[X86] Support -march=x86-64-v[234]

PR47686. These micro-architecture levels are defined in the x86-64 psABI:

https://gitlab.com/x86-psABIs/x86-64-ABI/-/commit/77566eb03bc6a326811cb7e9

GCC 11 will support these levels.

Note, -mtune=x86-64-v[234] are invalid and __builtin_cpu_is cannot be
used on them.

Reviewed By: craig.topper, RKSimon

Differential Revision: https://reviews.llvm.org/D89197

Author: MaskRay

clang/docs/ReleaseNotes.rst

@@ -188,7 +188,10 @@ X86 Support in Clang
 - The x86 intrinsics ``__rorb``, ``__rorw``, ``__rord``, ``__rorq`, ``_rotr``,
   ``_rotwr`` and ``_lrotr`` may now be used within constant expressions.
 
-- Support for -march=sapphirerapids was added.
+- Support for ``-march=sapphirerapids`` was added.
+
+- Support for ``-march=x86-64-v[234]`` has been added.
+  See :doc:`UsersManual` for details about these micro-architecture levels.
 
 - The -mtune command line option is no longer ignored for X86. This can be used
   to request microarchitectural optimizations independent on -march. -march=<cpu>

clang/docs/UsersManual.rst

@@ -3201,6 +3201,15 @@ and the ABI remains 32-bit but the assembler emits instructions
 appropriate for a CPU running in 16-bit mode, with address-size and
 operand-size prefixes to enable 32-bit addressing and operations.
 
+Several micro-architecture levels as specified by the x86-64 psABI are defined.
+They are cumulative in the sense that features from previous levels are
+implicitly included in later levels.
+
+- ``-march=x86-64``: CMOV, CMPXCHG8B, FPU, FXSR, MMX, FXSR, SCE, SSE, SSE2
+- ``-march=x86-64-v2``: (close to Nehalem) CMPXCHG16B, LAHF-SAHF, POPCNT, SSE3, SSE4.1, SSE4.2, SSSE3
+- ``-march=x86-64-v3``: (close to Haswell) AVX, AVX2, BMI1, BMI2, F16C, FMA, LZCNT, MOVBE, XSAVE
+- ``-march=x86-64-v4``: AVX512F, AVX512BW, AVX512CD, AVX512DQ, AVX512VL
+
 ARM
 ^^^
 

clang/lib/Basic/Targets/X86.cpp

@@ -506,6 +506,9 @@ void X86TargetInfo::getTargetDefines(const LangOptions &Opts,
   case CK_K8:
   case CK_K8SSE3:
   case CK_x86_64:
+  case CK_x86_64_v2:
+  case CK_x86_64_v3:
+  case CK_x86_64_v4:
     defineCPUMacros(Builder, "k8");
     break;
   case CK_AMDFAM10:
@@ -1312,6 +1315,9 @@ Optional<unsigned> X86TargetInfo::getCPUCacheLineSize() const {
     case CK_ZNVER2:
     // Deprecated
     case CK_x86_64:
+    case CK_x86_64_v2:
+    case CK_x86_64_v3:
+    case CK_x86_64_v4:
     case CK_Yonah:
     case CK_Penryn:
     case CK_Core2:
@@ -1456,7 +1462,7 @@ void X86TargetInfo::fillValidCPUList(SmallVectorImpl<StringRef> &Values) const {
 }
 
 void X86TargetInfo::fillValidTuneCPUList(SmallVectorImpl<StringRef> &Values) const {
-  llvm::X86::fillValidCPUArchList(Values);
+  llvm::X86::fillValidTuneCPUList(Values);
 }
 
