[x64][ia32] Reconcile supported extensions

This is similar in spirit to https://crrev.com/c/2808621, which is to
ensure that no matter what combination of --[no-]enable-{extension} flag
is passed, we end up with a set of supported extensions that make sense.

The 2 rules are:

- If a newer extension is supported (SSE4_2), older extensions are
supported (SSE4_1, SSSE3, SSE3),
- If an older extension is not supported (SSE4_1), new extensions are
not supported (SSE4_2, AVX)

Tests have been added to both ia32 and x64 to check that we follow these
above 2 rules.

We change the ProbeImpl to have a reconciliation step to ensure that we
stick to the 2 rules.

E.g. if --enable-avx --no-enable-sse4-2, we will first set AVX to
supported, then in the second step, fix-up AVX to unsupported. In this
sense, the --no version of the flags take priority. This more accurately
follows the intention of the flags.

Bug: chromium:1195579
Change-Id: I0390f24de9d203fe6bbd4cc02a23771a1f052618
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2818570
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Reviewed-by: Michael Achenbach <machenbach@chromium.org>
Reviewed-by: Clemens Backes <clemensb@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#73942}

src/codegen/cpu-features.h

@@ -109,6 +109,9 @@ class V8_EXPORT_PRIVATE CpuFeatures : public AllStatic {
     return (supported_ & (1u << f)) != 0;
   }
 
+  static void SetSupported(CpuFeature f) { supported_ |= 1u << f; }
+  static void SetUnsupported(CpuFeature f) { supported_ &= ~(1u << f); }
+
   static bool SupportsWasmSimd128();
 
   static inline bool SupportsOptimizer();

src/codegen/ia32/assembler-ia32.cc

@@ -138,31 +138,38 @@ void CpuFeatures::ProbeImpl(bool cross_compile) {
   // Only use statically determined features for cross compile (snapshot).
   if (cross_compile) return;
 
-  if (cpu.has_sse42() && FLAG_enable_sse4_2) supported_ |= 1u << SSE4_2;
-  if (cpu.has_sse41() && FLAG_enable_sse4_1) supported_ |= 1u << SSE4_1;
-  if (cpu.has_ssse3() && FLAG_enable_ssse3) supported_ |= 1u << SSSE3;
-  if (cpu.has_sse3() && FLAG_enable_sse3) supported_ |= 1u << SSE3;
-  if (cpu.has_avx() && FLAG_enable_avx && cpu.has_osxsave() &&
-      OSHasAVXSupport()) {
-    supported_ |= 1u << AVX;
+  if (cpu.has_sse42()) SetSupported(SSE4_2);
+  if (cpu.has_sse41()) SetSupported(SSE4_1);
+  if (cpu.has_ssse3()) SetSupported(SSSE3);
+  if (cpu.has_sse3()) SetSupported(SSE3);
+  if (cpu.has_avx() && cpu.has_osxsave() && OSHasAVXSupport()) {
+    SetSupported(AVX);
+    if (cpu.has_avx2()) SetSupported(AVX2);
+    if (cpu.has_fma3()) SetSupported(FMA3);
   }
-  if (cpu.has_avx2() && FLAG_enable_avx2 && IsSupported(AVX)) {
-    supported_ |= 1u << AVX2;
-  }
-  if (cpu.has_fma3() && FLAG_enable_fma3 && cpu.has_osxsave() &&
-      OSHasAVXSupport()) {
-    supported_ |= 1u << FMA3;
-  }
-  if (cpu.has_bmi1() && FLAG_enable_bmi1) supported_ |= 1u << BMI1;
-  if (cpu.has_bmi2() && FLAG_enable_bmi2) supported_ |= 1u << BMI2;
-  if (cpu.has_lzcnt() && FLAG_enable_lzcnt) supported_ |= 1u << LZCNT;
-  if (cpu.has_popcnt() && FLAG_enable_popcnt) supported_ |= 1u << POPCNT;
+
+  if (cpu.has_bmi1() && FLAG_enable_bmi1) SetSupported(BMI1);
+  if (cpu.has_bmi2() && FLAG_enable_bmi2) SetSupported(BMI2);
+  if (cpu.has_lzcnt() && FLAG_enable_lzcnt) SetSupported(LZCNT);
+  if (cpu.has_popcnt() && FLAG_enable_popcnt) SetSupported(POPCNT);
   if (strcmp(FLAG_mcpu, "auto") == 0) {
-    if (cpu.is_atom()) supported_ |= 1u << ATOM;
+    if (cpu.is_atom()) SetSupported(ATOM);
   } else if (strcmp(FLAG_mcpu, "atom") == 0) {
-    supported_ |= 1u << ATOM;
+    SetSupported(ATOM);
   }
 
