why does gcc need -fno-trapping-math to generate roundsd instructions?

why does gcc need -fno-trapping-math to generate roundsd instructions?

[Jay Foad]

I'm using GCC 4.8.1 with eglibc 2.17 on an Intel CPU with SSE 4.1.

Given this source:

#include <math.h>
double f(double x){return floor x;}

gcc -msse4.1 -O -S generates:

f:
subq $8, %rsp
call floor
addq $8, %rsp
ret

It seems I need to add -fno-trapping-math to get inline code:

f:
roundsd $1, %xmm0, %xmm0
ret

But why? If I run the first version under a debugger, the call to
"floor" eventually resolves to eglibc's __floor_sse41, which just does
this:

roundsd $0x1,%xmm0,%xmm0
retq

So what is -ftrapping-math really buying me? Why can't the compiler
generate roundsd inline in all cases?

Thanks,
Jay.

Re: why does gcc need -fno-trapping-math to generate roundsd instructions?

[Marc Glisse]

On Tue, 14 Jan 2014, Jay Foad wrote:

> I'm using GCC 4.8.1 with eglibc 2.17 on an Intel CPU with SSE 4.1.
>
> Given this source:
>
> #include <math.h>
> double f(double x){return floor x;}
>
> gcc -msse4.1 -O -S generates:
>
> f:
> subq $8, %rsp
> call floor
> addq $8, %rsp
> ret
>
> It seems I need to add -fno-trapping-math to get inline code:
>
> f:
> roundsd $1, %xmm0, %xmm0
> ret
>
> But why? If I run the first version under a debugger, the call to
> "floor" eventually resolves to eglibc's __floor_sse41, which just does
> this:
>
> roundsd $0x1,%xmm0,%xmm0
> retq
>
> So what is -ftrapping-math really buying me? Why can't the compiler
> generate roundsd inline in all cases?

First, I think you want to file a PR on bugzilla, or your message will 
just be forgotten.

The reason could be related to this note in the glibc manpage:

"SUSv2 and POSIX.1-2001 contain text about overflow (which might set
errno to ERANGE, or raise an FE_OVERFLOW exception).  In practice, the
result cannot overflow on any current machine, so this error-handling
stuff is just nonsense."

-- 
Marc Glisse