`.subsections_via_symbols` directive causes invalid symbol pruning in macOS x86-64 kernels

[topolarity]

// sasum.c
#include <stdio.h>
#include <cblas.h>
int main(void) {
  float x[5] = {3, -1, 2, -4, 1};
  printf("cblas_sasum = %g\n", cblas_sasum(5, x, 1));   // expect 11
  return 0;
}

$ clang -arch x86_64 -O2 -I$(BUILD_DIR) sasum.c -L$(BUILD_DIR) -lopenblas -Wl,-rpath,$(BUILD_DIR) -o sasum_dyn
$ ./sasum_dyn
cblas_sasum = 4.01598e-36

The correct output is:

$ ./sasum_dyn
cblas_sasum = 11

OpenBLAS built with make TARGET=NEHALEM NOFORTRAN=1 (on Apple Silicon, the full invocation is arch -x86_64 make -j6 ARCH=x86_64 BINARY=64 CC="clang -arch x86_64" HOSTCC="clang -arch x86_64" TARGET=NEHALEM NOFORTRAN=1 shared) with Apple clang 21.0.0

The problem is that even without -Wl,-dead-strip the linker will still try to deduplicate labels unless -Wl,-no_deduplicate is provided explicitly. This happens to .L998 and .L999 in the sasum_k kernel.

I am testing via Rosetta on AArch64 macOS hardware, but I suspect this reproduces on proper x86-64 macOS targets.

[topolarity]

Looks like this problem is specific to Apple's new linker from 2023. Building with -Wl,-ld_classic also works around the issue.

[martin-frbg]

Not sure what to make of this - your PRs look way too invasive for what may well be a bug in a specific linker version on an obsolescent target that (if I understand you correctly) gets triggered only by very specific non-default build options.

[topolarity]

your PRs look way too invasive

Agreed. The PRs are much bigger and more invasive than I would like...

I hope I can clarify the motivation. The goal for #6088 / #6087 is to allow the linker to remove unused code - #6090 is only needed to enable code pruning on macOS.

a bug in a specific linker version on an obsolescent target

x86_64-darwin happens to trigger this issue w/o dead-strip, but the problem is broader than that.

The issue here happens as soon as you enable .subsections_via_symbols with a new enough clang, on any macOS target AFAIK and so dead-strip breaks OpenBLAS on macOS (incl. arm64) without #6090

The first two PRs apply to all platforms (--Wl,--gc-sections on Linux / MinGW + -Wl,-dead-strip on macOS) and they are trying to allow the linker to strip code under DYNAMIC_ARCH. Without them, gotoblas references all kernels even if you e.g. only rely on dgemm.