[PowerPC] fatal error: error in backend: failed to perform tail call elimination on a call site marked musttail

[pkubaj]

FreeBSD 13.2-RELEASE
LLVM from devel/llvm-devel, version 17.0.d20230522,
PowerPC, PowerPC64, PowerPC64LE

Reproducer:

void name(int *params, int *program, int r, int g, int b, int a) {
  auto fn = (void (*)(int *, int *, int, int, int, int))1;
  [[clang::musttail]] return fn(params, program, r, g, b, a);
}

Compile with:

clang++-devel -cc1 -triple powerpc64le-portbld-freebsd13.2 -emit-obj SkOpts-735c83.cpp

Output:

fatal error: error in backend: failed to perform tail call elimination on a call site marked musttail

This causes Firefox 114 to fail to build.

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

[orcguru]

The error is due to following check:

llvm-project/llvm/lib/Target/PowerPC/PPCISelLowering.cpp

Line 4902 in 375d798

if (!Subtarget.isUsingPCRelativeCalls() &&

According to the comments above the check logic, ppc backend cannot do tail call elimination for this case. PPC backend has a call PPCTargetLowering::supportsTailCallFor() to check if tail call elimination is eligible.

Now looking at the reproducer, the "[[clang::musttail]]" attribute is verified by Sema::checkMustTailAttr(). At that time I don't know how to reach the PPC backend call to determine if that attribute is eligible, but the end result is the same: the source code won't compile on PPC backend, either due to a check from sema (if somebody add the backend check logic for PPC, but looks like not a good idea to me) or a check from PPC backend.

[pkubaj]

When setting amd64-unknown-freebsd13.2 triple, the code is built properly, so it's likely about a PPC backend.

[Long5hot]

There is one similiar case which i found..

int square(int num) {
    return num * num;
}

int checkNum(int num) {
  [[clang::musttail]]
  return square(num);
}

But it fails only on ppc32 not on ppc64..

Compile with:

clang++ ppc32crash.cpp --target=ppc32 -S

[daltenty]

Also fails on AIX 32-bit and 64-bit targets.

[nemanjai]

This seems to be a poorly thought-out attribute. We ultimately can't tail-call across DSO boundaries. This is a hard limitation on PPC (both Linux and AIX) due to the TOC pointer. For the front end, we could implement a Sema check to error out when a call site is marked with this attribute and the callee does not qualify for the dso_local runtime specifier.

[nemanjai]

The only PPC configuration that can always tail-call is Little Endian Linux with PC-Relative addressing turned on (default with -mcpu=power10).

[chenzheng1030]

@pkubaj PPC is not able to tail call every function marked with musttail attribute. I think Nemanjai gave the reason in above comments. So what do you think resolving this issue by not setting this attribute on PowerPC target? Thanks.

[pkubaj]

@chenzheng1030 disabling musttail on powerpc* is the current workaround, but this is not a proper solution. The compiler should not crash on musttail. If it can't tail call some code, it should print a warning, not crash. Crashing is not a proper behaviour.

[chenzheng1030]

@chenzheng1030 disabling musttail on powerpc* is the current workaround, but this is not a proper solution. The compiler should not crash on musttail. If it can't tail call some code, it should print a warning, not crash. Crashing is not a proper behaviour.

OK, let me try adding diagnosis for this attribute and ignore the attribute if it is not able to tailcall.