AMDGPU: Move v_readlane lane select from VGPR to SGPR

Summary:
Fix a compiler bug when the lane select happens to end up in a VGPR.

Clarify the semantic of the corresponding intrinsic to be that of
the corresponding GLSL: the lane select must be uniform across a
wave front, otherwise results are undefined.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D32343

llvm-svn: 301197

1 files changed, 13 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index d51110bcbd6..c5af8a1ad92 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2640,6 +2640,19 @@ void SIInstrInfo::legalizeOperandsVOP2(MachineRegisterInfo &MRI,
   if (isLegalRegOperand(MRI, InstrDesc.OpInfo[Src1Idx], Src1))
     return;
 
+  // Special case: V_READLANE_B32 accepts only immediate or SGPR operands for
+  // lane select. Fix up using V_READFIRSTLANE, since we assume that the lane
+  // select is uniform.
+  if (Opc == AMDGPU::V_READLANE_B32 && Src1.isReg() &&
+      RI.isVGPR(MRI, Src1.getReg())) {
+    unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+    const DebugLoc &DL = MI.getDebugLoc();
+    BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
+        .add(Src1);
+    Src1.ChangeToRegister(Reg, false);
+    return;
+  }
+
   // We do not use commuteInstruction here because it is too aggressive and will
   // commute if it is possible. We only want to commute here if it improves
   // legality. This can be called a fairly large number of times so don't waste