[llvm] r271255 - [AVX512] Fix intrinsic vcvtps2ph lowering.
Igor Breger via llvm-commits
llvm-commits at lists.llvm.org
Tue May 31 01:04:22 PDT 2016
Author: ibreger
Date: Tue May 31 03:04:21 2016
New Revision: 271255
URL: http://llvm.org/viewvc/llvm-project?rev=271255&view=rev
Log:
[AVX512] Fix intrinsic vcvtps2ph lowering.
Differential Revision: http://reviews.llvm.org/D20788
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/lib/Target/X86/X86InstrAVX512.td
llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll
llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics.ll
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=271255&r1=271254&r2=271255&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Tue May 31 03:04:21 2016
@@ -16978,6 +16978,7 @@ static SDValue getVectorMaskingNode(SDVa
case X86ISD::VTRUNC:
case X86ISD::VTRUNCS:
case X86ISD::VTRUNCUS:
+ case ISD::FP_TO_FP16:
// We can't use ISD::VSELECT here because it is not always "Legal"
// for the destination type. For example vpmovqb require only AVX512
// and vselect that can operate on byte element type require BWI
Modified: llvm/trunk/lib/Target/X86/X86InstrAVX512.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86InstrAVX512.td?rev=271255&r1=271254&r2=271255&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86InstrAVX512.td (original)
+++ llvm/trunk/lib/Target/X86/X86InstrAVX512.td Tue May 31 03:04:21 2016
@@ -248,12 +248,12 @@ multiclass AVX512_maskable<bits<8> O, Fo
string AttSrcAsm, string IntelSrcAsm,
dag RHS,
InstrItinClass itin = NoItinerary,
- bit IsCommutable = 0> :
+ bit IsCommutable = 0, SDNode Select = vselect> :
AVX512_maskable_common<O, F, _, Outs, Ins,
!con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
!con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
- (vselect _.KRCWM:$mask, RHS, _.RC:$src0), vselect,
+ (Select _.KRCWM:$mask, RHS, _.RC:$src0), Select,
"$src0 = $dst", itin, IsCommutable>;
// This multiclass generates the unconditional/non-masking, the masking and
@@ -5801,11 +5801,12 @@ let Predicates = [HasAVX512] in {
multiclass avx512_cvtps2ph<X86VectorVTInfo _dest, X86VectorVTInfo _src,
X86MemOperand x86memop> {
defm rr : AVX512_maskable<0x1D, MRMDestReg, _dest ,(outs _dest.RC:$dst),
- (ins _src.RC:$src1, i32u8imm:$src2),
- "vcvtps2ph", "$src2, $src1", "$src1, $src2",
+ (ins _src.RC:$src1, i32u8imm:$src2),
+ "vcvtps2ph", "$src2, $src1", "$src1, $src2",
(X86cvtps2ph (_src.VT _src.RC:$src1),
(i32 imm:$src2),
- (i32 FROUND_CURRENT))>, AVX512AIi8Base;
+ (i32 FROUND_CURRENT)),
+ NoItinerary, 0, X86select>, AVX512AIi8Base;
let hasSideEffects = 0, mayStore = 1 in {
def mr : AVX512AIi8<0x1D, MRMDestMem, (outs),
(ins x86memop:$dst, _src.RC:$src1, i32u8imm:$src2),
@@ -5821,11 +5822,12 @@ multiclass avx512_cvtps2ph<X86VectorVTIn
}
multiclass avx512_cvtps2ph_sae<X86VectorVTInfo _dest, X86VectorVTInfo _src> {
defm rb : AVX512_maskable<0x1D, MRMDestReg, _dest ,(outs _dest.RC:$dst),
- (ins _src.RC:$src1, i32u8imm:$src2),
- "vcvtps2ph", "$src2, {sae}, $src1", "$src1, {sae}, $src2",
+ (ins _src.RC:$src1, i32u8imm:$src2),
+ "vcvtps2ph", "$src2, {sae}, $src1", "$src1, {sae}, $src2",
(X86cvtps2ph (_src.VT _src.RC:$src1),
(i32 imm:$src2),
- (i32 FROUND_NO_EXC))>, EVEX_B, AVX512AIi8Base;
+ (i32 FROUND_NO_EXC)),
+ NoItinerary, 0, X86select>, EVEX_B, AVX512AIi8Base;
}
let Predicates = [HasAVX512] in {
defm VCVTPS2PHZ : avx512_cvtps2ph<v16i16x_info, v16f32_info, f256mem>,
Modified: llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll?rev=271255&r1=271254&r2=271255&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll Tue May 31 03:04:21 2016
@@ -604,13 +604,20 @@ define <16 x float> @test_x86_vcvtph2ps_
declare <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16>, <16 x float>, i16, i32) nounwind readonly
-
-define <16 x i16> @test_x86_vcvtps2ph_256(<16 x float> %a0) {
+define <16 x i16> @test_x86_vcvtps2ph_256(<16 x float> %a0, <16 x i16> %src, i16 %mask, <16 x i16> * %dst) {
; CHECK-LABEL: test_x86_vcvtps2ph_256:
; CHECK: ## BB#0:
-; CHECK-NEXT: vcvtps2ph $2, %zmm0, %ymm0
