[X86] Support folding in a k-register OR when creating KORTEST from scalar compare of a bitcast from vXi1.

This should allow us to remove the kortest intrinsic from IR and use compare+bitcast+or in IR instead.

llvm-svn: 324580

3 files changed, 18 insertions, 18 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 500a1d3ab03..5aaf2417e1b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -18149,7 +18149,15 @@ static SDValue EmitKTEST(SDValue Op0, SDValue Op1, ISD::CondCode CC,
   } else
     return SDValue();
 
-  SDValue KORTEST = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, Op0, Op0);
+  // If the input is an OR, we can combine it's operands into the KORTEST.
+  SDValue LHS = Op0;
+  SDValue RHS = Op0;
+  if (Op0.getOpcode() == ISD::OR && Op0.hasOneUse( && Op0.hasOneUse())) {
+    LHS = Op0.getOperand(0);
+    RHS = Op0.getOperand(1);
+  }
+
+  SDValue KORTEST = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
   return getSETCC(X86CC, KORTEST, dl, DAG);
 }
 
diff --git a/llvm/test/CodeGen/X86/avx512-mask-op.ll b/llvm/test/CodeGen/X86/avx512-mask-op.ll
index 4766c8b1560..6953cd9ed47 100644
--- a/llvm/test/CodeGen/X86/avx512-mask-op.ll
+++ b/llvm/test/CodeGen/X86/avx512-mask-op.ll
@@ -583,8 +583,7 @@ define void @test7(<8 x i1> %mask)  {
 ; SKX-NEXT:    vpmovw2m %xmm0, %k0
 ; SKX-NEXT:    movb $85, %al
 ; SKX-NEXT:    kmovd %eax, %k1
-; SKX-NEXT:    korb %k1, %k0, %k0
-; SKX-NEXT:    kortestb %k0, %k0
+; SKX-NEXT:    kortestb %k1, %k0
 ; SKX-NEXT:    retq
 ;
 ; AVX512BW-LABEL: test7:
@@ -606,8 +605,7 @@ define void @test7(<8 x i1> %mask)  {
 ; AVX512DQ-NEXT:    vptestmq %zmm0, %zmm0, %k0
 ; AVX512DQ-NEXT:    movb $85, %al
 ; AVX512DQ-NEXT:    kmovw %eax, %k1
-; AVX512DQ-NEXT:    korb %k1, %k0, %k0
-; AVX512DQ-NEXT:    kortestb %k0, %k0
+; AVX512DQ-NEXT:    kortestb %k1, %k0
 ; AVX512DQ-NEXT:    vzeroupper
 ; AVX512DQ-NEXT:    retq
 allocas:
@@ -1787,8 +1785,7 @@ define void @ktest_2(<32 x float> %in, float * %base) {
 ; SKX-NEXT:    vcmpltps %zmm3, %zmm0, %k1
 ; SKX-NEXT:    vcmpltps %zmm2, %zmm1, %k2
 ; SKX-NEXT:    kunpckwd %k1, %k2, %k1
-; SKX-NEXT:    kord %k1, %k0, %k0
-; SKX-NEXT:    kortestd %k0, %k0
+; SKX-NEXT:    kortestd %k1, %k0
 ; SKX-NEXT:    je LBB43_2
 ; SKX-NEXT:  ## %bb.1: ## %L1
 ; SKX-NEXT:    vmovaps %zmm0, (%rdi)
@@ -1813,8 +1810,7 @@ define void @ktest_2(<32 x float> %in, float * %base) {
 ; AVX512BW-NEXT:    vcmpltps %zmm3, %zmm0, %k1
 ; AVX512BW-NEXT:    vcmpltps %zmm2, %zmm1, %k2
 ; AVX512BW-NEXT:    kunpckwd %k1, %k2, %k1
-; AVX512BW-NEXT:    kord %k1, %k0, %k0
-; AVX512BW-NEXT:    kortestd %k0, %k0
+; AVX512BW-NEXT:    kortestd %k1, %k0
 ; AVX512BW-NEXT:    je LBB43_2
 ; AVX512BW-NEXT:  ## %bb.1: ## %L1
 ; AVX512BW-NEXT:    vmovaps %zmm0, (%rdi)
