[llvm] r274736 - [AVX512] Zero extend the result of vpcmpeq/vpcmpgt and similar intrinsics in the autoupgrade code. This currently results in worse codegen but is needed for correctness.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 6 23:11:07 PDT 2016
Author: ctopper
Date: Thu Jul 7 01:11:07 2016
New Revision: 274736
URL: http://llvm.org/viewvc/llvm-project?rev=274736&view=rev
Log:
[AVX512] Zero extend the result of vpcmpeq/vpcmpgt and similar intrinsics in the autoupgrade code. This currently results in worse codegen but is needed for correctness.
Modified:
llvm/trunk/lib/IR/AutoUpgrade.cpp
llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll
Modified: llvm/trunk/lib/IR/AutoUpgrade.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/IR/AutoUpgrade.cpp?rev=274736&r1=274735&r2=274736&view=diff
==============================================================================
--- llvm/trunk/lib/IR/AutoUpgrade.cpp (original)
+++ llvm/trunk/lib/IR/AutoUpgrade.cpp Thu Jul 7 01:11:07 2016
@@ -601,8 +601,9 @@ static Value *upgradeMaskedCompare(IRBui
for (unsigned i = 0; i != NumElts; ++i)
Indices[i] = i;
for (unsigned i = NumElts; i != 8; ++i)
- Indices[i] = NumElts;
- Cmp = Builder.CreateShuffleVector(Cmp, UndefValue::get(Cmp->getType()),
+ Indices[i] = NumElts + i % NumElts;
+ Cmp = Builder.CreateShuffleVector(Cmp,
+ Constant::getNullValue(Cmp->getType()),
Indices);
}
return Builder.CreateBitCast(Cmp, IntegerType::get(CI.getContext(),
Modified: llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll?rev=274736&r1=274735&r2=274736&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll (original)
+++ llvm/trunk/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll Thu Jul 7 01:11:07 2016
@@ -1002,6 +1002,8 @@ define i8 @test_pcmpeq_q_256(<4 x i64> %
; CHECK-LABEL: test_pcmpeq_q_256:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x28,0x29,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.q.256(<4 x i64> %a, <4 x i64> %b, i8 -1)
@@ -1013,6 +1015,8 @@ define i8 @test_mask_pcmpeq_q_256(<4 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x29,0x29,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.q.256(<4 x i64> %a, <4 x i64> %b, i8 %mask)
@@ -1048,6 +1052,8 @@ define i8 @test_pcmpgt_q_256(<4 x i64> %
; CHECK-LABEL: test_pcmpgt_q_256:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x28,0x37,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.256(<4 x i64> %a, <4 x i64> %b, i8 -1)
@@ -1059,6 +1065,8 @@ define i8 @test_mask_pcmpgt_q_256(<4 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x29,0x37,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.256(<4 x i64> %a, <4 x i64> %b, i8 %mask)
@@ -1071,6 +1079,8 @@ define i8 @test_pcmpeq_d_128(<4 x i32> %
; CHECK-LABEL: test_pcmpeq_d_128:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x76,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.d.128(<4 x i32> %a, <4 x i32> %b, i8 -1)
@@ -1082,6 +1092,8 @@ define i8 @test_mask_pcmpeq_d_128(<4 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x76,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.d.128(<4 x i32> %a, <4 x i32> %b, i8 %mask)
@@ -1094,6 +1106,10 @@ define i8 @test_pcmpeq_q_128(<2 x i64> %
; CHECK-LABEL: test_pcmpeq_q_128:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x08,0x29,0xc1]
+; CHECK-NEXT: kshiftlw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0e]
+; CHECK-NEXT: kshiftrw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0e]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.q.128(<2 x i64> %a, <2 x i64> %b, i8 -1)
@@ -1105,6 +1121,10 @@ define i8 @test_mask_pcmpeq_q_128(<2 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0x29,0xc1]
+; CHECK-NEXT: kshiftlw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0e]
+; CHECK-NEXT: kshiftrw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0e]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpeq.q.128(<2 x i64> %a, <2 x i64> %b, i8 %mask)
@@ -1117,6 +1137,8 @@ define i8 @test_pcmpgt_d_128(<4 x i32> %
; CHECK-LABEL: test_pcmpgt_d_128:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf1,0x7d,0x08,0x66,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.128(<4 x i32> %a, <4 x i32> %b, i8 -1)
@@ -1128,6 +1150,8 @@ define i8 @test_mask_pcmpgt_d_128(<4 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf1,0x7d,0x09,0x66,0xc1]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.d.128(<4 x i32> %a, <4 x i32> %b, i8 %mask)
@@ -1140,6 +1164,10 @@ define i8 @test_pcmpgt_q_128(<2 x i64> %
; CHECK-LABEL: test_pcmpgt_q_128:
; CHECK: ## BB#0:
; CHECK-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 ## encoding: [0x62,0xf2,0xfd,0x08,0x37,0xc1]
+; CHECK-NEXT: kshiftlw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0e]
+; CHECK-NEXT: kshiftrw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0e]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.128(<2 x i64> %a, <2 x i64> %b, i8 -1)
@@ -1151,6 +1179,10 @@ define i8 @test_mask_pcmpgt_q_128(<2 x i
; CHECK: ## BB#0:
; CHECK-NEXT: kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
; CHECK-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1} ## encoding: [0x62,0xf2,0xfd,0x09,0x37,0xc1]
+; CHECK-NEXT: kshiftlw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0e]
+; CHECK-NEXT: kshiftrw $14, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0e]
+; CHECK-NEXT: kshiftlw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x32,0xc0,0x0c]
+; CHECK-NEXT: kshiftrw $12, %k0, %k0 ## encoding: [0xc4,0xe3,0xf9,0x30,0xc0,0x0c]
; CHECK-NEXT: kmovw %k0, %eax ## encoding: [0xc5,0xf8,0x93,0xc0]
; CHECK-NEXT: retq ## encoding: [0xc3]
%res = call i8 @llvm.x86.avx512.mask.pcmpgt.q.128(<2 x i64> %a, <2 x i64> %b, i8 %mask)
More information about the llvm-commits
mailing list