[llvm] r202337 - R600/SI: Optimize SI_KILL for constant operands

[Michel Danzer]

Author: daenzer
Date: Wed Feb 26 19:47:09 2014
New Revision: 202337

URL: http://llvm.org/viewvc/llvm-project?rev=202337&view=rev
Log:
R600/SI: Optimize SI_KILL for constant operands

If the SI_KILL operand is constant, we can either clear the exec mask if
the operand is negative, or do nothing otherwise.

Reviewed-by: Tom Stellard <thomas.stellard at amd.com>

Modified:
    llvm/trunk/lib/Target/R600/SIInstructions.td
    llvm/trunk/lib/Target/R600/SILowerControlFlow.cpp
    llvm/trunk/test/CodeGen/R600/llvm.AMDGPU.kill.ll

Modified: llvm/trunk/lib/Target/R600/SIInstructions.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstructions.td?rev=202337&r1=202336&r2=202337&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstructions.td (original)
+++ llvm/trunk/lib/Target/R600/SIInstructions.td Wed Feb 26 19:47:09 2014
@@ -1333,7 +1333,7 @@ def SI_END_CF : InstSI <
 
 def SI_KILL : InstSI <
   (outs),
-  (ins VReg_32:$src),
+  (ins VSrc_32:$src),
   "SI_KILL $src",
   [(int_AMDGPU_kill f32:$src)]
 >;
@@ -1425,7 +1425,7 @@ def : Pat<
 
 def : Pat <
   (int_AMDGPU_kilp),
-  (SI_KILL (V_MOV_B32_e32 0xbf800000))
+  (SI_KILL 0xbf800000)
 >;
 
 /* int_SI_vs_load_input */

Modified: llvm/trunk/lib/Target/R600/SILowerControlFlow.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SILowerControlFlow.cpp?rev=202337&r1=202336&r2=202337&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SILowerControlFlow.cpp (original)
+++ llvm/trunk/lib/Target/R600/SILowerControlFlow.cpp Wed Feb 26 19:47:09 2014
@@ -55,6 +55,7 @@
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/Constants.h"
 
 using namespace llvm;
 
@@ -294,6 +295,7 @@ void SILowerControlFlowPass::Branch(Mach
 void SILowerControlFlowPass::Kill(MachineInstr &MI) {
   MachineBasicBlock &MBB = *MI.getParent();
   DebugLoc DL = MI.getDebugLoc();
+  const MachineOperand &Op = MI.getOperand(0);
 
   // Kill is only allowed in pixel / geometry shaders
   assert(MBB.getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType ==
@@ -301,10 +303,19 @@ void SILowerControlFlowPass::Kill(Machin
          MBB.getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType ==
          ShaderType::GEOMETRY);
 
-  // Clear this pixel from the exec mask if the operand is negative
-  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
-          .addImm(0)
-          .addOperand(MI.getOperand(0));
+  // Clear this thread from the exec mask if the operand is negative
+  if ((Op.isImm() || Op.isFPImm())) {
+    // Constant operand: Set exec mask to 0 or do nothing
+    if (Op.isImm() ? (Op.getImm() & 0x80000000) :
+        Op.getFPImm()->isNegative()) {
+      BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
+              .addImm(0);
+    }
+  } else {
+    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
+           .addImm(0)
+           .addOperand(Op);
+  }
 
   MI.eraseFromParent();
 }

Modified: llvm/trunk/test/CodeGen/R600/llvm.AMDGPU.kill.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/llvm.AMDGPU.kill.ll?rev=202337&r1=202336&r2=202337&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/llvm.AMDGPU.kill.ll (original)
+++ llvm/trunk/test/CodeGen/R600/llvm.AMDGPU.kill.ll Wed Feb 26 19:47:09 2014
@@ -1,13 +1,17 @@
 ; RUN: llc < %s -march=r600 -mcpu=verde -verify-machineinstrs | FileCheck --check-prefix=SI %s
 
-; SI-LABEL: @kill_gs
-; SI: V_CMPX_LE_F32
+; SI-LABEL: @kill_gs_const
+; SI-NOT: V_CMPX_LE_F32
+; SI: S_MOV_B64 exec, 0
 
-define void @kill_gs() #0 {
+define void @kill_gs_const() #0 {
 main_body:
   %0 = icmp ule i32 0, 3
   %1 = select i1 %0, float 1.000000e+00, float -1.000000e+00
   call void @llvm.AMDGPU.kill(float %1)
+  %2 = icmp ule i32 3, 0
+  %3 = select i1 %2, float 1.000000e+00, float -1.000000e+00
+  call void @llvm.AMDGPU.kill(float %3)
   ret void
 }