Bugs in MIPS FUNCTION_ARG_REGNO_P ?

[Mark Mitchell]

Jim Wilson wrote:

> On Tue, 2004-03-02 at 23:25, Mark Mitchell wrote:
> 
>>>>#define FUNCTION_ARG_REGNO_P(N)					\
>>>> ((IN_RANGE((N), GP_ARG_FIRST, GP_ARG_LAST)			\
>>>>   || (IN_RANGE((N), FP_ARG_FIRST, FP_ARG_LAST)		\
>>>>	&& ((N) % FP_INC == 0) && mips_abi != ABI_O64))		\
>>>>  && !fixed_regs[N])
> 
> 
> This changed between 3.2 and 3.3.  3.2 has
> #define FUNCTION_ARG_REGNO_P(N)                                 \
>   (((N) >= GP_ARG_FIRST && (N) <= GP_ARG_LAST)                  \
>    || (! TARGET_SOFT_FLOAT                                      \
>        && ((N) >= FP_ARG_FIRST && (N) <= FP_ARG_LAST)           \
>        && (TARGET_FLOAT64 || (0 == (N) % 2))))
> which will work.
> 
> Note that a || was changed to a && and a set of parentheses was dropped,
> which breaks the code.
> 
> The key point here is that O32 uses f12/f14 because we can't access odd
> numbered registers, but O64 uses f12/f13/f14/f15.  This assumes that O64
> always uses TARGET_FLOAT64, but that is probably a reasonable
> assumption.

Yes.

The current situation is optimization bugs waiting to happen.  We are 
lying to the optimizers, telling them that these registers cannot be 
used for argument-passing, even though they can be so used.

I think this line

   && ((N) % FP_INC == 0) && mips_abi != ABI_O64))		

should become:

   && (((N) % FP_INC == 0) || mips_abi == ABI_O64)))

which would say that the register is an argument register if it is even, 
*or* if we are on o64, because then all of $f12-$f15 are usable.

Do you agree?

I don't have a good o64 target environment for running DejaGNU tests, so 
I could make the change but I'd have a hard time testing it properly. 
Does anyone have an o64 system around to test on?

If not, how do you feel about considering than an obvious fix?

Thanks,

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com