This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [new-regalloc] fix_truncdi_nomemory, fix_truncdi_memory.


> Hello Michael, Jan.
> 
> I have successfully bootstrapped C,C++,f77,java with the following
> patch.
> (I have bootstrapped new regalloc with enabled pre-reload on
> i586-pc-linux-gnu.)
> 
> Few months ago I have already asked about constraints from
> fix_truncdi_nomemory and fix_truncdi_memory.
> Problem:
> (define_insn "fix_truncdi_memory"
>   [(set (match_operand:DI 0 "memory_operand" "=m")
> 	(fix:DI (match_operand 1 "register_operand" "f")))
> -----------------------------^^ No mode definition here (I know why)
>    (use (match_operand:HI 2 "memory_operand" "m"))
>    (use (match_operand:HI 3 "memory_operand" "m"))
>    (clobber (match_scratch:DF 4 "=&1f"))]
> ---------------------------^^ ----^^^ '=&1' refers to operands[1]
>   "TARGET_80387 && FLOAT_MODE_P (GET_MODE (operands[1]))
>    && (!SSE_FLOAT_MODE_P (GET_MODE (operands[1])) || !TARGET_64BIT)"
>   "* operands[5] = operands[4]; return output_fix_trunc (insn, operands);"
>   [(set_attr "type" "fistp")])
> 
> operands[4] have a mode definition (DFmode), but operands[4] can be
> equal to operands[1] which can have different mode - IMHO it's wrong.
> Also I can't understand differences between '=&1f' and just '=f'
> What is '=&1f':
> 1. =&1;
>    it's operand which will be same as operands[1] and "the
>    previous value is discarded and replaced" like gcc info say.
> 2. =&f;
>    it's any register exclude register from operands[1].
> 
> So, =&1f is any discarded and replaced register exclude operands[1]
> or operands[1] which will be discarded and replaced.
> IMHO: =&1f equal to =f.
> 
> Am I wrong ?
I see now the thread :)
The semantic of =&1f is that the operand can be eighter equivalent to
input operand 1 (so the register is clobbered in place) or some other FP
register, but different from all the input registers.
So & is ignored for "1" but holds for "f".

It is different from "=f" that may be same to other floating point input
register. As I see here, there is only one FP input operand, so the
semantic is equivalent for this pattern, but new regalloc should
probably keep the "=&1f" style working as it is usefull in other cases
and always worked and is used by some other ports too.

Honza


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]