+  // Ensure that supported cpu features make sense. E.g. it is wrong to support
+  // AVX but not SSE4_2, if we have --enable-avx and --no-enable-sse4-2, the
+  // code above would set AVX to supported, and SSE4_2 to unsupported, then the
+  // checks below will set AVX to unsupported.
+  if (!FLAG_enable_sse3) SetUnsupported(SSE3);
+  if (!FLAG_enable_ssse3 || !IsSupported(SSE3)) SetUnsupported(SSSE3);
+  if (!FLAG_enable_sse4_1 || !IsSupported(SSSE3)) SetUnsupported(SSE4_1);
+  if (!FLAG_enable_sse4_2 || !IsSupported(SSE4_1)) SetUnsupported(SSE4_2);
+  if (!FLAG_enable_avx || !IsSupported(SSE4_2)) SetUnsupported(AVX);
+  if (!FLAG_enable_avx2 || !IsSupported(AVX)) SetUnsupported(AVX2);
+  if (!FLAG_enable_fma3 || !IsSupported(AVX)) SetUnsupported(FMA3);
+
   // Set a static value on whether Simd is supported.
   // This variable is only used for certain archs to query SupportWasmSimd128()
   // at runtime in builtins using an extern ref. Other callers should use

src/codegen/x64/assembler-x64.cc

@@ -86,37 +86,40 @@ void CpuFeatures::ProbeImpl(bool cross_compile) {
 
   // Only use statically determined features for cross compile (snapshot).
   if (cross_compile) return;
-
-  if (cpu.has_sse42() && FLAG_enable_sse4_2) supported_ |= 1u << SSE4_2;
-  if (cpu.has_sse41() && FLAG_enable_sse4_1) {
-    supported_ |= 1u << SSE4_1;
-    supported_ |= 1u << SSSE3;
+  if (cpu.has_sse42()) SetSupported(SSE4_2);
+  if (cpu.has_sse41()) SetSupported(SSE4_1);
+  if (cpu.has_ssse3()) SetSupported(SSSE3);
+  if (cpu.has_sse3()) SetSupported(SSE3);
+  if (cpu.has_avx() && cpu.has_osxsave() && OSHasAVXSupport()) {
+    SetSupported(AVX);
+    if (cpu.has_avx2()) SetSupported(AVX2);
+    if (cpu.has_fma3()) SetSupported(FMA3);
   }
-  if (cpu.has_ssse3() && FLAG_enable_ssse3) supported_ |= 1u << SSSE3;
-  if (cpu.has_sse3() && FLAG_enable_sse3) supported_ |= 1u << SSE3;
+
   // SAHF is not generally available in long mode.
-  if (cpu.has_sahf() && FLAG_enable_sahf) supported_ |= 1u << SAHF;
-  if (cpu.has_avx() && FLAG_enable_avx && cpu.has_osxsave() &&
-      OSHasAVXSupport()) {
-    supported_ |= 1u << AVX;
-  }
-  if (cpu.has_avx2() && FLAG_enable_avx2 && IsSupported(AVX)) {
-    supported_ |= 1u << AVX2;
-  }
-  if (cpu.has_fma3() && FLAG_enable_fma3 && cpu.has_osxsave() &&
-      OSHasAVXSupport()) {
-    supported_ |= 1u << FMA3;
-  }
-  if (cpu.has_bmi1() && FLAG_enable_bmi1) supported_ |= 1u << BMI1;
-  if (cpu.has_bmi2() && FLAG_enable_bmi2) supported_ |= 1u << BMI2;
-  if (cpu.has_lzcnt() && FLAG_enable_lzcnt) supported_ |= 1u << LZCNT;
-  if (cpu.has_popcnt() && FLAG_enable_popcnt) supported_ |= 1u << POPCNT;
+  if (cpu.has_sahf() && FLAG_enable_sahf) SetSupported(SAHF);
+  if (cpu.has_bmi1() && FLAG_enable_bmi1) SetSupported(BMI1);
+  if (cpu.has_bmi2() && FLAG_enable_bmi2) SetSupported(BMI2);
+  if (cpu.has_lzcnt() && FLAG_enable_lzcnt) SetSupported(LZCNT);
+  if (cpu.has_popcnt() && FLAG_enable_popcnt) SetSupported(POPCNT);
   if (strcmp(FLAG_mcpu, "auto") == 0) {
-    if (cpu.is_atom()) supported_ |= 1u << ATOM;
+    if (cpu.is_atom()) SetSupported(ATOM);
   } else if (strcmp(FLAG_mcpu, "atom") == 0) {
-    supported_ |= 1u << ATOM;
+    SetSupported(ATOM);
   }
 
