[AMDGPU] Improve amdgcn_kill verification

[ro-i]

No description provided.

[llvmorg-github-actions]

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-backend-amdgpu

Author: Robert Imschweiler (ro-i)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/208220.diff

1 Files Affected:

• (modified) llvm/lib/IR/VerifierAMDGPU.cpp (+6-1)

diff --git a/llvm/lib/IR/VerifierAMDGPU.cpp b/llvm/lib/IR/VerifierAMDGPU.cpp
index 04cb214ef2520..ab63307570b28 100644
--- a/llvm/lib/IR/VerifierAMDGPU.cpp
+++ b/llvm/lib/IR/VerifierAMDGPU.cpp
@@ -145,7 +145,12 @@ void llvm::verifyAMDGPUIntrinsicCall(VerifierSupport &VS, Intrinsic::ID ID,
       Check(CBI->getNumIndirectDests() == 1,
             "callbr amdgcn_kill only supports one indirect dest");
       bool Unreachable = isa<UnreachableInst>(CBI->getIndirectDest(0)->begin());
-      CallInst *CI = dyn_cast<CallInst>(CBI->getIndirectDest(0)->begin());
+      // We assume that amdgcn_unreachable is only introduced by
+      // AMDGPUUnifyDivergentExitNodes, which replaces the block's original
+      // unreachable terminator by a call to amdgcn_unreachable + a return.
+      const Instruction *Term = CBI->getIndirectDest(0)->getTerminator();
+      const CallInst *CI =
+          Term ? dyn_cast_if_present<CallInst>(Term->getPrevNode()) : nullptr;
       Check(Unreachable ||
                 (CI && CI->getIntrinsicID() == Intrinsic::amdgcn_unreachable),
             "callbr amdgcn_kill indirect dest needs to be unreachable");

[shiltian]

Do we need any test?

[ro-i]

verifier has only negative tests, iirc? Could probably make a test in CodeGen/

[shiltian]

verifier has only negative tests, iirc?

Yes. I thought this PR is to capture something that might be missing before?

[ro-i]

yeah, makes sense. This also led me to relax the unreachable verification, as well, although that wasn't my initial intention (but ig it makes more sense)