Fix openmp on PowerPC64-BE-ELFv2 ABI on FreeBSD

[adalava]

Summary

openmp library incorrectly assumes OpenPower ELF v2 ABI is only used on little endian systems. This is not true, since ELFv2 ABI supports both big and little.

This patch correct this assumption by checking CALL_ELF variable instead of endiannes, and unbreaks openmp on FreeBSD/powerpc64-ELFv2 port (https://wiki.freebsd.org/powerpc/llvm-elfv2)

*I set the reviewers based on git log. Please feel free to change.

[hbae]

openmp/runtime/src/z_Linux_asm.S

1370

Change V2 to v2.

[jdoerfert]

Is there a good way to test this? If so, we should, if not, this seems fine to me (except the typo @hbae found).

[adalava]

In D67190#1670325, @jdoerfert wrote:

Is there a good way to test this? If so, we should, if not, this seems fine to me (except the typo @hbae found).

I checked openmp sanity by running the code bellow. It runs now, before patch it crashed with segmentation fault.

I'm not sure we could do an automated test on llvm side. Any idea/guidance is welcome.

#include <stdio.h>


void bar(int i, float p, double q){
	 printf("%d %f %f\n", i, p, q);
	 fflush(stdout);
 }

 void foo(int N) {
 	float p = 3;
 	double q = 5;
 	N = 7;

	#pragma omp parallel for firstprivate(q)
	for (int i = 2; i < N; i++) {
		bar(i, p, q);
	}
}


int main()
{
	foo(0);
}

[adalava]

update fixing the typo found during review.

thanks !

[AndreyChurbanov]

LGTM

[adalava]

As I'm not a commiter, may someone commit it, if it's the case, please?

note for myself: this patch is already integrated to FreeBSD src base.

openmp/runtime/src/kmp_platform.h

 #define KMP_OS_UNIX 1
 #endif
 
 /* ---------------------- Architecture recognition ------------------- */
 
 #define KMP_ARCH_X86 0
 #define KMP_ARCH_X86_64 0
 #define KMP_ARCH_AARCH64 0
-#define KMP_ARCH_PPC64_BE 0
-#define KMP_ARCH_PPC64_LE 0
-#define KMP_ARCH_PPC64 (KMP_ARCH_PPC64_LE || KMP_ARCH_PPC64_BE)
+#define KMP_ARCH_PPC64_ELFv1 0
+#define KMP_ARCH_PPC64_ELFv2 0
+#define KMP_ARCH_PPC64 (KMP_ARCH_PPC64_ELFv2 || KMP_ARCH_PPC64_ELFv1)
 #define KMP_ARCH_MIPS 0
 #define KMP_ARCH_MIPS64 0
 #define KMP_ARCH_RISCV64 0
 
 #if KMP_OS_WINDOWS
 #if defined(_M_AMD64) || defined(__x86_64)
 #undef KMP_ARCH_X86_64
 #define KMP_ARCH_X86_64 1
 #else
 #undef KMP_ARCH_X86
 #define KMP_ARCH_X86 1
 #endif
 #endif
 
 #if KMP_OS_UNIX
 #if defined __x86_64
 #undef KMP_ARCH_X86_64
 #define KMP_ARCH_X86_64 1
 #elif defined __i386
 #undef KMP_ARCH_X86
 #define KMP_ARCH_X86 1
 #elif defined __powerpc64__
-#if defined __LITTLE_ENDIAN__
-#undef KMP_ARCH_PPC64_LE
-#define KMP_ARCH_PPC64_LE 1
+#if defined(_CALL_ELF) && _CALL_ELF == 2
+#undef KMP_ARCH_PPC64_ELFv2
+#define KMP_ARCH_PPC64_ELFv2 1
 #else
-#undef KMP_ARCH_PPC64_BE
-#define KMP_ARCH_PPC64_BE 1
+#undef KMP_ARCH_PPC64_ELFv1
+#define KMP_ARCH_PPC64_ELFv1 1
 #endif
 #elif defined __aarch64__
 #undef KMP_ARCH_AARCH64
 #define KMP_ARCH_AARCH64 1
 #elif defined __mips__
 #if defined __mips64
 #undef KMP_ARCH_MIPS64
 #define KMP_ARCH_MIPS64 1

openmp/runtime/src/z_Linux_asm.S

 //  r5: tid
 //  r6: argc
 //  r7: p_argv
 //  r8: &exit_frame
 //
 // return:  r3  (always 1/TRUE)
 //
   .text
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   .abiversion 2
 # endif
   .globl  __kmp_invoke_microtask
 
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   .p2align  4
 # else
   .p2align  2
 # endif
 
   .type __kmp_invoke_microtask,@function
 
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
 __kmp_invoke_microtask:
 .Lfunc_begin0:
 .Lfunc_gep0:
   addis 2, 12, .TOC.-.Lfunc_gep0@ha
   addi 2, 2, .TOC.-.Lfunc_gep0@l
 .Lfunc_lep0:
   .localentry __kmp_invoke_microtask, .Lfunc_lep0-.Lfunc_gep0
 # else

 // the stack frame size, and so we keep a direct copy of r1 to access our
 // register save areas and restore the r1 value before returning.
   mr 31, 1
   .cfi_def_cfa_register r31
   .cfi_offset r31, -8
   .cfi_offset lr, 16
 
 // Compute the size necessary for the local stack frame.
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   li 12, 72
 # else
   li 12, 88
 # endif
   sldi 0, 6, 3
   add 12, 0, 12
   neg 12, 12
 
// remainder to the stack.
   addi 12, 12, -6
   mtctr 12
 
 // These are set to 8 bytes before the first desired store address (we're using
 // pre-increment loads and stores in the loop below). The parameter save area
 // for the microtask begins 48 + 8*8 == 112 bytes above r1 for ELFv1 and
 // 32 + 8*8 == 96 bytes above r1 for ELFv2.
   addi 4, 4, 40
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   addi 12, 1, 88
 # else
   addi 12, 1, 104
 # endif
 
 .Lnext:
   ldu 0, 8(4)
   stdu 0, 8(12)
   bdnz .Lnext
 
 .Lcall:
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   std 2, 24(1)
   mr 12, 3
 #else
   std 2, 40(1)
 // For ELFv1, we need to load the actual function address from the function descriptor.
   ld 12, 0(3)
   ld 2, 8(3)
   ld 11, 16(3)
 #endif
 
   addi 3, 31, -20
   addi 4, 31, -24
 
   mtctr 12
   bctrl
-# if KMP_ARCH_PPC64_LE
+# if KMP_ARCH_PPC64_ELFv2
   ld 2, 24(1)
 # else
   ld 2, 40(1)
 # endif
 
 # if OMPT_SUPPORT
   li 3, 0
   std 3, 0(30)