[llvm] [SLSR] Adding a cost model considering register pressure (PR #213808)
Igor Wodiany via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 13 03:13:30 PDT 2026
================
@@ -1362,11 +1376,124 @@ bool StraightLineStrengthReduceLegacyPass::runOnFunction(Function &F) {
return StraightLineStrengthReduce(DL, DT, SE, TTI).runOnFunction(F);
}
+// Go through all operands of instruction, and check if any operand is used in
+// another block different from the instruction's block and the other use is
+// not rewritable. return true if such operand is found, otherwise return false.
+bool StraightLineStrengthReduce::
+ hasOperandsUsedInNonRewritableUsersInAnotherBlock(
+ llvm::Instruction *Inst) const {
+ llvm::BasicBlock *InstBB = Inst->getParent();
+
+ for (Value *OpVal : Inst->operand_values()) {
+ auto *OpInst = dyn_cast<Instruction>(OpVal);
+ if (!OpInst)
+ continue;
+
+ for (const User *U : OpInst->users())
+ if (auto *UI = dyn_cast<Instruction>(U)) {
+ if (UI->isDebugOrPseudoInst())
+ continue;
+ if (UI->getParent() != InstBB && !hasRewritableCandidates(UI)) {
+ LLVM_DEBUG(dbgs()
+ << "Inst's operand is used in another block "
+ << "("
+ << (InstBB->hasName() ? InstBB->getName() : "unnamed")
+ << " -> "
+ << (UI->getParent() && UI->getParent()->hasName()
+ ? UI->getParent()->getName()
+ : "unnamed")
+ << ") " << *UI << "\n");
+ return true;
+ }
+ }
+ }
+ return false;
+}
+
+bool StraightLineStrengthReduce::hasRewritableCandidates(
+ const Instruction *Inst) const {
+ if (!RewriteCandidates.count(Inst))
+ return false;
+
+ for (const Candidate *C : RewriteCandidates.at(Inst))
+ if (C->Basis)
+ return true;
+
+ return false;
+}
+
+// Assign a monotonically increasing index to each (non-debug/pseudo)
+// instruction in BB so in-block distances can be queried in O(1) once built.
+static DenseMap<const Instruction *, int>
----------------
IgWod wrote:
Should the function take a map as an argument to potentially avoid a copy on return? Not sure what's a LLVM guidance on that.
https://github.com/llvm/llvm-project/pull/213808
More information about the llvm-commits
mailing list