From d19e8301745a25293edef20bfd4264b30277f24f Mon Sep 17 00:00:00 2001
From: Vincent Lejeune <vljn@ovi.com>
Date: Sat, 10 May 2014 19:18:25 +0000
Subject: [PATCH] R600/SI: Use pseudo instruction for fabs/clamp/fneg

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208478 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIISelLowering.cpp | 44 ++++++++++++++++++++++++++++++
 lib/Target/R600/SIInstructions.td  | 34 ++++++++++++++++++++---
 2 files changed, 74 insertions(+), 4 deletions(-)

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index cacff836774..1fae8cd592d 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -539,6 +539,50 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
       MIB.addOperand(MI->getOperand(i));
 
     MI->eraseFromParent();
+    break;
+  }
+  case AMDGPU::FABS_SI: {
+    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
+    const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
+    unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
+    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
+            Reg)
+            .addImm(0x7fffffff);
+    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_AND_B32_e32),
+            MI->getOperand(0).getReg())
+            .addReg(MI->getOperand(1).getReg())
+            .addReg(Reg);
+    MI->eraseFromParent();
+    break;
+  }
+  case AMDGPU::FNEG_SI: {
+    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
+    const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
+    unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
+    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_MOV_B32_e32),
+            Reg)
+            .addImm(0x80000000);
+    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_XOR_B32_e32),
+            MI->getOperand(0).getReg())
+            .addReg(MI->getOperand(1).getReg())
+            .addReg(Reg);
+    MI->eraseFromParent();
+    break;
+  }
+  case AMDGPU::FCLAMP_SI: {
+    const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
+    BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F32_e64),
+            MI->getOperand(0).getReg())
+            .addOperand(MI->getOperand(1))
+            .addImm(0) // SRC1
+            .addImm(0) // ABS
+            .addImm(1) // CLAMP
+            .addImm(0) // OMOD
+            .addImm(0); // NEG
+    MI->eraseFromParent();
   }
   }
   return BB;
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index e686ea71d6f..7d6624b0404 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1951,10 +1951,18 @@ def : BitConvert <v16f32, v16i32, VReg_512>;
 /********** Src & Dst modifiers **********/
 /********** =================== **********/
 
+def FCLAMP_SI : AMDGPUShaderInst <
+  (outs VReg_32:$dst),
+  (ins VSrc_32:$src0),
+  "FCLAMP_SI $dst, $src0",
+  []
+> {
+  let usesCustomInserter = 1;
+}
+
 def : Pat <
   (int_AMDIL_clamp f32:$src, (f32 FP_ZERO), (f32 FP_ONE)),
-  (V_ADD_F32_e64 $src, (i32 0 /* SRC1 */),
-   0 /* ABS */, 1 /* CLAMP */, 0 /* OMOD */, 0 /* NEG */)
+  (FCLAMP_SI f32:$src)
 >;
 
 /********** ================================ **********/
@@ -1973,14 +1981,32 @@ def : Pat <
   (V_OR_B32_e32 $src, (V_MOV_B32_e32 0x80000000)) /* Set sign bit */
 >;
 
+def FABS_SI : AMDGPUShaderInst <
+  (outs VReg_32:$dst),
+  (ins VSrc_32:$src0),
+  "FABS_SI $dst, $src0",
+  []
+> {
+  let usesCustomInserter = 1;
+}
+
 def : Pat <
   (fabs f32:$src),
-  (V_AND_B32_e32 $src, (V_MOV_B32_e32 0x7fffffff)) /* Clear sign bit */
+  (FABS_SI f32:$src)
 >;
 
+def FNEG_SI : AMDGPUShaderInst <
+  (outs VReg_32:$dst),
+  (ins VSrc_32:$src0),
+  "FNEG_SI $dst, $src0",
+  []
+> {
+  let usesCustomInserter = 1;
+}
+
 def : Pat <
   (fneg f32:$src),
-  (V_XOR_B32_e32 $src, (V_MOV_B32_e32 0x80000000)) /* Toggle sign bit */
+  (FNEG_SI f32:$src)
 >;
 
 /********** ================== **********/