 ArrayRef<const char *> X86TargetInfo::getGCCRegNames() const {

clang/lib/Basic/Targets/X86.h

@@ -314,7 +314,7 @@ class LLVM_LIBRARY_VISIBILITY X86TargetInfo : public TargetInfo {
     // Allow 32-bit only CPUs regardless of 64-bit mode unlike isValidCPUName.
     // NOTE: gcc rejects 32-bit mtune CPUs in 64-bit mode. But being lenient
     // since mtune was ignored by clang for so long.
-    return llvm::X86::parseArchX86(Name) != llvm::X86::CK_None;
+    return llvm::X86::parseTuneCPU(Name) != llvm::X86::CK_None;
   }
 
   void fillValidCPUList(SmallVectorImpl<StringRef> &Values) const override;

clang/test/CodeGen/attr-target-x86.c

@@ -32,6 +32,10 @@ int __attribute__((target("arch=lakemont,mmx"))) use_before_def(void) {
 
 int __attribute__((target("tune=sandybridge"))) walrus(int a) { return 4; }
 
+void __attribute__((target("arch=x86-64-v2"))) x86_64_v2() {}
+void __attribute__((target("arch=x86-64-v3"))) x86_64_v3() {}
+void __attribute__((target("arch=x86-64-v4"))) x86_64_v4() {}
+
 // Check that we emit the additional subtarget and cpu features for foo and not for baz or bar.
 // CHECK: baz{{.*}} #0
 // CHECK: foo{{.*}} #1
@@ -59,3 +63,10 @@ int __attribute__((target("tune=sandybridge"))) walrus(int a) { return 4; }
 // CHECK: #7 = {{.*}}"target-cpu"="lakemont" "target-features"="+cx8,+mmx"
 // CHECK-NOT: tune-cpu
 // CHECK: #8 = {{.*}}"target-cpu"="i686" "target-features"="+cx8,+x87" "tune-cpu"="sandybridge"
+
+// CHECK: "target-cpu"="x86-64-v2"
+// CHECK-SAME: "target-features"="+cx16,+cx8,+fxsr,+mmx,+popcnt,+sahf,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87"
+// CHECK: "target-cpu"="x86-64-v3"
+// CHECK-SAME: "target-features"="+avx,+avx2,+bmi,+bmi2,+cx16,+cx8,+f16c,+fma,+fxsr,+lzcnt,+mmx,+movbe,+popcnt,+sahf,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave"
+// CHECK: "target-cpu"="x86-64-v4"
+// CHECK-SAME: "target-features"="+avx,+avx2,+avx512bw,+avx512cd,+avx512dq,+avx512f,+avx512vl,+bmi,+bmi2,+cx16,+cx8,+f16c,+fma,+fxsr,+lzcnt,+mmx,+movbe,+popcnt,+sahf,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave"

clang/test/Driver/x86-march.c

@@ -175,3 +175,12 @@
 // RUN: %clang -target x86_64-unknown-unknown -c -### %s -march=znver2 2>&1 \
 // RUN:   | FileCheck %s -check-prefix=znver2
 // znver2: "-target-cpu" "znver2"
+
+// RUN: %clang -target x86_64 -c -### %s -march=x86-64 2>&1 | FileCheck %s --check-prefix=x86-64
+// x86-64: "-target-cpu" "x86-64"
+// RUN: %clang -target x86_64 -c -### %s -march=x86-64-v2 2>&1 | FileCheck %s --check-prefix=x86-64-v2
+// x86-64-v2: "-target-cpu" "x86-64-v2"
+// RUN: %clang -target x86_64 -c -### %s -march=x86-64-v3 2>&1 | FileCheck %s --check-prefix=x86-64-v3
+// x86-64-v3: "-target-cpu" "x86-64-v3"
+// RUN: %clang -target x86_64 -c -### %s -march=x86-64-v4 2>&1 | FileCheck %s --check-prefix=x86-64-v4
+// x86-64-v4: "-target-cpu" "x86-64-v4"