diff --git a/llvm/test/CodeGen/X86/avx512-schedule.ll b/llvm/test/CodeGen/X86/avx512-schedule.ll
index 69d0a62f281..1055570de07 100755
--- a/llvm/test/CodeGen/X86/avx512-schedule.ll
+++ b/llvm/test/CodeGen/X86/avx512-schedule.ll
@@ -7030,8 +7030,7 @@ define void @vcmp_test7(<8 x i1> %mask)  {
 ; GENERIC-NEXT:    vpmovw2m %xmm0, %k0 # sched: [1:0.33]
 ; GENERIC-NEXT:    movb $85, %al # sched: [1:0.33]
 ; GENERIC-NEXT:    kmovd %eax, %k1 # sched: [1:0.33]
-; GENERIC-NEXT:    korb %k1, %k0, %k0 # sched: [1:1.00]
-; GENERIC-NEXT:    kortestb %k0, %k0 # sched: [1:1.00]
+; GENERIC-NEXT:    kortestb %k1, %k0 # sched: [1:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; SKX-LABEL: vcmp_test7:
@@ -7040,8 +7039,7 @@ define void @vcmp_test7(<8 x i1> %mask)  {
 ; SKX-NEXT:    vpmovw2m %xmm0, %k0 # sched: [1:1.00]
 ; SKX-NEXT:    movb $85, %al # sched: [1:0.25]
 ; SKX-NEXT:    kmovd %eax, %k1 # sched: [1:1.00]
-; SKX-NEXT:    korb %k1, %k0, %k0 # sched: [1:1.00]
-; SKX-NEXT:    kortestb %k0, %k0 # sched: [3:1.00]
+; SKX-NEXT:    kortestb %k1, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 allocas:
   %a= or <8 x i1> %mask, <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>
@@ -7683,8 +7681,7 @@ define void @ktest_2(<32 x float> %in, float * %base) {
 ; GENERIC-NEXT:    vcmpltps %zmm3, %zmm0, %k1 # sched: [3:1.00]
 ; GENERIC-NEXT:    vcmpltps %zmm2, %zmm1, %k2 # sched: [3:1.00]
 ; GENERIC-NEXT:    kunpckwd %k1, %k2, %k1 # sched: [1:1.00]
-; GENERIC-NEXT:    kord %k1, %k0, %k0 # sched: [1:1.00]
-; GENERIC-NEXT:    kortestd %k0, %k0 # sched: [1:1.00]
+; GENERIC-NEXT:    kortestd %k1, %k0 # sched: [1:1.00]
 ; GENERIC-NEXT:    je .LBB411_2 # sched: [1:1.00]
 ; GENERIC-NEXT:  # %bb.1: # %L1
 ; GENERIC-NEXT:    vmovaps %zmm0, (%rdi) # sched: [1:1.00]
@@ -7703,14 +7700,13 @@ define void @ktest_2(<32 x float> %in, float * %base) {
 ; SKX-NEXT:    vmovups 64(%rdi), %zmm3 # sched: [8:0.50]
 ; SKX-NEXT:    vcmpltps %zmm0, %zmm2, %k1 # sched: [3:1.00]
 ; SKX-NEXT:    vcmpltps %zmm1, %zmm3, %k2 # sched: [3:1.00]
+; SKX-NEXT:    kunpckwd %k1, %k2, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    vmovups 68(%rdi), %zmm2 {%k2} {z} # sched: [8:0.50]
 ; SKX-NEXT:    vmovups 4(%rdi), %zmm3 {%k1} {z} # sched: [8:0.50]
-; SKX-NEXT:    kunpckwd %k1, %k2, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    vcmpltps %zmm3, %zmm0, %k1 # sched: [3:1.00]
 ; SKX-NEXT:    vcmpltps %zmm2, %zmm1, %k2 # sched: [3:1.00]
 ; SKX-NEXT:    kunpckwd %k1, %k2, %k1 # sched: [3:1.00]
-; SKX-NEXT:    kord %k1, %k0, %k0 # sched: [1:1.00]
-; SKX-NEXT:    kortestd %k0, %k0 # sched: [3:1.00]
+; SKX-NEXT:    kortestd %k1, %k0 # sched: [3:1.00]
 ; SKX-NEXT:    je .LBB411_2 # sched: [1:0.50]
 ; SKX-NEXT:  # %bb.1: # %L1
 ; SKX-NEXT:    vmovaps %zmm0, (%rdi) # sched: [1:1.00]