+; CHECK-NEXT: kmovw %edi, %k1
+; CHECK-NEXT: vcvtps2ph $2, %zmm0, %ymm1 {%k1}
+; CHECK-NEXT: vcvtps2ph $2, %zmm0, %ymm2 {%k1} {z}
+; CHECK-NEXT: vcvtps2ph $2, %zmm0, (%rsi)
+; CHECK-NEXT: vpaddw %ymm1, %ymm2, %ymm0
; CHECK-NEXT: retq
- %res = call <16 x i16> @llvm.x86.avx512.mask.vcvtps2ph.512(<16 x float> %a0, i32 2, <16 x i16> zeroinitializer, i16 -1)
+ %res1 = call <16 x i16> @llvm.x86.avx512.mask.vcvtps2ph.512(<16 x float> %a0, i32 2, <16 x i16> zeroinitializer, i16 -1)
+ %res2 = call <16 x i16> @llvm.x86.avx512.mask.vcvtps2ph.512(<16 x float> %a0, i32 2, <16 x i16> zeroinitializer, i16 %mask)
+ %res3 = call <16 x i16> @llvm.x86.avx512.mask.vcvtps2ph.512(<16 x float> %a0, i32 2, <16 x i16> %src, i16 %mask)
+ store <16 x i16> %res1, <16 x i16> * %dst
+ %res = add <16 x i16> %res2, %res3
ret <16 x i16> %res
}
Modified: llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics.ll?rev=271255&r1=271254&r2=271255&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics.ll Tue May 31 03:04:21 2016
@@ -6931,24 +6931,41 @@ define <8 x float> @test_x86_vcvtph2ps_2
declare <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16>, <8 x float>, i8) nounwind readonly
-define <8 x i16> @test_x86_vcvtps2ph_128(<4 x float> %a0) {
+define <8 x i16> @test_x86_vcvtps2ph_128(<4 x float> %a0, i8 %mask, <8 x i16> %src) {
; CHECK-LABEL: test_x86_vcvtps2ph_128:
; CHECK: ## BB#0:
+; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT: vcvtps2ph $2, %xmm0, %xmm1 {%k1} ## encoding: [0x62,0xf3,0x7d,0x09,0x1d,0xc1,0x02]
+; CHECK-NEXT: vcvtps2ph $2, %xmm0, %xmm2 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0x89,0x1d,0xc2,0x02]
; CHECK-NEXT: vcvtps2ph $2, %xmm0, %xmm0 ## encoding: [0x62,0xf3,0x7d,0x08,0x1d,0xc0,0x02]
+; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xfd,0xc2]
+; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## encoding: [0xc5,0xf1,0xfd,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
- %res = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.128(<4 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 -1)
+ %res1 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.128(<4 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 -1)
+ %res2 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.128(<4 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 %mask)
+ %res3 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.128(<4 x float> %a0, i32 2, <8 x i16> %src, i8 %mask)
+ %res0 = add <8 x i16> %res1, %res2
+ %res = add <8 x i16> %res3, %res0
ret <8 x i16> %res
}
-
declare <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.128(<4 x float>, i32, <8 x i16>, i8) nounwind readonly
-define <8 x i16> @test_x86_vcvtps2ph_256(<8 x float> %a0) {
+define <8 x i16> @test_x86_vcvtps2ph_256(<8 x float> %a0, i8 %mask, <8 x i16> %src) {
; CHECK-LABEL: test_x86_vcvtps2ph_256:
; CHECK: ## BB#0:
+; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; CHECK-NEXT: vcvtps2ph $2, %ymm0, %xmm1 {%k1} ## encoding: [0x62,0xf3,0x7d,0x29,0x1d,0xc1,0x02]
+; CHECK-NEXT: vcvtps2ph $2, %ymm0, %xmm2 {%k1} {z} ## encoding: [0x62,0xf3,0x7d,0xa9,0x1d,0xc2,0x02]
; CHECK-NEXT: vcvtps2ph $2, %ymm0, %xmm0 ## encoding: [0x62,0xf3,0x7d,0x28,0x1d,0xc0,0x02]
+; CHECK-NEXT: vpaddw %xmm2, %xmm0, %xmm0 ## encoding: [0xc5,0xf9,0xfd,0xc2]
+; CHECK-NEXT: vpaddw %xmm0, %xmm1, %xmm0 ## encoding: [0xc5,0xf1,0xfd,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
- %res = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.256(<8 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 -1)
+ %res1 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.256(<8 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 -1)
+ %res2 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.256(<8 x float> %a0, i32 2, <8 x i16> zeroinitializer, i8 %mask)
+ %res3 = call <8 x i16> @llvm.x86.avx512.mask.vcvtps2ph.256(<8 x float> %a0, i32 2, <8 x i16> %src, i8 %mask)
+ %res0 = add <8 x i16> %res1, %res2
+ %res = add <8 x i16> %res3, %res0
ret <8 x i16> %res
}
More information about the llvm-commits
mailing list