Test failures on PPC64 with GCC 15 (Power9 baseline)

[schopin-pro]

Hi,

On the upcoming Ubuntu 25.10, when compiling and running highway using the Power9 baseline with GCC 15, we get the following failures:

The following tests FAILED:
	405 - HwyRandomTestGroup/HwyRandomTest.TestSeeding/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	406 - HwyRandomTestGroup/HwyRandomTest.TestMultiThreadSeeding/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	407 - HwyRandomTestGroup/HwyRandomTest.TestRandomUint64/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	408 - HwyRandomTestGroup/HwyRandomTest.TestNextNRandomUint64/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	409 - HwyRandomTestGroup/HwyRandomTest.TestNextFixedNRandomUint64/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	410 - HwyRandomTestGroup/HwyRandomTest.TestCachedXorshiro/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	411 - HwyRandomTestGroup/HwyRandomTest.TestUniformDist/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	412 - HwyRandomTestGroup/HwyRandomTest.TestNextNUniformDist/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	413 - HwyRandomTestGroup/HwyRandomTest.TestNextFixedNUniformDist/PPC9  # GetParam() = 281474976710656 (ILLEGAL)
	414 - HwyRandomTestGroup/HwyRandomTest.TestUniformCachedXorshiro/PPC9  # GetParam() = 281474976710656 (ILLEGAL)

Tests run fine if we either drop back to -mcpu=power8 or use GCC 14.

The issue appears both in our packaged version and in a fresh git clone.

[schopin-pro]

Ubuntu bug: https://bugs.launchpad.net/ubuntu/+source/highway/+bug/2121375

[jan-wassenberg]

Thanks for raising the issue. Do I understand correctly (from Heinrich's GCC bug report) that these errors are resolved by the Highway commit "pre-release workaround for PPC TestIntegerMaskedDiv: Unpredictable1"?
If so, would be good to raise a GCC bug for that as well - very likely a compiler bug (Our workaround prevents the compiler from evaluating the expression at compile time).

Is the issue here that we want to build a pre-1.3.0 version of Highway with GCC 15? We generally recommend using the latest release if that's feasible.

[schopin-pro]

I think there's some confusion here, most likely because we've been dealing with a series of build failures in a single bug on the Ubuntu bug.

Heinrich is dealing with a RISCV internal compiler error. It probably doesn't have anything to do with the failure at hand, which is on the IBM PowerPC architecture. Sadly, the commit you mention doesn't fix the issue, since it was already on the main branch by the time I tested on a fresh git clone.

My educated guess is that it's a bug in the GCC codegen, but I don't really have the time to dig into it, especially since I don't know much about either the GCC codebase, Highway, or PPC64LE. I did ping some folks internally that could lend a hand, though.

[jan-wassenberg]

Got it. Thanks for pinging. Given that it worked in gcc 14, a compiler bug is quite likely.

It would be an option to (temporarily) disable POWER9 codegen on GCC 15 if that's helpful.

Please also consider filing this as a GCC bug.

[Scottcjn]

I dug into this and found a concrete, reproducible failure mode — though I want to be upfront that it may be a distinct issue from the compile-time-evaluation miscompile discussed above, since I could only test under emulation (no real Power9 hardware).

Setup: cross-built random_test with powerpc64le-linux-gnu-g++ 15.2 at -mcpu=power9, ran under qemu-ppc64le -cpu power9.

Symptom: SIGILL (illegal instruction). Disassembling the faulting address, the illegal instruction is a pla (Power10 prefixed Load-Address) sitting in a linker-generated long-branch/PLT stub — not in Highway's or gcc's codegen. The stubs come from ld's --power10-stubs, which defaults to auto and emits Power10 stubs even though everything was compiled -mcpu=power9. pla is illegal on Power9, hence the trap.

Workaround that fixes it for me: linking with -Wl,--no-power10-stubs makes all 16 HwyRandomTest/PPC9 cases pass under qemu -cpu power9 (confirmed for both static and dynamic builds). -cpu power10 also passes (it implements pla), which is consistent.

Important caveats — please don't take this as confirmed root cause:

• This is a SIGILL from linker stubs, not a wrong-value miscompile. A minimal C reproducer shows gcc 15 at -mcpu=power9 does not emit pcrel/prefixed instructions in its own codegen (-mpcrel is disabled by default), so this isn't a gcc codegen regression.
• A gcc 14 build at -mcpu=power9 also emits some Power10 pla stubs here — they just didn't happen to land in the executed path in my build, so the gcc14-passes/gcc15-fails split might be incidental code layout under emulation rather than a fundamental compiler difference.
• Because I'm on qemu, I can't be sure this is the same failure you hit on real Power9 silicon vs. a separate linker-stub problem.

Suggested confirmation: if anyone can re-run on actual Power9 hardware with -DCMAKE_EXE_LINKER_FLAGS="-Wl,--no-power10-stubs" (or --power10-stubs=no), that would tell us whether this resolves the original ctest failures — i.e., whether they were illegal-instruction crashes (this issue) or genuine wrong values (the compile-time-eval theory). Happy to help narrow it further.