clang/test/Driver/x86-mtune.c

@@ -40,3 +40,8 @@
 // RUN:   | FileCheck %s -check-prefix=marchmtune
 // marchmtune: "-target-cpu" "core2"
 // mmarchmtune: "-tune-cpu" "nehalem"
+
+// RUN: not %clang %s -target x86_64 -E -mtune=x86-64-v2 2>&1 | FileCheck %s --check-prefix=INVALID
+// RUN: not %clang %s -target x86_64 -E -mtune=x86-64-v3 2>&1 | FileCheck %s --check-prefix=INVALID
+// RUN: not %clang %s -target x86_64 -E -mtune=x86-64-v4 2>&1 | FileCheck %s --check-prefix=INVALID
+// INVALID: error: unknown target CPU '{{.*}}'

clang/test/Misc/target-invalid-cpu-note.c

@@ -25,7 +25,7 @@
 // X86-SAME: athlon, athlon-tbird, athlon-xp, athlon-mp, athlon-4, k8, athlon64,
 // X86-SAME: athlon-fx, opteron, k8-sse3, athlon64-sse3, opteron-sse3, amdfam10,
 // X86-SAME: barcelona, btver1, btver2, bdver1, bdver2, bdver3, bdver4, znver1, znver2,
-// X86-SAME: x86-64, geode
+// X86-SAME: x86-64, x86-64-v2, x86-64-v3, x86-64-v4, geode{{$}}
 
 // RUN: not %clang_cc1 -triple x86_64--- -target-cpu not-a-cpu -fsyntax-only %s 2>&1 | FileCheck %s --check-prefix X86_64
 // X86_64: error: unknown target CPU 'not-a-cpu'
@@ -35,7 +35,8 @@
 // X86_64-SAME: core-avx2, broadwell, skylake, skylake-avx512, skx, cascadelake, cooperlake, cannonlake,
 // X86_64-SAME: icelake-client, icelake-server, tigerlake, sapphirerapids, knl, knm, k8, athlon64, athlon-fx, opteron, k8-sse3,
 // X86_64-SAME: athlon64-sse3, opteron-sse3, amdfam10, barcelona, btver1,
-// X86_64-SAME: btver2, bdver1, bdver2, bdver3, bdver4, znver1, znver2, x86-64
+// X86_64-SAME: btver2, bdver1, bdver2, bdver3, bdver4, znver1, znver2,
+// X86_64-SAME: x86-64, x86-64-v2, x86-64-v3, x86-64-v4{{$}}
 
 // RUN: not %clang_cc1 -triple i386--- -tune-cpu not-a-cpu -fsyntax-only %s 2>&1 | FileCheck %s --check-prefix TUNE_X86
 // TUNE_X86: error: unknown target CPU 'not-a-cpu'
@@ -49,7 +50,7 @@
 // TUNE_X86-SAME: athlon, athlon-tbird, athlon-xp, athlon-mp, athlon-4, k8, athlon64,
 // TUNE_X86-SAME: athlon-fx, opteron, k8-sse3, athlon64-sse3, opteron-sse3, amdfam10,
 // TUNE_X86-SAME: barcelona, btver1, btver2, bdver1, bdver2, bdver3, bdver4, znver1, znver2,
-// TUNE_X86-SAME: x86-64, geode
+// TUNE_X86-SAME: x86-64, geode{{$}}
 
 // RUN: not %clang_cc1 -triple x86_64--- -tune-cpu not-a-cpu -fsyntax-only %s 2>&1 | FileCheck %s --check-prefix TUNE_X86_64
 // TUNE_X86_64: error: unknown target CPU 'not-a-cpu'
@@ -63,7 +64,7 @@
 // TUNE_X86_64-SAME: athlon, athlon-tbird, athlon-xp, athlon-mp, athlon-4, k8, athlon64,
 // TUNE_X86_64-SAME: athlon-fx, opteron, k8-sse3, athlon64-sse3, opteron-sse3, amdfam10,
 // TUNE_X86_64-SAME: barcelona, btver1, btver2, bdver1, bdver2, bdver3, bdver4, znver1, znver2,
-// TUNE_X86_64-SAME: x86-64, geode
+// TUNE_X86_64-SAME: x86-64, geode{{$}}
 
