[llvm] r284891 - [AMDGPU] Perform uchar to float combine for ISD::SINT_TO_FP

[Konstantin Zhuravlyov via llvm-commits]

Author: kzhuravl
Date: Fri Oct 21 17:10:03 2016
New Revision: 284891

URL: http://llvm.org/viewvc/llvm-project?rev=284891&view=rev
Log:
[AMDGPU] Perform uchar to float combine for ISD::SINT_TO_FP

This will prevent following regression when enabling i16 support (D18049):
  test/CodeGen/AMDGPU/cvt_f32_ubyte.ll

Differential Revision: https://reviews.llvm.org/D25805

Modified:
    llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

Modified: llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp?rev=284891&r1=284890&r2=284891&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp Fri Oct 21 17:10:03 2016
@@ -233,6 +233,7 @@ SITargetLowering::SITargetLowering(const
   setTargetDAGCombine(ISD::AND);
   setTargetDAGCombine(ISD::OR);
   setTargetDAGCombine(ISD::XOR);
+  setTargetDAGCombine(ISD::SINT_TO_FP);
   setTargetDAGCombine(ISD::UINT_TO_FP);
   setTargetDAGCombine(ISD::FCANONICALIZE);
 
@@ -3520,19 +3521,27 @@ SDValue SITargetLowering::PerformDAGComb
   case AMDGPUISD::CVT_F32_UBYTE2:
   case AMDGPUISD::CVT_F32_UBYTE3: {
     unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
+
     SDValue Src = N->getOperand(0);
+    SDValue Srl = N->getOperand(0);
+    if (Srl.getOpcode() == ISD::ZERO_EXTEND)
+      Srl = Srl.getOperand(0);
 
     // TODO: Handle (or x, (srl y, 8)) pattern when known bits are zero.
-    if (Src.getOpcode() == ISD::SRL) {
+    if (Srl.getOpcode() == ISD::SRL) {
       // cvt_f32_ubyte0 (srl x, 16) -> cvt_f32_ubyte2 x
       // cvt_f32_ubyte1 (srl x, 16) -> cvt_f32_ubyte3 x
       // cvt_f32_ubyte0 (srl x, 8) -> cvt_f32_ubyte1 x
 
-      if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src.getOperand(1))) {
+      if (const ConstantSDNode *C =
+              dyn_cast<ConstantSDNode>(Srl.getOperand(1))) {
+        Srl = DAG.getZExtOrTrunc(Srl.getOperand(0), SDLoc(Srl.getOperand(0)),
+                                 EVT(MVT::i32));
+
         unsigned SrcOffset = C->getZExtValue() + 8 * Offset;
         if (SrcOffset < 32 && SrcOffset % 8 == 0) {
           return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0 + SrcOffset / 8, DL,
-                             MVT::f32, Src.getOperand(0));
+                             MVT::f32, Srl);
         }
       }
     }
@@ -3550,7 +3559,7 @@ SDValue SITargetLowering::PerformDAGComb
 
     break;
   }
-
+  case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP: {
     return performUCharToFloatCombine(N, DCI);
   }