Inline assembly error: "requires more registers than available" with Clang 21.1.8 (`strsm_kernel_LT.c.o` build failure)

[nicolarossi]

Hello,

I am encountering a build failure when compiling OpenBLAS with Clang 21.1.8, targeting SKYLAKE and using AddressSanitizer. The specific error occurs when building the strsm_kernel_LT.c.o target.

Steps to reproduce:

cmake -G Ninja -DCMAKE_CXX_COMPILER=/opt/clang-21.1.8/bin/clang++ -DTARGET=SKYLAKE -DCCOMMON_OPT="-fsanitize=address" ..
ninja -- kernel/CMakeFiles/kernel.dir/CMakeFiles/strsm_kernel_LT.c.o

Error message:

OpenBLAS/build_clang21/kernel/CMakeFiles/strsm_kernel_LT.c'
In file included from /home/[nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/build_clang21/kernel/CMakeFiles/strsm_kernel_LT.c:9](http://nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/build_clang21/kernel/CMakeFiles/strsm_kernel_LT.c:9):
/home/[nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/kernel/x86_64/strsm_kernel_8x4_haswell_LT.c:221:32](http://nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/kernel/x86_64/strsm_kernel_8x4_haswell_LT.c:221:32): error: inline assembly requires more registers than available
  221 |   for(;n_count>11;n_count-=12) COMPUTE(12)
      |                                ^
/home/[nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/kernel/x86_64/strsm_kernel_8x4_haswell_LT.c:160:5](http://nicola.rossi@mavensecurities.com/fix_openmp_clang_ims/OpenBLAS/kernel/x86_64/strsm_kernel_8x4_haswell_LT.c:160:5): note: expanded from macro 'COMPUTE'
  160 |     "movq %0,%%r15; movq %1,%%r14; movq %7,%%r13; movq %6,%%r12; salq $2,%%r12; movq %10,%%r11;"

It appears that the inline assembly in strsm_kernel_8x4_haswell_LT.c is requesting more registers than are available when compiling with Clang 21 as C compiler and with AddressSanitizer enabled.

Additional information:

• Compiler: Clang 21.1.8
• AddressSanitizer enabled via -fsanitize=address
• Target: HASWELL

Is there a known workaround or fix for this issue? Any advice for making this kernel build successfully with ASan enabled would be much appreciated.

Thank you!

[martin-frbg]

Haven't tried with AddressSanitizer specifically, but a similar issue kept popping up with the SkylakeX SGEMM kernel. The workaround there was to add -mllvm -exhaustive-register-search to make the the compiler try harder. If you look at the issue tracker for LLVM, there are a bunch of unresolved tickets mentioning this error message...