Re: [patch] fix reload errors in fisttp float128 conversions

[Eric Christopher <echristo at apple dot com>]

> The use of this macro (and SCALAR_MODE_FLOAT_P) to limit operands to  
> X87 modes through i386.md (and i386.c) is wrong. We should define  
> new macro (say, X87_FLOAT_MODE_P, just to be consistent with  
> SSE_FLOAT_MODE_P) and explicitly check all X87 supported modes in it.
>
> At this moment, I suggest to review all FLOAT_MODE_P and  
> SCALAR_FLOAT_MODE_P uses in i386.md directory. It looks to me, that  
> all uses have to be changed to this new X87_FLOAT_MODE_P macro, as I  
> don't remeber x86 targets having HW DFP insns ;).

OK, I did this in i386.md. I was going to do it in i386.c, but was a  
little more involved than I had time for at the moment to verify each  
use. More to the point a few places may actually use FLOAT_MODE_P or  
SCALAR_FLOAT_MODE_P correctly. The uses in the md file I did verify  
and believe to be correct.

Tested on i386 and x86_64-darwin. OK for mainline? 4.2?

-eric

2007-04-09  Eric Christopher  <echristo@gmail.com>

	* config/i386/i386.h (X87_FLOAT_MODE_P): New.
	* config/i386/i386.md: Use. Replace uses of
	FLOAT_MODE_P.
	* config/i386/i386.c (output_fix_trunc): Assert that
	we don't have TFmode operand[1].

Attachment: i386.diff.txt
Description: Text document