[AMDGPU] gfx1010 premlane instructions

Differential Revision: https://reviews.llvm.org/D63202

llvm-svn: 363185

1 files changed, 20 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 775b35d6521..dc35bf4e3eb 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -3790,6 +3790,26 @@ void SIInstrInfo::legalizeOperandsVOP3(MachineRegisterInfo &MRI,
     AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2)
   };
 
+  if (Opc == AMDGPU::V_PERMLANE16_B32 ||
+      Opc == AMDGPU::V_PERMLANEX16_B32) {
+    // src1 and src2 must be scalar
+    MachineOperand &Src1 = MI.getOperand(VOP3Idx[1]);
+    MachineOperand &Src2 = MI.getOperand(VOP3Idx[2]);
+    const DebugLoc &DL = MI.getDebugLoc();
+    if (Src1.isReg() && !RI.isSGPRClass(MRI.getRegClass(Src1.getReg()))) {
+      unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+      BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
+        .add(Src1);
+      Src1.ChangeToRegister(Reg, false);
+    }
+    if (Src2.isReg() && !RI.isSGPRClass(MRI.getRegClass(Src2.getReg()))) {
+      unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+      BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
+        .add(Src2);
+      Src2.ChangeToRegister(Reg, false);
+    }
+  }
+
   // Find the one SGPR operand we are allowed to use.
   int ConstantBusLimit = ST.getConstantBusLimit(Opc);
   int LiteralLimit = ST.hasVOP3Literal() ? 1 : 0;