x86: Add Avoid_STOSB tunable to allow NT memset without ERMS

The goal of this flag is to allow targets which don't prefer/have ERMS
to still access the non-temporal memset implementation.

There are 4 cases for tuning memset:
    1) `Avoid_STOSB && Avoid_Non_Temporal_Memset`
        - Memset with temporal stores
    2) `Avoid_STOSB && !Avoid_Non_Temporal_Memset`
        - Memset with temporal/non-temporal stores. Non-temporal path
          goes through `rep stosb` path. We accomplish this by setting
          `x86_rep_stosb_threshold` to
          `x86_memset_non_temporal_threshold`.
    3) `!Avoid_STOSB && Avoid_Non_Temporal_Memset`
        - Memset with temporal stores/`rep stosb`
    3) `!Avoid_STOSB && !Avoid_Non_Temporal_Memset`
        - Memset with temporal stores/`rep stosb`/non-temporal stores.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>

sysdeps/x86/cpu-features.c

@@ -1119,6 +1119,10 @@ https://www.intel.com/content/www/us/en/support/articles/000059422/processors.ht
   if (CPU_FEATURES_CPU_P (cpu_features, CMOV))
     cpu_features->preferred[index_arch_I686] |= bit_arch_I686;
 
+  /* No ERMS, we want to avoid stosb for memset.  */
+  if (!CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+    cpu_features->preferred[index_arch_Avoid_STOSB] |= bit_arch_Avoid_STOSB;
+
 #if !HAS_CPUID
 no_cpuid:
 #endif

sysdeps/x86/cpu-tunables.c

@@ -193,6 +193,8 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp)
 						11);
 	      CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features, Prefer_FSRM,
 						11);
+	      CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features, Avoid_STOSB,
+						11);
 	      CHECK_GLIBC_IFUNC_PREFERRED_NEED_BOTH (n, cpu_features,
 						     Slow_SSE4_2,
 						     SSE4_2,

sysdeps/x86/dl-cacheinfo.h

@@ -1041,18 +1041,42 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
        slightly better than ERMS.  */
     rep_stosb_threshold = SIZE_MAX;
 
+  /*
+     For memset, the non-temporal implementation is only accessed through the
+     stosb code. ie:
+     ```
+     if (size >= rep_stosb_thresh)
+     {
+    	if (size >= non_temporal_thresh)
+     {
+     do_non_temporal ();
+     }
+    	do_stosb ();
+     }
+     do_normal_vec_loop ();
+     ```
+     So if we prefer non-temporal, set `rep_stosb_thresh = non_temporal_thresh`
+     to enable the implementation. If `rep_stosb_thresh = non_temporal_thresh`,
+    `rep stosb` will never be used.
+   */
+  TUNABLE_SET_WITH_BOUNDS (x86_memset_non_temporal_threshold,
+			   memset_non_temporal_threshold,
+			   minimum_non_temporal_threshold, SIZE_MAX);
+  /* Do `rep_stosb_thresh = non_temporal_thresh` after setting/getting the
+     final value of `x86_memset_non_temporal_threshold`. In some cases this can
+     be a matter of correctness.  */
+  if (CPU_FEATURES_ARCH_P (cpu_features, Avoid_STOSB))
+    rep_stosb_threshold
+	= TUNABLE_GET (x86_memset_non_temporal_threshold, long int, NULL);
+  TUNABLE_SET_WITH_BOUNDS (x86_rep_stosb_threshold, rep_stosb_threshold, 1,
+			   SIZE_MAX);
   TUNABLE_SET_WITH_BOUNDS (x86_data_cache_size, data, 0, SIZE_MAX);
   TUNABLE_SET_WITH_BOUNDS (x86_shared_cache_size, shared, 0, SIZE_MAX);
   TUNABLE_SET_WITH_BOUNDS (x86_non_temporal_threshold, non_temporal_threshold,
 			   minimum_non_temporal_threshold,
 			   maximum_non_temporal_threshold);
-  TUNABLE_SET_WITH_BOUNDS (x86_memset_non_temporal_threshold,
-			   memset_non_temporal_threshold,
-			   minimum_non_temporal_threshold, SIZE_MAX);
   TUNABLE_SET_WITH_BOUNDS (x86_rep_movsb_threshold, rep_movsb_threshold,
 			   minimum_rep_movsb_threshold, SIZE_MAX);
-  TUNABLE_SET_WITH_BOUNDS (x86_rep_stosb_threshold, rep_stosb_threshold, 1,
-			   SIZE_MAX);
 
