From: Elena Demikhovsky <elena.demikhovsky@intel.com>
Date: Mon, 28 Mar 2016 07:47:58 +0000 (+0000)
Subject: AVX-512: Fixed ICMP instruction selection for i1 operands
X-Git-Tag: llvmorg-3.9.0-rc1~10773
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=83f0647d85e37806df88546063bef26e6a9bb01f;p=platform%2Fupstream%2Fllvm.git

AVX-512: Fixed ICMP instruction selection for i1 operands

ICMP instruction selection fails on SKX and KNL for i1 operand.
I use XOR to resolve:
(A == B) is equivalent to (A xor B) == 0

Differential Revision: http://reviews.llvm.org/D18511

llvm-svn: 264566
---

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a06faba..98c020c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -15523,11 +15523,15 @@ SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
       return SetCC;
     }
   }
-  if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
-      (CC == ISD::SETEQ || CC == ISD::SETNE)) {
-
-    ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
-    return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
+  if (Op0.getValueType() == MVT::i1 && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
+    if (isOneConstant(Op1)) {
+      ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
+      return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
+    }
+    if (!isNullConstant(Op1)) {
+      SDValue Xor = DAG.getNode(ISD::XOR, dl, MVT::i1, Op0, Op1);
+      return DAG.getSetCC(dl, VT, Xor, DAG.getConstant(0, dl, MVT::i1), CC);
+    }
   }
 
   bool isFP = Op1.getSimpleValueType().isFloatingPoint();
diff --git a/llvm/test/CodeGen/X86/avx512-cmp.ll b/llvm/test/CodeGen/X86/avx512-cmp.ll
index 6e0d185..68989b1 100644
--- a/llvm/test/CodeGen/X86/avx512-cmp.ll
+++ b/llvm/test/CodeGen/X86/avx512-cmp.ll
@@ -1,8 +1,19 @@
-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl --show-mc-encoding | FileCheck %s
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl  | FileCheck %s --check-prefix=ALL --check-prefix=KNL
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skx  | FileCheck %s --check-prefix=ALL --check-prefix=SKX
 
-; CHECK-LABEL: test1
-; CHECK: vucomisd {{.*}}encoding: [0x62
 define double @test1(double %a, double %b) nounwind {
+; ALL-LABEL: test1:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vucomisd %xmm1, %xmm0
+; ALL-NEXT:    jne LBB0_1
+; ALL-NEXT:    jnp LBB0_2
+; ALL-NEXT:  LBB0_1: ## %l1
+; ALL-NEXT:    vsubsd %xmm1, %xmm0, %xmm0
+; ALL-NEXT:    retq
+; ALL-NEXT:  LBB0_2: ## %l2
+; ALL-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
+; ALL-NEXT:    retq
   %tobool = fcmp une double %a, %b
   br i1 %tobool, label %l1, label %l2
 
@@ -14,9 +25,17 @@ l2:
   ret double %c1
 }
 
-; CHECK-LABEL: test2
-; CHECK: vucomiss {{.*}}encoding: [0x62
 define float @test2(float %a, float %b) nounwind {
+; ALL-LABEL: test2:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vucomiss %xmm0, %xmm1
+; ALL-NEXT:    jbe LBB1_2
+; ALL-NEXT:  ## BB#1: ## %l1
+; ALL-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; ALL-NEXT:    retq
+; ALL-NEXT:  LBB1_2: ## %l2
+; ALL-NEXT:    vaddss %xmm1, %xmm0, %xmm0
+; ALL-NEXT:    retq
   %tobool = fcmp olt float %a, %b
   br i1 %tobool, label %l1, label %l2
 
@@ -29,18 +48,35 @@ l2:
 }
 
 ; FIXME: Can use vcmpeqss and extract from the mask here in AVX512.
-; CHECK-LABEL: test3
-; CHECK: vucomiss {{.*}}encoding: [0x62
 define i32 @test3(float %a, float %b) {
+; ALL-LABEL: test3:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vucomiss %xmm1, %xmm0
+; ALL-NEXT:    setnp %al
+; ALL-NEXT:    sete %cl
+; ALL-NEXT:    andb %al, %cl
+; ALL-NEXT:    movzbl %cl, %eax
+; ALL-NEXT:    retq
 
   %cmp10.i = fcmp oeq float %a, %b
   %conv11.i = zext i1 %cmp10.i to i32
   ret i32 %conv11.i
 }
 
-; CHECK-LABEL: test5
-; CHECK: ret
 define float @test5(float %p) #0 {
+; ALL-LABEL: test5:
+; ALL:       ## BB#0: ## %entry
+; ALL-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; ALL-NEXT:    vucomiss %xmm1, %xmm0
+; ALL-NEXT:    jne LBB3_1
+; ALL-NEXT:    jnp LBB3_2
+; ALL-NEXT:  LBB3_1: ## %if.end
+; ALL-NEXT:    seta %al
+; ALL-NEXT:    movzbl %al, %eax
+; ALL-NEXT:    leaq {{.*}}(%rip), %rcx
+; ALL-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; ALL-NEXT:  LBB3_2: ## %return
+; ALL-NEXT:    retq
 entry:
   %cmp = fcmp oeq float %p, 0.000000e+00
   br i1 %cmp, label %return, label %if.end
@@ -55,21 +91,25 @@ return:                                           ; preds = %if.end, %entry
   ret float %retval.0
 }
 
