Re: Fix rs6000 fix_trunc TFmode

[Geoffrey Keating <geoffk at apple dot com>]

On 11/01/2004, at 3:20 PM, David Edelsohn wrote:

> 	Is there a reason that fix_trunctfsi2 cannot use fix_truncdfsi2
> instead of duplicating most of fix_truncdfsi2_internal in
> fix_trunctfsi2_internal?

Yes:

> 	fix_trunctfsi2 currently only is valid for TARGET_POWERPC ||
> TARGET_POWER2, which provide fctiwz instruction.  This means that GCC
> tries to call __fixtfsi when the processor does not support that
> intruction.  That libgcc function currently is not built and should not
> need to be.  Using fix_truncdfsi2 would fall back to the exisitng 
> __itrunc
> function.  Currently the rs6000-ldouble-1.c test is failing on AIX (in
> common mode without PowerPC instructions) due to this problem.
>
> 	In other words, could the define_expand be something like:
>
> (define_expand "fix_trunctfsi2"
>   [(parallel [(set (match_operand:SI 0 "gpc_reg_operand" "")
>                    (fix:SI (match_operand:TF 1 "gpc_reg_operand" "")))
>               (clobber (match_dup 2))
>               (clobber (match_dup 3))])]
>   "(DEFAULT_ABI == ABI_AIX || DEFAULT_ABI == ABI_DARWIN)
>    && TARGET_HARD_FLOAT && TARGET_FPRS && TARGET_LONG_DOUBLE_128"
> {
>   emit_insn (gen_fix_trunc_helper (operands[2], operands[1], 
> operands[3]));
>   emit_insn (gen_fix_truncdfsi (operands[0], operands[2]));
>   DONE;
> })

That would prevent RTL constant folding.