R600/SI: Use s_movk_i32

llvm-svn: 221922

3 files changed, 17 insertions, 2 deletions

diff --git a/llvm/lib/Target/R600/SIInstrInfo.td b/llvm/lib/Target/R600/SIInstrInfo.td
index 6e098cd1dc1..b84a2b12149 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.td
+++ b/llvm/lib/Target/R600/SIInstrInfo.td
@@ -368,12 +368,12 @@ class SOPC_64<bits<7> op, string opName, PatLeaf cond = COND_NULL>
   : SOPC_Helper<op, SSrc_64, i64, opName, cond>;
 
 class SOPK_32 <bits<5> op, string opName, list<dag> pattern> : SOPK <
-  op, (outs SReg_32:$dst), (ins i16imm:$src0),
+  op, (outs SReg_32:$dst), (ins u16imm:$src0),
   opName#" $dst, $src0", pattern
 >;
 
 class SOPK_64 <bits<5> op, string opName, list<dag> pattern> : SOPK <
-  op, (outs SReg_64:$dst), (ins i16imm:$src0),
+  op, (outs SReg_64:$dst), (ins u16imm:$src0),
   opName#" $dst, $src0", pattern
 >;
 
diff --git a/llvm/lib/Target/R600/SILowerControlFlow.cpp b/llvm/lib/Target/R600/SILowerControlFlow.cpp
index 59270ee062e..9702565c462 100644
--- a/llvm/lib/Target/R600/SILowerControlFlow.cpp
+++ b/llvm/lib/Target/R600/SILowerControlFlow.cpp
@@ -586,6 +586,8 @@ bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
     MachineBasicBlock::iterator Start = MBB.getFirstNonPHI();
     const MCInstrDesc &SMovK = TII->get(AMDGPU::S_MOVK_I32);
 
+    assert(isInt<16>(StackOffset) && isInt<16>(StackSizeBytes));
+
     BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_LO)
       .addImm(StackOffset);
 
diff --git a/llvm/lib/Target/R600/SIShrinkInstructions.cpp b/llvm/lib/Target/R600/SIShrinkInstructions.cpp
index 0b9e7ca9666..45e83f54e7c 100644
--- a/llvm/lib/Target/R600/SIShrinkInstructions.cpp
+++ b/llvm/lib/Target/R600/SIShrinkInstructions.cpp
@@ -189,6 +189,19 @@ bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
       Next = std::next(I);
       MachineInstr &MI = *I;
 
+      // Try to use S_MOVK_I32, which will save 4 bytes for small immediates.
+      if (MI.getOpcode() == AMDGPU::S_MOV_B32) {
+        const MachineOperand &Src = MI.getOperand(1);
+
+        // TODO: Handle FPImm?
+        if (Src.isImm()) {
+          if (isInt<16>(Src.getImm()) && !TII->isInlineConstant(Src)) {
+            MI.setDesc(TII->get(AMDGPU::S_MOVK_I32));
+            continue;
+          }
+        }
+      }
+
       if (!TII->hasVALU32BitEncoding(MI.getOpcode()))
         continue;