 // RUN: not %clang_cc1 -triple nvptx--- -target-cpu not-a-cpu -fsyntax-only %s 2>&1 | FileCheck %s --check-prefix NVPTX
 // NVPTX: error: unknown target CPU 'not-a-cpu'

clang/test/Preprocessor/predefined-arch-macros-x86.c

@@ -0,0 +1,54 @@
+// RUN: %clang -target x86_64 -march=x86-64 -E -dM %s > %tv1
+// RUN: FileCheck %s --check-prefix=X86_64_V1 < %tv1
+
+// X86_64_V1: #define __MMX__ 1
+// X86_64_V1: #define __SSE2_MATH__ 1
+// X86_64_V1: #define __SSE2__ 1
+// X86_64_V1: #define __SSE_MATH__ 1
+// X86_64_V1: #define __SSE__ 1
+// X86_64_V1: #define __amd64 1
+// X86_64_V1: #define __amd64__ 1
+// X86_64_V1: #define __k8 1
+// X86_64_V1: #define __k8__ 1
+// X86_64_V1: #define __x86_64 1
+// X86_64_V1: #define __x86_64__ 1
+
+// RUN: %clang -target x86_64 -march=x86-64-v2 -E -dM %s > %tv2
+// RUN: diff %tv1 %tv2 > %t.txt || true
+// RUN: FileCheck %s --check-prefix=X86_64_V2 < %t.txt
+
+/// v2 is close to Nehalem.
+// X86_64_V2:      #define __GCC_HAVE_SYNC_COMPARE_AND_SWAP_16 1
+// X86_64_V2:      #define __LAHF_SAHF__ 1
+// X86_64_V2:      #define __POPCNT__ 1
+// X86_64_V2:      #define __SSE3__ 1
+// X86_64_V2-NEXT: #define __SSE4_1__ 1
+// X86_64_V2-NEXT: #define __SSE4_2__ 1
+// X86_64_V2:      #define __SSSE3__ 1
+
+/// v3 is close to Haswell.
+// RUN: %clang -target x86_64 -march=x86-64-v3 -E -dM %s > %tv3
+// RUN: diff %tv2 %tv3 > %t.txt || true
+// RUN: FileCheck %s --check-prefix=X86_64_V3 < %t.txt
+
+// X86_64_V3:      #define __AVX2__ 1
+// X86_64_V3-NEXT: #define __AVX__ 1
+// X86_64_V3:      #define __BMI2__ 1
+// X86_64_V3-NEXT: #define __BMI__ 1
+// X86_64_V3:      #define __F16C__ 1
+// X86_64_V3:      #define __FMA__ 1
+// X86_64_V3:      #define __LZCNT__ 1
+// X86_64_V3:      #define __MOVBE__ 1
+// X86_64_V3:      #define __XSAVE__ 1
+
+/// v4 is close to the AVX-512 level implemented by Xeon Scalable Processors.
+// RUN: %clang -target x86_64 -march=x86-64-v4 -E -dM %s > %tv4
+// RUN: diff %tv3 %tv4 > %t.txt || true
+// RUN: FileCheck %s --check-prefix=X86_64_V4 < %t.txt
+
+// X86_64_V4:      #define __AVX512BW__ 1
+// X86_64_V4-NEXT: #define __AVX512CD__ 1
+// X86_64_V4-NEXT: #define __AVX512DQ__ 1
+// X86_64_V4-NEXT: #define __AVX512F__ 1
+// X86_64_V4-NEXT: #define __AVX512VL__ 1
+// X86_64_V4-NOT:  #define __AVX512{{.*}}

clang/test/Preprocessor/predefined-arch-macros.c

@@ -2263,21 +2263,6 @@
 // CHECK_X86_64_M32: #define __k8__ 1
 // CHECK_X86_64_M32: #define i386 1
 
