libomp :: misc_bugs/many-microtask-args.c FAILs on SPARC

[rorth]

Once PR #138523 gets in, one openmp failure remains on SPARC (both Solaris/sparcv9 and Linux/sparc64):

  libomp :: misc_bugs/many-microtask-args.c

The log shows

# executed command: /var/llvm/local-sparc64-debug-openmp/projects/openmp/runtime/test/misc_bugs/Output/many-microtask-args.c.tmp
# .---command stderr------------
# | Too many args to microtask: 17!

with the last message repeated several times. The message is from z_Linux_util.cpp (__kmp_invoke_microtask), so far I have no idea where that additional 17th arg is from. I found it's ultimately passed to __kmp_fork_call which is generated in clang/lib/CodeGen/CGOpenMPRuntime.cpp (CGOpenMPRuntime::emitParallelCall).

@jprotze @mjklemm @TerryLWilmarth Any suggestions on what may be going on?

[jprotze]

The test has:

  #pragma omp parallel for firstprivate(i1, i2, i3, i4, i5, i6, i7, i8, i9, i10, i11, i12, i13, i14, i15, i16) reduction(+:r)
  for (i = 0; i < i16; i++) {
    r += i + i1 + i2 + i3 + i4 + i5 + i6 + i7 + i8 + i9 + i10 + i11 + i12 + i13 + i14 + i15 + i16;
  }

According to OpenMP semantics, each thread has a private copy of all variables in the firstprivate clause and also in the reduction clause. The firstprivate variables are initialized with the value from outside the parallel, so they are passed in by value. The reduction variable is basically an output variable, so r is passed into each thread by reference. So, in total the parallel region has semantically 17 arguments.

The compiler generates a function like, which must be called by __kmp_invoke_microtask:

void main.omp_outlined_debug__(int *gtid, int *tid, int i1, int i2, ... , int i16, int* r){
  int lb, ub, step;
  __kmpc_get_bounds(0, i16, &lb, &ub, &step);
  for (int i = lb; i < ub; i+=step) {
    *r += i + i1 + i2 + i3 + i4 + i5 + i6 + i7 + i8 + i9 + i10 + i11 + i12 + i13 + i14 + i15 + i16;
  }
}

For several architectures, __kmp_invoke_microtask is defined in z_Linux_asm.S to expand the argc+argv[] into a list of function parameters according to the calling convention of the architecture. The code in z_Linux_util.cpp is just the fallback code for architectures that don't define the function as asm. The code in z_Linux_util.cpp is limited to support 15 parameters.

I'm not sure and did not check, but I would hope that the codegen for these architectures limits the parameters for the outlined function to 15 parameters and probably packs the remaining parameters as an array into the last parameter?

So, you can either add asm code that implements the __kmp_invoke_microtask for sparc as sketched in the comments in z_Linux_asm.S or check the codegen, whether it can generate code for the fallback solution.

[rorth]

Thanks for the thorough explanation. While I won't get to this soon, I ran a quick experiment, disabling the assembler versions of __kmp_invoke_microtask on x86, rebuilding libomp.so and rerunning the openmp tests: now I get the same error on i386-pc-solaris2.11, so this isn't SPARC-specific but related to the C implementation of __kmp_invoke_microtask.

[jprotze]

Thanks for the thorough explanation. While I won't get to this soon, I ran a quick experiment, disabling the assembler versions of __kmp_invoke_microtask on x86, rebuilding libomp.so and rerunning the openmp tests: now I get the same error on i386-pc-solaris2.11, so this isn't SPARC-specific but related to the C implementation of __kmp_invoke_microtask.

While adding more cases for the C implementation would allow the test to pass, it would break the intention of the test: check whether the implementation allows to specify more scoped variables than the C implementation can directly handle, since there will always be some limit with the C function.