLE_P (cpu_features, BMI2)
       && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
     {
       if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)