-// RUN: %clang -march=x86-64 -m64 -E -dM %s -o - 2>&1 \
-// RUN:     -target i386-unknown-linux \
-// RUN:   | FileCheck -match-full-lines %s -check-prefix=CHECK_X86_64_M64
-// CHECK_X86_64_M64: #define __MMX__ 1
-// CHECK_X86_64_M64: #define __SSE2_MATH__ 1
-// CHECK_X86_64_M64: #define __SSE2__ 1
-// CHECK_X86_64_M64: #define __SSE_MATH__ 1
-// CHECK_X86_64_M64: #define __SSE__ 1
-// CHECK_X86_64_M64: #define __amd64 1
-// CHECK_X86_64_M64: #define __amd64__ 1
-// CHECK_X86_64_M64: #define __k8 1
-// CHECK_X86_64_M64: #define __k8__ 1
-// CHECK_X86_64_M64: #define __x86_64 1
-// CHECK_X86_64_M64: #define __x86_64__ 1
-
 // RUN: %clang -march=k8 -m32 -E -dM %s -o - 2>&1 \
 // RUN:     -target i386-unknown-linux \
 // RUN:   | FileCheck -match-full-lines %s -check-prefix=CHECK_K8_M32

clang/test/Sema/builtin-cpu-supports.c

@@ -15,6 +15,11 @@ int main() {
 
   if (__builtin_cpu_is("int")) // expected-error {{invalid cpu name for builtin}}
     a("intel");
+
+  (void)__builtin_cpu_is("x86-64");    // expected-error {{invalid cpu name for builtin}}
+  (void)__builtin_cpu_is("x86-64-v2"); // expected-error {{invalid cpu name for builtin}}
+  (void)__builtin_cpu_is("x86-64-v3"); // expected-error {{invalid cpu name for builtin}}
+  (void)__builtin_cpu_is("x86-64-v4"); // expected-error {{invalid cpu name for builtin}}
 #else
   if (__builtin_cpu_supports("vsx")) // expected-error {{use of unknown builtin}}
     a("vsx");

llvm/docs/ReleaseNotes.rst

@@ -105,7 +105,7 @@ During this release ...
 * The 'mpx' feature was removed from the backend. It had been removed from clang
   frontend in 10.0. Mention of the 'mpx' feature in an IR file will print a
   message to stderr, but IR should still compile.
-* Support for -march=sapphirerapids was added.
+* Support for ``-march=sapphirerapids`` and ``-march=x86-64-v[234]`` has been added.
 * The assembler now has support for {disp32} and {disp8} pseudo prefixes for
   controlling displacement size for memory operands and jump displacements. The
   assembler also supports the .d32 and .d8 mnemonic suffixes to do the same.

llvm/include/llvm/Support/X86TargetParser.h

@@ -121,17 +121,24 @@ enum CPUKind {
   CK_ZNVER1,
   CK_ZNVER2,
   CK_x86_64,
+  CK_x86_64_v2,
+  CK_x86_64_v3,
+  CK_x86_64_v4,
   CK_Geode,
 };
 
 /// Parse \p CPU string into a CPUKind. Will only accept 64-bit capable CPUs if
 /// \p Only64Bit is true.
 CPUKind parseArchX86(StringRef CPU, bool Only64Bit = false);
+CPUKind parseTuneCPU(StringRef CPU, bool Only64Bit = false);
 
 /// Provide a list of valid CPU names. If \p Only64Bit is true, the list will
 /// only contain 64-bit capable CPUs.
 void fillValidCPUArchList(SmallVectorImpl<StringRef> &Values,
                           bool Only64Bit = false);
+/// Provide a list of valid -mtune names.
+void fillValidTuneCPUList(SmallVectorImpl<StringRef> &Values,
+                          bool Only64Bit = false);
 
