[x86] Simplify vector selection if condition value type matches vselect value type and true value is all ones or false value is all zeros.

This transformation worked if selector is produced by SETCC, however SETCC is needed only if we consider to swap operands. So I replaced SETCC check for this case.
Added tests for vselect of <X x i1> values.

llvm-svn: 220777

3 files changed, 56 insertions, 12 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 656c1de..d8ffc36 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -22479,22 +22479,22 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
       return DAG.getNode(Opc, DL, VT, LHS, RHS);
   }
 
-  // Simplify vector selection if the selector will be produced by CMPP*/PCMP*.
-  if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
-      // Check if SETCC has already been promoted
-      TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT &&
-      // Check that condition value type matches vselect operand type
-      CondVT == VT) { 
-
+  // Simplify vector selection if condition value type matches vselect
+  // operand type
+  if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
     assert(Cond.getValueType().isVector() &&
            "vector select expects a vector selector!");
 
     bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
     bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
 
-    if (!TValIsAllOnes && !FValIsAllZeros) {
-      // Try invert the condition if true value is not all 1s and false value
-      // is not all 0s.
+    // Try invert the condition if true value is not all 1s and false value
+    // is not all 0s.
+    if (!TValIsAllOnes && !FValIsAllZeros &&
+        // Check if the selector will be produced by CMPP*/PCMP*
+        Cond.getOpcode() == ISD::SETCC &&
+        // Check if SETCC has already been promoted
+        TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
       bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
       bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
 
diff --git a/llvm/test/CodeGen/X86/avx512-select.ll b/llvm/test/CodeGen/X86/avx512-select.ll
index e222113..0dbf286d 100644
--- a/llvm/test/CodeGen/X86/avx512-select.ll
+++ b/llvm/test/CodeGen/X86/avx512-select.ll
@@ -48,3 +48,47 @@ define <16 x double> @select04(<16 x double> %a, <16 x double> %b) {
   %sel = select <16 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false>, <16 x double> %a, <16 x double> %b
   ret <16 x double> %sel
 }
+
+; CHECK-LABEL: select05
+; CHECK: kmovw   %esi, %k0
+; CHECK-NEXT: kmovw   %edi, %k1
+; CHECK-NEXT: korw    %k1, %k0, %k0
+; CHECK-NEXT: kmovw   %k0, %eax
+define i8 @select05(i8 %a.0, i8 %m) {
+  %mask = bitcast i8 %m to <8 x i1>
+  %a = bitcast i8 %a.0 to <8 x i1>
+  %r = select <8 x i1> %mask, <8 x i1> <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>, <8 x i1> %a
+  %res = bitcast <8 x i1> %r to i8
+  ret i8 %res;
+}
+
+; CHECK-LABEL: select06
+; CHECK: kmovw   %esi, %k0
+; CHECK-NEXT: kmovw   %edi, %k1
+; CHECK-NEXT: kandw    %k1, %k0, %k0
+; CHECK-NEXT: kmovw   %k0, %eax
+define i8 @select06(i8 %a.0, i8 %m) {
+  %mask = bitcast i8 %m to <8 x i1>
+  %a = bitcast i8 %a.0 to <8 x i1>
+  %r = select <8 x i1> %mask, <8 x i1> %a, <8 x i1> zeroinitializer
+  %res = bitcast <8 x i1> %r to i8
+  ret i8 %res;
+}
+
+; CHECK-LABEL: select07
+; CHECK-DAG:  kmovw   %edx, %k0
+; CHECK-DAG:  kmovw   %edi, %k1
+; CHECK-DAG:  kmovw   %esi, %k2
+; CHECK: kandw %k0, %k1, %k1
+; CHECK-NEXT: knotw    %k0, %k0
+; CHECK-NEXT: kandw    %k0, %k2, %k0
+; CHECK-NEXT: korw %k0, %k1, %k0
+; CHECK-NEXT: kmovw   %k0, %eax
+define i8 @select07(i8 %a.0, i8 %b.0, i8 %m) {
+  %mask = bitcast i8 %m to <8 x i1>
+  %a = bitcast i8 %a.0 to <8 x i1>
+  %b = bitcast i8 %b.0 to <8 x i1>
+  %r = select <8 x i1> %mask, <8 x i1> %a, <8 x i1> %b
+  %res = bitcast <8 x i1> %r to i8
+  ret i8 %res;
+}
diff --git a/llvm/test/CodeGen/X86/vselect-avx.ll b/llvm/test/CodeGen/X86/vselect-avx.ll
index 2d7ccf3..7926b0c 100644
--- a/llvm/test/CodeGen/X86/vselect-avx.ll
+++ b/llvm/test/CodeGen/X86/vselect-avx.ll
@@ -14,8 +14,8 @@ target triple = "x86_64-apple-macosx"
 ; <rdar://problem/18675020>
 
 ; CHECK-LABEL: test:
-; CHECK: vmovdqa {{.*#+}} xmm1 = [65533,124,125,14807]
-; CHECK: vmovdqa {{.*#+}} xmm1 = [65535,0,0,65535]
+; CHECK: vmovdqa {{.*#+}} xmm0 = [65535,0,0,65535]
+; CHECK: vmovdqa {{.*#+}} xmm2 = [65533,124,125,14807]
 ; CHECK: ret
 define void @test(<4 x i16>* %a, <4 x i16>* %b) {
 body: