[PATCH] D73879: AMDGPU/GlobalISel: Avoid handling registers twice in waterfall loops

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: nhaehnle, kerbowa.
Herald added subscribers: Petar.Avramovic, hiraditya, t-tye, tpr, dstuttard, rovka, yaxunl, wdng, jvesely, kzhuravl.
Herald added a project: LLVM.

When multiple instructions are moved into a waterfall loop, it's
possible some of them re-use the same operands. Avoid creating
multiple sequences of readfirstlanes for them. None of the current
uses will hit this, but will be used in a future patch.


https://reviews.llvm.org/D73879

Files:
  llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp


Index: llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -747,6 +747,10 @@
   SmallVector<Register, 4> InitResultRegs;
   SmallVector<Register, 4> PhiRegs;
 
+  // Track use registers which have already been expanded with a readfirstlane
+  // sequence. This may have multiple uses if moving a sequence.
+  DenseMap<Register, Register> WaterfalledRegMap;
+
   MachineBasicBlock &MBB = B.getMBB();
   MachineFunction *MF = &B.getMF();
 
@@ -853,9 +857,18 @@
       if (!Op.isReg() || Op.isDef())
         continue;
 
-      if (!SGPROperandRegs.count(Op.getReg()))
+      Register OldReg = Op.getReg();
+      if (!SGPROperandRegs.count(OldReg))
         continue;
 
+      // See if we already processed this register in another instruction in the
+      // sequence.
+      auto OldVal = WaterfalledRegMap.find(OldReg);
+      if (OldVal != WaterfalledRegMap.end()) {
+        Op.setReg(OldVal->second);
+        continue;
+      }
+
       LLT OpTy = MRI.getType(Op.getReg());
       unsigned OpSize = OpTy.getSizeInBits();
 
@@ -1001,6 +1014,9 @@
 
         MRI.setRegBank(Op.getReg(), AMDGPU::SGPRRegBank);
       }
+
+      // Make sure we don't re-process this register again.
+      WaterfalledRegMap.insert(std::make_pair(OldReg, Op.getReg()));
     }
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D73879.242033.patch
Type: text/x-patch
Size: 1446 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200203/833a1ea0/attachment.bin>