[llvm] 0fac1c1 - [X86] Allow Yz inline assembly constraint to choose ymm0 or zmm0 when avx/avx512 are enabled and type is 256 or 512 bits
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Tue May 5 21:12:45 PDT 2020
Author: Craig Topper
Date: 2020-05-05T21:12:30-07:00
New Revision: 0fac1c19128106b6d65d349a5ed22fedacab520c
URL: https://github.com/llvm/llvm-project/commit/0fac1c19128106b6d65d349a5ed22fedacab520c
DIFF: https://github.com/llvm/llvm-project/commit/0fac1c19128106b6d65d349a5ed22fedacab520c.diff
LOG: [X86] Allow Yz inline assembly constraint to choose ymm0 or zmm0 when avx/avx512 are enabled and type is 256 or 512 bits
gcc supports selecting ymm0/zmm0 for the Yz constraint when used with 256 or 512 bit vector types.
Fixes PR45806
Differential Revision: https://reviews.llvm.org/D79448
Added:
Modified:
clang/lib/Basic/Targets/X86.cpp
clang/test/CodeGen/x86-inline-asm-v-constraint.c
llvm/lib/Target/X86/X86ISelLowering.cpp
llvm/test/CodeGen/X86/inline-asm-avx-v-constraint.ll
llvm/test/CodeGen/X86/inline-asm-avx512f-v-constraint.ll
Removed:
################################################################################
diff --git a/clang/lib/Basic/Targets/X86.cpp b/clang/lib/Basic/Targets/X86.cpp
index 747d07692fee..b0fecd6fd1c1 100644
--- a/clang/lib/Basic/Targets/X86.cpp
+++ b/clang/lib/Basic/Targets/X86.cpp
@@ -1898,8 +1898,14 @@ bool X86TargetInfo::validateOperandSize(const llvm::StringMap<bool> &FeatureMap,
return Size <= 64;
case 'z':
case '0':
- // XMM0
- if (FeatureMap.lookup("sse"))
+ // XMM0/YMM/ZMM0
+ if (FeatureMap.lookup("avx512f"))
+ // ZMM0 can be used if target supports AVX512F.
+ return Size <= 512U;
+ else if (FeatureMap.lookup("avx"))
+ // YMM0 can be used if target supports AVX.
+ return Size <= 256U;
+ else if (FeatureMap.lookup("sse"))
return Size <= 128U;
return false;
case 'i':
diff --git a/clang/test/CodeGen/x86-inline-asm-v-constraint.c b/clang/test/CodeGen/x86-inline-asm-v-constraint.c
index d335e4b6a0f7..215cccfa443e 100644
--- a/clang/test/CodeGen/x86-inline-asm-v-constraint.c
+++ b/clang/test/CodeGen/x86-inline-asm-v-constraint.c
@@ -28,3 +28,28 @@ __m512 testZMM(__m512 _zmm0, __m512 _zmm1) {
#endif
return _zmm0;
}
+
+// SSE: call <4 x float> asm "pcmpeqd $0, $0", "=^Yz,~{dirflag},~{fpsr},~{flags}"()
+__m128 testXMM0(void) {
+ __m128 xmm0;
+ __asm__("pcmpeqd %0, %0" :"=Yz"(xmm0));
+ return xmm0;
+}
+
+// AVX: call <8 x float> asm "vpcmpeqd $0, $0, $0", "=^Yz,~{dirflag},~{fpsr},~{flags}"()
+__m256 testYMM0(void) {
+ __m256 ymm0;
+#ifdef AVX
+ __asm__("vpcmpeqd %0, %0, %0" :"=Yz"(ymm0));
+#endif
+ return ymm0;
+}
+
+// AVX512: call <16 x float> asm "vpternlogd $$255, $0, $0, $0", "=^Yz,~{dirflag},~{fpsr},~{flags}"()
+__m512 testZMM0(void) {
+ __m512 zmm0;
+#ifdef AVX512
+ __asm__("vpternlogd $255, %0, %0, %0" :"=Yz"(zmm0));
+#endif
+ return zmm0;
+}
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 57db959a6743..35c6d24ff80a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -48046,7 +48046,9 @@ TargetLowering::ConstraintWeight
// XMM0
case 'z':
case '0':
- if ((type->getPrimitiveSizeInBits() == 128) && Subtarget.hasSSE1())
+ if (((type->getPrimitiveSizeInBits() == 128) && Subtarget.hasSSE1()) ||
+ ((type->getPrimitiveSizeInBits() == 256) && Subtarget.hasAVX()) ||
+ ((type->getPrimitiveSizeInBits() == 512) && Subtarget.hasAVX512()))
return CW_SpecificReg;
return CW_Invalid;
// Conditional OpMask regs (AVX512)
@@ -48496,6 +48498,8 @@ X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
if (Subtarget.hasAVX())
return std::make_pair(0U, &X86::VR256RegClass);
break;
+ case MVT::v64i8:
+ case MVT::v32i16:
case MVT::v8f64:
case MVT::v16f32:
case MVT::v16i32:
@@ -48521,7 +48525,42 @@ X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
case 'z':
case '0':
if (!Subtarget.hasSSE1()) break;
- return std::make_pair(X86::XMM0, &X86::VR128RegClass);
+ switch (VT.SimpleTy) {
+ default: break;
+ // Scalar SSE types.
