clang doesn't support fp stack in inline asm yet

[nunoplopes]

Bugzilla Link	3970
Resolution	FIXED
Resolved on	Apr 19, 2009 12:32
Version	unspecified
OS	Linux
CC	@efriedma-quic

Extended Description

$ clang-c -c a.c

with a.c:

double floor(double x)
{
register long double val;

__asm __volatile (
		"frndint\n"
		:
		"=t" (val)
		:
		"0" (x)
		);

}

gives:
"Unsupported asm: input constraint with a matching output constraint of incompatible type!"

[efriedma-quic]

There's a missing diagnostic here, but llvm-gcc also gives an error...

[nunoplopes]

FYI, I got this code from a header of glibc 2.8:

530 "/usr/include/bits/mathinline.h" 3 4

double floor(double __x)
{
register long double __value;
register int __ignore;
unsigned short int __cw;
unsigned short int __cwtmp;
__asm __volatile ("fnstcw %3\n\t"
"movzwl %3, %1\n\t"
"andl $0xf3ff, %1\n\t"
"orl $0x0400, %1\n\t"
"movw %w1, %2\n\t"
"fldcw %2\n\t"
"frndint\n\t"
"fldcw %3"
: "=t" (__value), "=&q" (__ignore), "=m" (__cwtmp), "=m" (__cw)
: "0" (__x));
return __value;
}

[efriedma-quic]

Does compiling with -D__NO_MATH_INLINES help for the original code? If so, I'll add that to X86TargetInfo::getTargetDefines. (Related: bug 879.)

[lattner]

Yeah, that's probably the best solution for now.

[nunoplopes]

Does compiling with -D__NO_MATH_INLINES help for the original code? If so,
I'll add that to X86TargetInfo::getTargetDefines. (Related: bug 879.)

yes it does!

[lattner]

Ok, here ya go:
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20090413/015858.html
http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20090413/015859.html