Add non-SSE wrapper for __kmp_{load,store}_mxcsr

[dim]

Summary

To be able to successfully build OpenMP on FreeBSD/i386, which still
uses i486 as its default processor, I had to provide wrappers for the
__kmp_load_mxcsr and __kmp_store_mxcsr functions.

If the compiler signals that SSE is not available, loading and storing
mxcsr does not make sense anway, so in that case the inline functions
are empty. This gives the minimum amount of code churn.

See also https://svnweb.freebsd.org/changeset/base/345283

[dim]

Ping :)

[hfinkel]

Does the code that uses __kmp_load_mxcsr expect the pointed-to value to be initialized?

[dim]

In D60916#1486011, @hfinkel wrote:

Does the code that uses __kmp_load_mxcsr expect the pointed-to value to be initialized?

Hm, that's a good one. I expect that callers will just save the value somewhere in a struct, to restore it later via __kmp_load_mxcsr, which is a nop. But it is probably best to set the pointed-to value to zero, just to be safe.

[jlpeyton]

runtime/src/kmp.h

1259

This needs to play well with Windows. Can we stick these in the above operating system split section:

#if KMP_OS_UNIX
// Code for __kmp_x86_cpuid, __kmp_load_x87...
// Your current patch 
#else 
// Windows still has these as external function in assembly file
...
static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*(p)); }
static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
#endif // KMP_OS_UNIX

[dim]

Address review comments:

• Assign zero to pointed-to value in __kmp_store_mxcsr()
• Use SSE specific stuff in KMP_OS_UNIX part only

[dim]

So does this look better now?

[jlpeyton]

LGTM

runtime/src/kmp.h

#if KMP_MIC
   __asm__ __volatile__("fstenv %0\n\t" // store FP env
                        "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
                        "fldenv %0\n\t" // load FP env back
                        : "+m"(fpu_state), "+m"(fpu_state.sw));
 #else
   __asm__ __volatile__("fnclex");
 #endif // KMP_MIC
 }
+#if __SSE__
+static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
+static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
+#else
+static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
+static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
+#endif
 #else
 // Windows still has these as external functions in assembly file
 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
 extern void __kmp_clear_x87_fpu_status_word();
-#endif // KMP_OS_UNIX
-
-#define __kmp_load_mxcsr(p) _mm_setcsr(*(p))
+static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
+#endif // KMP_OS_UNIX
 
 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
 
 #if KMP_ARCH_X86
 extern void __kmp_x86_pause(void);
 #elif KMP_MIC
 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
 // regression after removal of extra PAUSE from spin loops. Changing