[clang] [llvm] [VectorCombine] Trim low end of loads used in shufflevector rebroadcasts. (PR #149093)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 20 23:48:15 PST 2026
================
@@ -4789,8 +4794,13 @@ bool VectorCombine::shrinkLoadForShuffles(Instruction &I) {
return false;
// Create new load of smaller vector.
+ Value *NewPtr =
+ LowOffset > 0u
+ ? Builder.CreateInBoundsPtrAdd(PtrOp, Builder.getInt64(LowOffset))
----------------
arsenm wrote:
Should use constant of correct address space size. If this is a 32-bit pointer it adds an extra step
https://github.com/llvm/llvm-project/pull/149093
More information about the llvm-commits
mailing list