+ case MVT::f32:
+ case MVT::i32:
+ return std::make_pair(X86::XMM0, &X86::FR32RegClass);
+ case MVT::f64:
+ case MVT::i64:
+ return std::make_pair(X86::XMM0, &X86::FR64RegClass);
+ case MVT::f128:
+ case MVT::v16i8:
+ case MVT::v8i16:
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v4f32:
+ case MVT::v2f64:
+ return std::make_pair(X86::XMM0, &X86::VR128RegClass);
+ // AVX types.
+ case MVT::v32i8:
+ case MVT::v16i16:
+ case MVT::v8i32:
+ case MVT::v4i64:
+ case MVT::v8f32:
+ case MVT::v4f64:
+ if (Subtarget.hasAVX())
+ return std::make_pair(X86::YMM0, &X86::VR256RegClass);
+ break;
+ case MVT::v8f64:
+ case MVT::v16f32:
+ case MVT::v16i32:
+ case MVT::v8i64:
+ if (Subtarget.hasAVX512())
+ return std::make_pair(X86::ZMM0, &X86::VR512_0_15RegClass);
+ break;
+ }
+ break;
case 'k':
// This register class doesn't allocate k0 for masked vector operation.
if (Subtarget.hasAVX512()) {
diff --git a/llvm/test/CodeGen/X86/inline-asm-avx-v-constraint.ll b/llvm/test/CodeGen/X86/inline-asm-avx-v-constraint.ll
index 2c8de16fd372..e5b2b1905cb8 100644
--- a/llvm/test/CodeGen/X86/inline-asm-avx-v-constraint.ll
+++ b/llvm/test/CodeGen/X86/inline-asm-avx-v-constraint.ll
@@ -134,3 +134,13 @@ entry:
ret <8 x float> %0
}
+define <8 x float> @testYMM0() {
+; CHECK: vpcmpeqd %ymm0, %ymm0, %ymm0
+entry:
+ %ymm0 = alloca <8 x float>, align 32
+ %0 = call <8 x float> asm "vpcmpeqd $0, $0, $0", "=^Yz,~{dirflag},~{fpsr},~{flags}"()
+ store <8 x float> %0, <8 x float>* %ymm0, align 32
+ %1 = load <8 x float>, <8 x float>* %ymm0, align 32
+ ret <8 x float> %1
+}
+
diff --git a/llvm/test/CodeGen/X86/inline-asm-avx512f-v-constraint.ll b/llvm/test/CodeGen/X86/inline-asm-avx512f-v-constraint.ll
index 019973ba935b..a71255052760 100644
--- a/llvm/test/CodeGen/X86/inline-asm-avx512f-v-constraint.ll
+++ b/llvm/test/CodeGen/X86/inline-asm-avx512f-v-constraint.ll
@@ -70,3 +70,12 @@ entry:
ret <16 x float> %0
}
+define <16 x float> @testZMM0() {
+entry:
+; CHECK: vpternlogd $255, %zmm0, %zmm0, %zmm0
+ %zmm0 = alloca <16 x float>, align 64
+ %0 = call <16 x float> asm "vpternlogd $$255, $0, $0, $0", "=^Yz,~{dirflag},~{fpsr},~{flags}"()
+ store <16 x float> %0, <16 x float>* %zmm0, align 64
+ %1 = load <16 x float>, <16 x float>* %zmm0, align 64
+ ret <16 x float> %1
+}
More information about the llvm-commits
mailing list