x86_64/uftrace segmentation fault

[kraj]

Bugzilla Link	43626
Version	9.0
OS	Linux
Attachments	test case and script to run it
CC	@RKSimon,@zygoloid

Extended Description

attached sample makes clang crash

[Endilll]

Still crashing as of post-17 trunk: https://godbolt.org/z/7beb6xYrh
I wonder if this is yet another instance of disabling SSE2 when its registers are required by x86_64 calling convention.
Reduced by C-Reduce:

void print_diff_percent() {
  __pr_out("%s%+7.2f%s%%", "", 999.99, "");
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

[Endilll]

CC @phoebewang

[phoebewang]

I wonder if this is yet another instance of disabling SSE2 when its registers are required by x86_64 calling convention.

I think so. GCC will still generate SSE2 instruction even though SSE2 is disabled. I think we just need a better diagnosis instead of crash directly.

[Endilll]

CC @AaronBallman

[AaronBallman]

Oh this is neat. https://godbolt.org/z/hTK96nWaT

Not asserting: -x c -Xclang -target-feature -Xclang -sse2
Asserting: -x c -Xclang -target-feature -Xclang -sse2 -Wno-implicit-function-declaration

So I think the issue is that the implicit function declaration is a K&R C function, so when we allow the definition to be implicitly declared, we trigger the backend assertion.

I wonder if the best approach here is to make it an error to attempt to disable sse2 (or any target features) that are required for ABI purposes for any given target? CC @MaskRay for opinions

[MaskRay]

This is likely a duplicate of #29774. I have a reply at #29774 (comment)
I do wish that we can reject -mno-sse -mno-sse2 for x86-64, but the options are misused (and even by Linux kernel) and the projects may complain.

[AaronBallman]

This is likely a duplicate of #29774. I have a reply at #29774 (comment) I do wish that we can reject -mno-sse -mno-sse2 for x86-64, but the options are misused (and even by Linux kernel) and the projects may complain.

I think people misusing the options is exactly why we should start to diagnose it (we've already gotten multiple reports from users about crashing, so this is something people are hitting repeatedly). Is there a way we can do a gentle introduction (like warn "this combination of options will be a hard error in Clang N") so that we can start moving the needle in the correct direction? Could @nickdesaulniers or @nathanchance help us with the Linux kernel side of things?