-; CHECK-LABEL: test6
-; CHECK: cmpl
-; CHECK-NOT: kmov
-; CHECK: ret
 define i32 @test6(i32 %a, i32 %b) {
+; ALL-LABEL: test6:
+; ALL:       ## BB#0:
+; ALL-NEXT:    cmpl %esi, %edi
+; ALL-NEXT:    sete %al
+; ALL-NEXT:    movzbl %al, %eax
+; ALL-NEXT:    retq
   %cmp = icmp eq i32 %a, %b
   %res = zext i1 %cmp to i32
   ret i32 %res
 }
 
-; CHECK-LABEL: test7
-; CHECK: vucomisd
-; CHECK-NOT: kmov
-; CHECK: ret
 define i32 @test7(double %x, double %y) #2 {
+; ALL-LABEL: test7:
+; ALL:       ## BB#0: ## %entry
+; ALL-NEXT:    vucomisd %xmm1, %xmm0
+; ALL-NEXT:    setne %al
+; ALL-NEXT:    movzbl %al, %eax
+; ALL-NEXT:    retq
 entry:
   %0 = fcmp one double %x, %y
   %or = zext i1 %0 to i32
@@ -77,6 +117,16 @@ entry:
 }
 
 define i32 @test8(i32 %a1, i32 %a2, i32 %a3) {
+; ALL-LABEL: test8:
+; ALL:       ## BB#0:
+; ALL-NEXT:    testl %edx, %edx
+; ALL-NEXT:    movl $1, %eax
+; ALL-NEXT:    cmovel %eax, %edx
+; ALL-NEXT:    cmpl $-2147483648, %esi ## imm = 0xFFFFFFFF80000000
+; ALL-NEXT:    cmovnel %edx, %eax
+; ALL-NEXT:    cmpl $-1, %edi
+; ALL-NEXT:    cmovnel %edx, %eax
+; ALL-NEXT:    retq
   %tmp1 = icmp eq i32 %a1, -1
   %tmp2 = icmp eq i32 %a2, -2147483648
   %tmp3 = and i1 %tmp1, %tmp2
@@ -86,11 +136,17 @@ define i32 @test8(i32 %a1, i32 %a2, i32 %a3) {
   ret i32 %res
 }
 
-; CHECK-LABEL: test9
-; CHECK: testb
-; CHECK-NOT: kmov
-; CHECK: ret
 define i32 @test9(i64 %a) {
+; ALL-LABEL: test9:
+; ALL:       ## BB#0:
+; ALL-NEXT:    testb $1, %dil
+; ALL-NEXT:    jne LBB7_2
+; ALL-NEXT:  ## BB#1: ## %A
+; ALL-NEXT:    movl $6, %eax
+; ALL-NEXT:    retq
+; ALL-NEXT:  LBB7_2: ## %B
+; ALL-NEXT:    movl $7, %eax
+; ALL-NEXT:    retq
  %b = and i64 %a, 1
  %cmp10.i = icmp eq i64 %b, 0
  br i1 %cmp10.i, label %A, label %B
@@ -99,3 +155,37 @@ A:
 B:
  ret i32 7
 }
+
+define i32 @test10(i64 %b, i64 %c, i1 %d) {
+; ALL-LABEL: test10:
+; ALL:       ## BB#0:
+; ALL-NEXT:    andl $1, %edx
+; ALL-NEXT:    kmovw %edx, %k0
+; ALL-NEXT:    cmpq %rsi, %rdi
+; ALL-NEXT:    sete %al
+; ALL-NEXT:    andl $1, %eax
+; ALL-NEXT:    kmovw %eax, %k1
+; ALL-NEXT:    korw %k1, %k0, %k1
+; ALL-NEXT:    kxorw %k1, %k0, %k0
+; ALL-NEXT:    kmovw %k0, %eax
+; ALL-NEXT:    andl $1, %eax
+; ALL-NEXT:    testb %al, %al
+; ALL-NEXT:    je LBB8_1
+; ALL-NEXT:  ## BB#2: ## %if.end.i
+; ALL-NEXT:    movl $6, %eax
+; ALL-NEXT:    retq
+; ALL-NEXT:  LBB8_1: ## %if.then.i
+; ALL-NEXT:    movl $5, %eax
+; ALL-NEXT:    retq
+
+  %cmp8.i = icmp eq i64 %b, %c
+  %or1 = or i1 %d, %cmp8.i
+  %xor1 = xor i1 %d, %or1
+  br i1 %xor1, label %if.end.i, label %if.then.i
+
+if.then.i:
+ ret i32 5
+
+if.end.i:
+  ret i32 6
+}