Re: [RFC] Prototype for SSE2 calling convention libm on ia32

[Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>]

Richard Henderson wrote:
> On Tue, Jun 07, 2005 at 07:21:18PM +0200, Richard Guenther wrote:
> 
>>It does so by first exposing a new function attribute "fpregparm" that
>>changes calling conventions for a single function, and second, patching
>>optabs and builtin decls and honouring it at expansion time.
> 
> 
> I think we need a different name than fpregparm.  There was another
> patch that was submitted that would pass fp arguments on the fpu stack.
> It didn't get reviewed, but...
> 
> Perhaps "sseregparm".

Ok that was my first choice, too.

> 
>>	* calls.c (expand_call_1): New function from expand_call
>>	with extra dispatch location argument.
>>	(expand_call): Wrapper to expand_call_1 with NULL_RTX
>>	dispatch location.
>>	* builtins.c (expand_builtin_mathfn_3): Honour optabs
>>	libfunc entry via expand_call_1.
> 
> 
> What in the world is this for?!?

Basically we need to be able to tell the middle-end to emit
a libcall for __builtin_sin to __libm_sse2_sin rather than
to sin.  Using expand_call, we always get sin, using
emit_library_call_value we get a libcall to __libm_sse2_sin,
but the backend never sees the fn decl or type for changing
the ABI.  So I chose to pass powi_optab->handlers[(int) mode].libfunc
as optional extra argument to expand_call (and being lazy for the
prototype I didn't want to change all callers of expand_call, so
I wrapped it).

> 
>>+   if (!TARGET_SSE2)
>>+     {
>>+       error ("%qs attribute only can be used with -msse2",
>>+ 	     IDENTIFIER_POINTER (name));
> 
> 
> No, all data types are supported with moves in -msse1.  See commentary
> elsewhere about minimizing the number of abis.

Basically the problem is for f.i.

double foo(double) __attribute__((fpregparm));
double bar(double x) { return foo(x); }

we ICE compiling without -msse2 enabled (because some pattern does
not match its constraints).  So I chose to error out for the attribute
without -msse2 and require headers to be appropriately #ifdef'ed.  The
other possibility is to ignore the attribute in this case and use
fpstack calling conventions and break at runtime.  My preference was
breaking at compile time.

> Perhaps you only want to enable some of the libcalls with sse2, but 
> this attribute is user visible.  And I'd like this attribute to go
> in separate from any libcall tweakage.

Yes, I'll split it apart.

> 
>>+   /* Use SSE registers to pass SFmode and DFmode arguments if requested
>>+      by the fpregparm attribute.  */
>>+   if (fntype && lookup_attribute ("fpregparm", TYPE_ATTRIBUTES (fntype)))
>>+     {
>>+       cum->sse_nregs = 8;
>>+       cum->float_in_sse = true;
>>+     }
>>+ 
>>    /* For local functions, pass SFmode (and DFmode for SSE2) arguments
>>       in SSE registers even for 32-bit mode and not just 3, but up to
>>       8 SSE arguments in registers.  */
>>!   else if (!TARGET_64BIT && !cum->maybe_vaarg && !cum->fastcall
>>        && cum->sse_nregs == SSE_REGPARM_MAX && fndecl
>>        && TARGET_SSE_MATH && flag_unit_at_a_time && !profile_flag)
> 
> 
> We should handle this in a separate function like ix86_function_regparm.

Ok, will do.

Thanks,
Richard.