 /// Get the key feature prioritizing target multiversioning.
 ProcessorFeatures getKeyFeature(CPUKind Kind);

llvm/lib/Support/X86TargetParser.cpp

@@ -137,6 +137,15 @@ constexpr FeatureBitset FeaturesNocona =
 
 // Basic 64-bit capable CPU.
 constexpr FeatureBitset FeaturesX86_64 = FeaturesPentium4 | Feature64BIT;
+constexpr FeatureBitset FeaturesX86_64_V2 = FeaturesX86_64 | FeatureSAHF |
+                                            FeaturePOPCNT | FeatureSSE4_2 |
+                                            FeatureCMPXCHG16B;
+constexpr FeatureBitset FeaturesX86_64_V3 =
+    FeaturesX86_64_V2 | FeatureAVX2 | FeatureBMI | FeatureBMI2 | FeatureF16C |
+    FeatureFMA | FeatureLZCNT | FeatureMOVBE | FeatureXSAVE;
+constexpr FeatureBitset FeaturesX86_64_V4 = FeaturesX86_64_V3 |
+                                            FeatureAVX512BW | FeatureAVX512CD |
+                                            FeatureAVX512DQ | FeatureAVX512VL;
 
 // Intel Core CPUs
 constexpr FeatureBitset FeaturesCore2 =
@@ -383,10 +392,15 @@ constexpr ProcInfo Processors[] = {
   { {"znver2"}, CK_ZNVER2, FEATURE_AVX2, FeaturesZNVER2 },
   // Generic 64-bit processor.
   { {"x86-64"}, CK_x86_64, ~0U, FeaturesX86_64 },
+  { {"x86-64-v2"}, CK_x86_64_v2, ~0U, FeaturesX86_64_V2 },
+  { {"x86-64-v3"}, CK_x86_64_v3, ~0U, FeaturesX86_64_V3 },
+  { {"x86-64-v4"}, CK_x86_64_v4, ~0U, FeaturesX86_64_V4 },
   // Geode processors.
   { {"geode"}, CK_Geode, ~0U, FeaturesGeode },
 };
 
+constexpr const char *NoTuneList[] = {"x86-64-v2", "x86-64-v3", "x86-64-v4"};
+
 X86::CPUKind llvm::X86::parseArchX86(StringRef CPU, bool Only64Bit) {
   for (const auto &P : Processors)
     if (P.Name == CPU && (P.Features[FEATURE_64BIT] || !Only64Bit))
@@ -395,13 +409,27 @@ X86::CPUKind llvm::X86::parseArchX86(StringRef CPU, bool Only64Bit) {
   return CK_None;
 }
 
+X86::CPUKind llvm::X86::parseTuneCPU(StringRef CPU, bool Only64Bit) {
+  if (llvm::is_contained(NoTuneList, CPU))
+    return CK_None;
+  return parseArchX86(CPU, Only64Bit);
+}
+
 void llvm::X86::fillValidCPUArchList(SmallVectorImpl<StringRef> &Values,
                                      bool Only64Bit) {
   for (const auto &P : Processors)
     if (!P.Name.empty() && (P.Features[FEATURE_64BIT] || !Only64Bit))
       Values.emplace_back(P.Name);
 }
 
+void llvm::X86::fillValidTuneCPUList(SmallVectorImpl<StringRef> &Values,
+                                     bool Only64Bit) {
+  for (const ProcInfo &P : Processors)
+    if (!P.Name.empty() && (P.Features[FEATURE_64BIT] || !Only64Bit) &&
+        !llvm::is_contained(NoTuneList, P.Name))
+      Values.emplace_back(P.Name);
+}
+
 ProcessorFeatures llvm::X86::getKeyFeature(X86::CPUKind Kind) {
   // FIXME: Can we avoid a linear search here? The table might be sorted by
   // CPUKind so we could binary search?