[llvm] [AMDGPU] Add IR LiveReg type-based optimization (PR #66838)

[Matt Arsenault via llvm-commits]

================
@@ -102,14 +184,261 @@ bool AMDGPULateCodeGenPrepare::runOnFunction(Function &F) {
   AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
   UA = &getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
 
+  // "Optimize" the virtual regs that cross basic block boundaries. When
+  // building the SelectionDAG, vectors of illegal types that cross basic blocks
+  // will be scalarized and widened, with each scalar living in its
+  // own physical register. To work around this, this optimization converts the
+  // vectors to equivalent vectors of legal type (which are converted back
+  // before uses in subsequent blocks), to pack the bits into fewer physical
+  // registers (used in CopyToReg/CopyFromReg pairs).
+  LiveRegOptimizer LRO(Mod);
+
   bool Changed = false;
   for (auto &BB : F)
-    for (Instruction &I : llvm::make_early_inc_range(BB))
+    for (Instruction &I : make_early_inc_range(BB)) {
       Changed |= visit(I);
+      if (!LRO.shouldReplaceUses(I))
+        continue;
+      Changed |= LRO.replaceUses(I);
+    }
 
+  Changed |= LRO.replacePHIs();
   return Changed;
 }
 
+bool LiveRegOptimizer::replaceUses(Instruction &I) {
+  bool MadeChange = false;
+
+  struct ConvertUseInfo {
+    Instruction *Converted;
+    SmallVector<Instruction *, 4> Users;
+  };
+  DenseMap<BasicBlock *, ConvertUseInfo> InsertedConversionMap;
+
+  ConversionCandidateInfo FromCCI(&I, I.getParent(),
+                                  std::next(I.getIterator()));
+  FromCCI.setNewType(getCompatibleType(FromCCI.getLiveRegDef()));
+  for (auto IUser = I.user_begin(); IUser != I.user_end(); IUser++) {
----------------
arsenm wrote:

This can be a range loop 

https://github.com/llvm/llvm-project/pull/66838