help removing unnecessary zero-extension

help removing unnecessary zero-extension

[William Tambe]

Given following program:

    unsigned char var;
    int main() {
        return var;
    }

And compiled using:
pu32-elf-gcc -O3 -c -save-temps test.c

Unnecessary zero-extension gets generated after a memory byte load
which already zero-extend.

LOAD_EXTEND_OP has been defined as follow:
#define LOAD_EXTEND_OP(M) ZERO_EXTEND

Find complete port at:
https://github.com/fontamsoc/gcc/commit/45840063
And machine description at:
https://github.com/fontamsoc/gcc/blob/45840063/gcc/config/pu32/pu32.md

Any idea what else can be tried to prevent the unnecessary zero-extension ?

Re: help removing unnecessary zero-extension

[Richard Sandiford]

William Tambe via Gcc-help <gcc-help@gcc.gnu.org> writes:
> Given following program:
>
>     unsigned char var;
>     int main() {
>         return var;
>     }
>
> And compiled using:
> pu32-elf-gcc -O3 -c -save-temps test.c
>
> Unnecessary zero-extension gets generated after a memory byte load
> which already zero-extend.
>
> LOAD_EXTEND_OP has been defined as follow:
> #define LOAD_EXTEND_OP(M) ZERO_EXTEND
>
> Find complete port at:
> https://github.com/fontamsoc/gcc/commit/45840063
> And machine description at:
> https://github.com/fontamsoc/gcc/blob/45840063/gcc/config/pu32/pu32.md
>
> Any idea what else can be tried to prevent the unnecessary zero-extension ?

(Thanks for sharing the links.  Unfortunately I can't look at unsubmitted
code for copyright reasons, so the below is just a guess.)

If you define LOAD_EXTEND_OP, it's still better to have a define_insn
that can zero_extend a memory source operand to a wider register
destination operand.  Ideally there should be one instruction that
handles both registers and memory -- rather than than two separate
instructions -- since that helps the register allocator to produce
better results.

E.g. the aarch64 pattern for this operation is:

(define_insn "*zero_extend<SHORT:mode><GPI:mode>2_aarch64"
  [(set (match_operand:GPI 0 "register_operand" "=r,r,w,r")
        (zero_extend:GPI (match_operand:SHORT 1 "nonimmediate_operand" "r,m,m,w")))]
  ""
  "@
   and\t%<GPI:w>0, %<GPI:w>1, <SHORT:short_mask>
   ldr<SHORT:size>\t%w0, %1
   ldr\t%<SHORT:size>0, %1
   umov\t%w0, %1.<SHORT:size>[0]"
  [(set_attr "type" "logic_imm,load_4,f_loads,neon_to_gp")
   (set_attr "arch" "*,*,fp,fp")]
)

which is quite complicated, but it's the first two alternatives
that matter here.

Thanks,
Richard