   unsigned long int rep_movsb_stop_threshold;
   /* Setting the upper bound of ERMS to the computed value of

sysdeps/x86/include/cpu-features-preferred_feature_index_1.def

@@ -34,3 +34,4 @@ BIT (MathVec_Prefer_No_AVX512)
 BIT (Prefer_FSRM)
 BIT (Avoid_Short_Distance_REP_MOVSB)
 BIT (Avoid_Non_Temporal_Memset)
+BIT (Avoid_STOSB)

sysdeps/x86/tst-hwcap-tunables.c

@@ -60,7 +60,8 @@ static const struct test_t
     /* Disable everything.  */
     "-Prefer_ERMS,-Prefer_FSRM,-AVX,-AVX2,-AVX512F,-AVX512VL,"
     "-SSE4_1,-SSE4_2,-SSSE3,-Fast_Unaligned_Load,-ERMS,"
-    "-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset",
+    "-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,"
+    "-Avoid_STOSB",
     test_1,
     array_length (test_1)
   },
@@ -68,7 +69,8 @@ static const struct test_t
     /* Same as before, but with some empty suboptions.  */
     ",-,-Prefer_ERMS,-Prefer_FSRM,-AVX,-AVX2,-AVX512F,-AVX512VL,"
     "-SSE4_1,-SSE4_2,-SSSE3,-Fast_Unaligned_Load,,-,"
-    "-ERMS,-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,-,",
+    "-ERMS,-AVX_Fast_Unaligned_Load,-Avoid_Non_Temporal_Memset,"
+    "-Avoid_STOSB,-,",
     test_1,
     array_length (test_1)
   }

sysdeps/x86_64/multiarch/ifunc-memset.h

@@ -46,6 +46,13 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned)
 extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned_erms)
   attribute_hidden;
 
+static inline int
+prefer_erms_nt_impl (const struct cpu_features *cpu_features)
+{
+  return CPU_FEATURE_USABLE_P (cpu_features, ERMS)
+	 || !CPU_FEATURES_ARCH_P (cpu_features, Avoid_Non_Temporal_Memset);
+}
+
 static inline void *
 IFUNC_SELECTOR (void)
 {
@@ -61,7 +68,7 @@ IFUNC_SELECTOR (void)
 	  && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
 	  && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2))
 	{
-	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+	  if (prefer_erms_nt_impl (cpu_features))
 	    return OPTIMIZE (avx512_unaligned_erms);
 
 	  return OPTIMIZE (avx512_unaligned);
@@ -76,7 +83,7 @@ IFUNC_SELECTOR (void)
 	  && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
 	  && X86_ISA_CPU_FEATURE_USABLE_P (cpu_features, BMI2))
 	{
-	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+	  if (prefer_erms_nt_impl (cpu_features))
 	    return OPTIMIZE (evex_unaligned_erms);
 
 	  return OPTIMIZE (evex_unaligned);
@@ -84,7 +91,7 @@ IFUNC_SELECTOR (void)
 
       if (CPU_FEATURE_USABLE_P (cpu_features, RTM))
 	{
-	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+	  if (prefer_erms_nt_impl (cpu_features))
 	    return OPTIMIZE (avx2_unaligned_erms_rtm);
 
 	  return OPTIMIZE (avx2_unaligned_rtm);
@@ -93,14 +100,15 @@ IFUNC_SELECTOR (void)
       if (X86_ISA_CPU_FEATURES_ARCH_P (cpu_features,
 				       Prefer_No_VZEROUPPER, !))
 	{
-	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+	  if (prefer_erms_nt_impl (cpu_features))
 	    return OPTIMIZE (avx2_unaligned_erms);
 
 	  return OPTIMIZE (avx2_unaligned);
 	}
     }
 
-  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))
+  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS)
+      || !CPU_FEATURES_ARCH_P (cpu_features, Avoid_Non_Temporal_Memset))
     return OPTIMIZE (sse2_unaligned_erms);
 
   return OPTIMIZE (sse2_unaligned);