[PATCH] D111500: [InstSimplify] Simplify intrinsic comparisons with domain knoweldge

Mon Oct 11 12:12:26 PDT 2021

nikic added inline comments.

================
Comment at: llvm/lib/Analysis/InstructionSimplify.cpp:609
+  // fold %cmp = icmp slt i32 %tid, %ntid to true.
+  if (Inst0->getIntrinsicID() == Intrinsic::nvvm_read_ptx_sreg_tid_x &&
+      Inst1->getIntrinsicID() == Intrinsic::nvvm_read_ptx_sreg_ntid_x)
----------------
tra wrote:
> What if LLVM has been compiled without NVPTX back-end? I'm not sure that NVVM intrinsics will be available then.
> 
> Perhaps we should re-visit enabling NVVMIntrRange.cpp pass, again. This should make it possible for LLVM to figure this optimization, and more.
I believe intrinsics are always included, even if the target is disabled. But I also don't think we have precedent for target intrinsic handling in InstSimplify, so adding @spatel and @lebedev.ri for that. Though I don't really see a problem with it.

We do provide InstCombine hooks (instCombineIntrinsic in TTI), but those work directly on the intrinsic. You could use that to replace NVVMIntrRange I believe. Though I don't think that would cover the particular use-case here, because range metadata is not sufficient to derive this result.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D111500/new/

https://reviews.llvm.org/D111500