[llvm] [AMDGPU] Eliminate likely-spurious execz checks (PR #117567)

[Fabian Ritter via llvm-commits]

================
@@ -0,0 +1,201 @@
+//===- AMDGPUAnnotateVaryingBranchWeights.cpp -----------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// Estimate if conditional branches for which SIAnnotateControlFlow introduced
+// amdgcn_if or amdgcn_else intrinsics are likely to have different outcomes for
+// the threads of each wavefront. If that is the case, BranchWeight metadata is
+// added to signal that "then" and "else" blocks are both likely to be executed.
+// This may introduce branch weights that would be self-contradictory in a
+// non-SIMT setting.
+//
+// A consequence of this is that SIPreEmitPeephole is more likely to eliminate
+// s_cbranch_execz instructions that were introduced to skip these blocks when
+// no thread in the wavefront is active for them.
+//
+// Should only run after SIAnnotateControlFlow.
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUTargetMachine.h"
+#include "GCNSubtarget.h"
+#include "llvm/Analysis/LazyValueInfo.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/Analysis.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/IR/ProfDataUtils.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Target/TargetMachine.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "amdgpu-annotate-varying-branch-weights"
+
+namespace {
+
+class AMDGPUAnnotateVaryingBranchWeightsImpl {
+public:
+  AMDGPUAnnotateVaryingBranchWeightsImpl() = delete;
+  AMDGPUAnnotateVaryingBranchWeightsImpl(const GCNSubtarget &ST,
+                                         const TargetTransformInfo &TTI)
+      : ST(ST), TTI(TTI) {
+    // Determine weights that signal that a branch is very likely to be
+    // predicted correctly, i.e., whose ratio exceeds
+    // TTI.getPredictableBranchThreshold().
+    auto BranchProbThreshold = TTI.getPredictableBranchThreshold();
+    LikelyWeight = BranchProbThreshold.getNumerator();
+    UnlikelyWeight = BranchProbThreshold.getDenominator() - LikelyWeight;
+    if (UnlikelyWeight > 0)
+      --UnlikelyWeight;
+  }
+
+  bool run(Function &F);
+
+private:
+  const GCNSubtarget &ST;
+  const TargetTransformInfo &TTI;
+  uint32_t LikelyWeight;
+  uint32_t UnlikelyWeight;
+  ValueMap<const Value *, bool> LikelyVaryingCache;
+
+  /// Heuristically check if it is likely that a wavefront has dynamically
+  /// varying values for V.
+  bool isLikelyVarying(const Value *V);
+
+  /// Set branch weights that signal that the "true" successor of Term is the
+  /// likely destination, if no prior weights are present.
+  /// Return true if weights were set.
+  bool setTrueSuccessorLikely(BranchInst *Term);
+};
+
+class AMDGPUAnnotateVaryingBranchWeightsLegacy : public FunctionPass {
+public:
+  static char ID;
+  AMDGPUAnnotateVaryingBranchWeightsLegacy() : FunctionPass(ID) {
+    initializeAMDGPUAnnotateVaryingBranchWeightsLegacyPass(
+        *PassRegistry::getPassRegistry());
+  }
+
+  StringRef getPassName() const override {
+    return "AMDGPU Annotate Varying Branch Weights";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<TargetPassConfig>();
+    AU.setPreservesCFG();
+  }
+
+  bool runOnFunction(Function &F) override {
+    TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
+    const TargetMachine &TM = TPC.getTM<TargetMachine>();
+    const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
+    const TargetTransformInfo &TTI = TM.getTargetTransformInfo(F);
+    return AMDGPUAnnotateVaryingBranchWeightsImpl(ST, TTI).run(F);
+  }
+};
+
+} // end anonymous namespace
+
+char AMDGPUAnnotateVaryingBranchWeightsLegacy::ID = 0;
+
+INITIALIZE_PASS_BEGIN(AMDGPUAnnotateVaryingBranchWeightsLegacy, DEBUG_TYPE,
+                      "Annotate Varying Branch Weights", false, false)
+INITIALIZE_PASS_END(AMDGPUAnnotateVaryingBranchWeightsLegacy, DEBUG_TYPE,
+                    "Annotate Varying Branch Weights", false, false)
+
+FunctionPass *llvm::createAMDGPUAnnotateVaryingBranchWeightsLegacyPass() {
+  return new AMDGPUAnnotateVaryingBranchWeightsLegacy();
+}
+
+PreservedAnalyses
+AMDGPUAnnotateVaryingBranchWeightsPass::run(Function &F,
+                                            FunctionAnalysisManager &AM) {
+  const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
+  const TargetTransformInfo &TTI = TM.getTargetTransformInfo(F);
+  bool Changed = AMDGPUAnnotateVaryingBranchWeightsImpl(ST, TTI).run(F);
+
+  if (!Changed)
+    return PreservedAnalyses::all();
+
+  PreservedAnalyses PA;
+  PA.preserveSet<CFGAnalyses>();
+  return PA;
+}
+
+bool AMDGPUAnnotateVaryingBranchWeightsImpl::isLikelyVarying(const Value *V) {
+  // Check if V is a source of divergence or if it transitively uses one.
+  if (TTI.isSourceOfDivergence(V))
+    return true;
+
+  auto *U = dyn_cast<User>(V);
+  if (!U)
+    return false;
+
+  // Have we already checked V?
+  auto CacheEntry = LikelyVaryingCache.find(V);
+  if (CacheEntry != LikelyVaryingCache.end())
+    return CacheEntry->second;
+
+  // Does it use a likely varying Value?
+  bool Result = false;
+  for (const auto &Use : U->operands()) {
+    Result |= isLikelyVarying(Use);
----------------
ritter-x2a wrote:

With the updated heuristic in 8cf0cbf8fb44c9630cb48f51f3b2e22b68598272, recursion does not happen for phis any more, that should rule out endless loops.

https://github.com/llvm/llvm-project/pull/117567