R600/SI: replace SI_V_CNDLT with a pattern

It actually fixes quite a bunch of piglit tests.

This is a candidate for the mesa-stable branch.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@175756 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 2f304eb8e9833e99d7b9e6900e6db14b708de354..212e3f23a8496b7663abd082ebd8145e9c761d43 100644 (file)
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -81,9 +81,6 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
   case AMDGPU::SI_WQM:
     LowerSI_WQM(MI, *BB, I, MRI);
     break;
-  case AMDGPU::SI_V_CNDLT:
-    LowerSI_V_CNDLT(MI, *BB, I, MRI);
-    break;
   }
   return BB;
 }
@@ -127,25 +124,6 @@ void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
   MI->eraseFromParent();
 }
 
-void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
-    MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
-  unsigned VCC = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
-
-  BuildMI(BB, I, BB.findDebugLoc(I),
-          TII->get(AMDGPU::V_CMP_GT_F32_e32),
-          VCC)
-          .addImm(0)
-          .addOperand(MI->getOperand(1));
-
-  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32_e32))
-          .addOperand(MI->getOperand(0))
-          .addOperand(MI->getOperand(3))
-          .addOperand(MI->getOperand(2))
-          .addReg(VCC);
-
-  MI->eraseFromParent();
-}
-
 EVT SITargetLowering::getSetCCResultType(EVT VT) const {
   return MVT::i1;
 }

diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index a8429b7cbe3dd5ecc254c54f9a725d5ab5f9cb59..5d048f85007a937c5e6ad5e25804cb5b5f5671a4 100644 (file)
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -29,8 +29,6 @@ class SITargetLowering : public AMDGPUTargetLowering {
               MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
   void LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
               MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
-  void LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
-              MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
 
   SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 1e5f94ff983437f42844cc8e63a771fb4fd3f6ff..2a4128568412589b875eac3c60724cb154402ae6 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -990,13 +990,6 @@ def LOAD_CONST : AMDGPUShaderInst <
 
 let usesCustomInserter = 1 in {
 
-def SI_V_CNDLT : InstSI <
-  (outs VReg_32:$dst),
-  (ins VReg_32:$src0, VReg_32:$src1, VReg_32:$src2),
-  "SI_V_CNDLT $dst, $src0, $src1, $src2",
-  [(set VReg_32:$dst, (int_AMDGPU_cndlt VReg_32:$src0, VReg_32:$src1, VReg_32:$src2))]
->;
-
 def SI_INTERP : InstSI <
   (outs VReg_32:$dst),
   (ins VReg_32:$i, VReg_32:$j, i32imm:$attr_chan, i32imm:$attr, SReg_32:$params),
@@ -1086,6 +1079,11 @@ def SI_KILL : InstSI <
 
 } // end IsCodeGenOnly, isPseudo
 
+def : Pat<
+  (int_AMDGPU_cndlt VReg_32:$src0, VReg_32:$src1, VReg_32:$src2),
+  (V_CNDMASK_B32_e64 VReg_32:$src2, VReg_32:$src1, (V_CMP_GT_F32_e64 0, VReg_32:$src0))
+>;
+
 def : Pat <
   (int_AMDGPU_kilp),
   (SI_KILL (V_MOV_B32_e32 0xbf800000))