[AMDGPU] Sprintf test fails after amdgpu-attributor pass

[jhuber6]

The IR after the amdgpu-attributor pass causes the return value of the sprintf call to fail in this unit test https://godbolt.org/z/h8x47PjKv.

llvm-project/libc/test/integration/src/stdio/sprintf_size_test.cpp

Line 4 in c1cec8c

// See https://llvm.org/LICENSE.txt for license information.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Joseph Huber (jhuber6)

The IR after the `amdgpu-attributor` pass causes the return value of the `sprintf` call to fail in this unit test https://godbolt.org/z/h8x47PjKv. https://github.com/llvm/llvm-project/blob/c1cec8c0dc5b0296f0bc86745b867ff72c0a21e3/libc/test/integration/src/stdio/sprintf_size_test.cpp#L4

[arsenm]

Attach IR?

[jhuber6]

Attach IR?

It's in the godbolt link, it's like 10,000 lines because I haven't reduced it at all.

[arsenm]

This is broken:

    >   %296 = icmp eq ptr addrspace(5) %5, null
    <   %296 = icmp eq ptr addrspace(5) %5, addrspacecast (ptr null to ptr addrspace(5))

I'm guessing this is the AddressSpaceAA mishandling null

[shiltian]

I'm not sure if AAAddressSpace can cause this issue, because it only runs on pointer arguments of load or store like instructions, and this doesn't look like one.

FWIW, I think this issue occurred after not running AMDGPUAttributor twice in LTO.

[jhuber6]

I'm not sure if AAAddressSpace can cause this issue, because it only runs on pointer arguments of load or store like instructions, and this doesn't look like one.

FWIW, I think this issue occurred after not running AMDGPUAttributor twice in LTO.

The test case passes before -passes=amdgpu-attributor so it needs to be something in the pass.

[shiltian]

It’s not actually related to AAAddressSpace. Enabling it just exposes an issue in the attributor framework. I’ll fix it soon.