+  // Ensure that supported cpu features make sense. E.g. it is wrong to support
+  // AVX but not SSE4_2, if we have --enable-avx and --no-enable-sse4-2, the
+  // code above would set AVX to supported, and SSE4_2 to unsupported, then the
+  // checks below will set AVX to unsupported.
+  if (!FLAG_enable_sse3) SetUnsupported(SSE3);
+  if (!FLAG_enable_ssse3 || !IsSupported(SSE3)) SetUnsupported(SSSE3);
+  if (!FLAG_enable_sse4_1 || !IsSupported(SSSE3)) SetUnsupported(SSE4_1);
+  if (!FLAG_enable_sse4_2 || !IsSupported(SSE4_1)) SetUnsupported(SSE4_2);
+  if (!FLAG_enable_avx || !IsSupported(SSE4_2)) SetUnsupported(AVX);
+  if (!FLAG_enable_avx2 || !IsSupported(AVX)) SetUnsupported(AVX2);
+  if (!FLAG_enable_fma3 || !IsSupported(AVX)) SetUnsupported(FMA3);
+
   // Set a static value on whether Simd is supported.
   // This variable is only used for certain archs to query SupportWasmSimd128()
   // at runtime in builtins using an extern ref. Other callers should use

test/cctest/test-assembler-ia32.cc

@@ -1550,6 +1550,35 @@ TEST(DeoptExitSizeIsFixed) {
   }
 }
 
+TEST(CpuFeatures_ProbeImpl) {
+  // Support for a newer extension implies support for the older extensions.
+  CHECK_IMPLIES(CpuFeatures::IsSupported(FMA3), CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(AVX2), CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(AVX),
+                CpuFeatures::IsSupported(SSE4_2));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSE4_2),
+                CpuFeatures::IsSupported(SSE4_1));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSE4_1),
+                CpuFeatures::IsSupported(SSSE3));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSSE3),
+                CpuFeatures::IsSupported(SSE3));
+
+  // Check the reverse, if an older extension is not supported, a newer
+  // extension cannot be supported.
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE3),
+                !CpuFeatures::IsSupported(SSSE3));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSSE3),
+                !CpuFeatures::IsSupported(SSE4_1));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE4_1),
+                !CpuFeatures::IsSupported(SSE4_2));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE4_2),
+                !CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(AVX),
+                !CpuFeatures::IsSupported(AVX2));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(AVX),
+                !CpuFeatures::IsSupported(FMA3));
+}
+
 #undef __
 
 }  // namespace internal

test/cctest/test-assembler-x64.cc

@@ -2521,6 +2521,35 @@ TEST(AssemblerX64vmovups) {
   CHECK_EQ(-1.5, f.Call(1.5, -1.5));
 }
 
+TEST(CpuFeatures_ProbeImpl) {
+  // Support for a newer extension implies support for the older extensions.
+  CHECK_IMPLIES(CpuFeatures::IsSupported(FMA3), CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(AVX2), CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(AVX),
+                CpuFeatures::IsSupported(SSE4_2));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSE4_2),
+                CpuFeatures::IsSupported(SSE4_1));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSE4_1),
+                CpuFeatures::IsSupported(SSSE3));
+  CHECK_IMPLIES(CpuFeatures::IsSupported(SSSE3),
+                CpuFeatures::IsSupported(SSE3));
+
+  // Check the reverse, if an older extension is not supported, a newer
+  // extension cannot be supported.
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE3),
+                !CpuFeatures::IsSupported(SSSE3));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSSE3),
+                !CpuFeatures::IsSupported(SSE4_1));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE4_1),
+                !CpuFeatures::IsSupported(SSE4_2));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(SSE4_2),
+                !CpuFeatures::IsSupported(AVX));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(AVX),
+                !CpuFeatures::IsSupported(AVX2));
+  CHECK_IMPLIES(!CpuFeatures::IsSupported(AVX),
+                !CpuFeatures::IsSupported(FMA3));
+}
+
 #undef __
 
 }  // namespace internal