x86-64: Require BMI2 for AVX2 strncmp implementation
authorAurelien Jarno <aurelien@aurel32.net>
Mon, 3 Oct 2022 21:46:11 +0000 (23:46 +0200)
committerAurelien Jarno <aurelien@aurel32.net>
Mon, 3 Oct 2022 21:46:11 +0000 (23:46 +0200)
The AVX2 strncmp implementations uses 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>
sysdeps/x86_64/multiarch/ifunc-impl-list.c
sysdeps/x86_64/multiarch/strncmp.c

index a42b0a4..aebef3d 100644 (file)
@@ -1176,13 +1176,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, strncmp,
              X86_IFUNC_IMPL_ADD_V4 (array, i, strncmp,
                                     (CPU_FEATURE_USABLE (AVX512VL)
-                                     && CPU_FEATURE_USABLE (AVX512BW)),
+                                     && CPU_FEATURE_USABLE (AVX512BW)
+                                     && CPU_FEATURE_USABLE (BMI2)),
                                     __strncmp_evex)
              X86_IFUNC_IMPL_ADD_V3 (array, i, strncmp,
-                                    CPU_FEATURE_USABLE (AVX2),
+                                    (CPU_FEATURE_USABLE (AVX2)
+                                     && CPU_FEATURE_USABLE (BMI2)),
                                     __strncmp_avx2)
              X86_IFUNC_IMPL_ADD_V3 (array, i, strncmp,
                                     (CPU_FEATURE_USABLE (AVX2)
+                                     && CPU_FEATURE_USABLE (BMI2)
                                      && CPU_FEATURE_USABLE (RTM)),
                                     __strncmp_avx2_rtm)
              X86_IFUNC_IMPL_ADD_V2 (array, i, strncmp,
index 4ebe4bd..c4f8b6b 100644 (file)
@@ -41,12 +41,12 @@ IFUNC_SELECTOR (void)
   const struct cpu_features *cpu_features = __get_cpu_features ();
 
   if (X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX2)
+      && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2)
       && X86_ISA_CPU_FEATURES_ARCH_P (cpu_features,
                                      AVX_Fast_Unaligned_Load, ))
     {
       if (X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
-         && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
-         && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2))
+         && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
        return OPTIMIZE (evex);
 
       if (CPU_FEATURE_USABLE_P (cpu_features, RTM))