AVX-512: Implemented CMOV for 512-bit vectors

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193747 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5dbef0f6fc334d72bc4307e706cdadfb41f42695..7648ddc6818ba53ea0ec15ee9055b61717364a2b 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -15750,6 +15750,9 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case X86::CMOV_V8F32:
   case X86::CMOV_V4F64:
   case X86::CMOV_V4I64:
   case X86::CMOV_V8F32:
   case X86::CMOV_V4F64:
   case X86::CMOV_V4I64:

+  case X86::CMOV_V16F32:
+  case X86::CMOV_V8F64:
+  case X86::CMOV_V8I64:

   case X86::CMOV_GR16:
   case X86::CMOV_GR32:
   case X86::CMOV_RFP32:
   case X86::CMOV_GR16:
   case X86::CMOV_GR32:
   case X86::CMOV_RFP32:


@@ -16633,8 +16636,9 @@ static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
       return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
   }
 
       return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
   }
 

-  if (Subtarget->hasAVX512() && VT.isVector() &&
-      Cond.getValueType().getVectorElementType() == MVT::i1) {
+  EVT CondVT = Cond.getValueType();
+  if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
+      CondVT.getVectorElementType() == MVT::i1) {

     // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
     // lowering on AVX-512. In this case we convert it to
     // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
     // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
     // lowering on AVX-512. In this case we convert it to
     // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.

diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 9e1cdc90571b14b24e0a94573bc5c76902a64d1f..7d10b67bfe6dfeba7af1aa775077457fda18f077 100644 (file)
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -884,6 +884,24 @@ let Uses = [EFLAGS], usesCustomInserter = 1 in {
                     [(set VR256:$dst,
                       (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
                                           EFLAGS)))]>;
                     [(set VR256:$dst,
                       (v4i64 (X86cmov VR256:$t, VR256:$f, imm:$cond,
                                           EFLAGS)))]>;

+  def CMOV_V8I64 : I<0, Pseudo,
+                    (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
+                    "#CMOV_V8I64 PSEUDO!",
+                    [(set VR512:$dst,
+                      (v8i64 (X86cmov VR512:$t, VR512:$f, imm:$cond,
+                                          EFLAGS)))]>;
+  def CMOV_V8F64 : I<0, Pseudo,
+                    (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
+                    "#CMOV_V8F64 PSEUDO!",
+                    [(set VR512:$dst,
+                      (v8f64 (X86cmov VR512:$t, VR512:$f, imm:$cond,
+                                          EFLAGS)))]>;
+  def CMOV_V16F32 : I<0, Pseudo,
+                    (outs VR512:$dst), (ins VR512:$t, VR512:$f, i8imm:$cond),
+                    "#CMOV_V16F32 PSEUDO!",
+                    [(set VR512:$dst,
+                      (v16f32 (X86cmov VR512:$t, VR512:$f, imm:$cond,
+                                          EFLAGS)))]>;

 }
 
 
 }
 
 

diff --git a/test/CodeGen/X86/avx512-select.ll b/test/CodeGen/X86/avx512-select.ll
new file mode 100644 (file)
index 0000000..d2d6681
--- /dev/null
+++ b/test/CodeGen/X86/avx512-select.ll
@@ -0,0 +1,22 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl  | FileCheck %s
+
+; CHECK-LABEL: select00
+; CHECK: vmovaps
+; CHECK-NEXT: LBB
+define <16 x i32> @select00(i32 %a, <16 x i32> %b) nounwind {
+  %cmpres = icmp eq i32 %a, 255
+  %selres = select i1 %cmpres, <16 x i32> zeroinitializer, <16 x i32> %b
+  %res = xor <16 x i32> %b, %selres
+  ret <16 x i32> %res
+}
+
+; CHECK-LABEL: select01
+; CHECK: vmovaps
+; CHECK-NEXT: LBB
+define <8 x i64> @select01(i32 %a, <8 x i64> %b) nounwind {
+  %cmpres = icmp eq i32 %a, 255
+  %selres = select i1 %cmpres, <8 x i64> zeroinitializer, <8 x i64> %b
+  %res = xor <8 x i64> %b, %selres
+  ret <8 x i64> %res
+}
+