AMDGPU: Reduce number of copies emitted

Instead of always inserting a copy in case
the super register is itself a subregister,
only extract to the super reg class if this is
actually the case.

This shouldn't really change codegen, but
makes looking at the output of SIFixSGPRCopies
easier to read.

llvm-svn: 248467

1 files changed, 9 insertions, 5 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index c7218dedddc..e2b6d3c1403 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1556,17 +1556,21 @@ unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
                                          unsigned SubIdx,
                                          const TargetRegisterClass *SubRC)
                                          const {
-  assert(SuperReg.isReg());
-
-  unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
+  MachineBasicBlock *MBB = MI->getParent();
+  DebugLoc DL = MI->getDebugLoc();
   unsigned SubReg = MRI.createVirtualRegister(SubRC);
 
+  if (SuperReg.getSubReg() == AMDGPU::NoSubRegister) {
+    BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
+      .addReg(SuperReg.getReg(), 0, SubIdx);
+    return SubReg;
+  }
+
   // Just in case the super register is itself a sub-register, copy it to a new
   // value so we don't need to worry about merging its subreg index with the
   // SubIdx passed to this function. The register coalescer should be able to
   // eliminate this extra copy.
-  MachineBasicBlock *MBB = MI->getParent();
-  DebugLoc DL = MI->getDebugLoc();
+  unsigned NewSuperReg = MRI.createVirtualRegister(SuperRC);
 
   BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
     .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());