================
@@ -2018,9 +2018,42 @@ bool GVNPass::performLoopLoadPRE(LoadInst *Load,
if (LoadPtr->canBeFreed())
return false;
- // TODO: Support critical edge splitting if blocker has more than 1
successor.
+ // The reload is inserted at the end of the chosen in-loop block so its value
+ // reaches the header along the path back to the latch. If that block has
+ // multiple successors, inserting there would also run the load on paths that
+ // leave the loop. Split the critical edge to the unique in-loop successor so
+ // the reload only runs on the path that feeds the header.
+ BasicBlock *InsertBlock = LoopBlock;
+ if (LoopBlock->getTerminator()->getNumSuccessors() != 1) {
+ if (isa<IndirectBrInst>(LoopBlock->getTerminator()))
+ return false;
+
+ BasicBlock *InLoopSucc = nullptr;
+ for (BasicBlock *Succ : successors(LoopBlock)) {
+ if (!L->contains(Succ))
+ continue;
+ // Bail if there is more than one in-loop successor: it is unclear which
+ // edge feeds the header.
+ if (InLoopSucc)
+ return false;
----------------
madhur13490 wrote:
If the blocker has more than one in-loop successor, it's ambiguous which edge
carries the reloaded value back to the header; splitting just one would leave
the load unavailable along the other in-loop path. Rather than pick arbitrarily
we bail. It wasn't tested before — I've added
`test_load_on_cold_path_multi_inloop_succ` which exercises exactly this path
and confirms the load is not PRE'd.
https://github.com/llvm/llvm-project/pull/206694
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
