[AMDGPU] Remove calling conv check on entry function

[jofrn]

It is undefined behavior to call a function with a mismatched calling convention. Rather than crash on this behavior, it should compile.

This LLVM defect was identified via the AMD Fuzzing project.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: None (jofrn)

Changes

It is undefined behavior to call a function with a mismatched calling convention. Rather than crash on this behavior, it should compile.

---

Full diff: https://github.com/llvm/llvm-project/pull/162080.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp (-7)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp b/llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp
index 0ea9add891111..b03d50f2d451d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp
@@ -261,13 +261,6 @@ AMDGPUResourceUsageAnalysisImpl::analyzeResourceUsage(
 
         const Function *Callee = getCalleeFunction(*CalleeOp);
 
-        // Avoid crashing on undefined behavior with an illegal call to a
-        // kernel. If a callsite's calling convention doesn't match the
-        // function's, it's undefined behavior. If the callsite calling
-        // convention does match, that would have errored earlier.
-        if (Callee && AMDGPU::isEntryFunctionCC(Callee->getCallingConv()))
-          report_fatal_error("invalid call to entry function");
-
         auto isSameFunction = [](const MachineFunction &MF, const Function *F) {
           return F == &MF.getFunction();
         };

[arsenm]

Needs test

[CatherineMoore]

@CatherineMoore for awareness

[shiltian]

Is there anything less fatal that will not crash the backend? I think the comment wanted it to behave like that, but the description of the PR seems to say other way.

[jofrn]

Is there anything less fatal that will not crash the backend? I think the comment wanted it to behave like that, but the description of the PR seems to say other way.

The original comment is contradictory: it says to avoid crashing on undefined behavior but then crashes with report_fatal_error(). While crashing is still undefined behavior, it's less fatal to compile it (even if the result is incorrect) than to crash the backend. We have a lint check to catch the mismatched cc.