[Issue]: memory/build performance issues due to amdgpu-early-inline-all/amdgpu-function-calls in hipcc

[AngryLoki]

Problem Description

Hi,

Right now hipcc adds -mllvm -amdgpu-early-inline-all=true -mllvm -amdgpu-function-calls=false both in Perl and binary versions:
https://github.com/search?q=repo%3AROCm%2Fllvm-project%20amdgpu-early-inline-all&type=code

In LLVM 18 release I noticed a major memory/performance regression due to these flags and reported to upstream llvm-project.
Performance was partially improved, but overall developers there say that these flags should not be used:
llvm#86332 (comment)

In https://github.com/search?q=org%3AROCm+SWDEV-442724+OR+SWDEV-459556+OR+SWDEV-460260&type=commits I see that these arguments were removed from ROCm/HIP and ROCm/clr. Could you remove them from hipcc too?

Operating System

Gentoo

CPU

GPU

ROCm Version

ROCm 6.3.0

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[ppanchad-amd]

Hi @AngryLoki. Internal ticket has been created to investigate your issue. Thanks!

[schung-amd]

Hi @AngryLoki, is this happening in the wild? The hipcc scripts were changed to not add these flags by default and should only add them if --hipcc-no-func-supp is passed: feba9fa.

[AngryLoki]

Sorry, I missed the moment when funcSupp was enabled. The issue does not affect hipcc anymore. These flags are still added in various places (notably, composable_kernel), but there are no issues in hipcc. Thank you for correcting me