On Wed, Apr 14, 2021 at 12:46:43PM -0500, Segher Boessenkool wrote:
> The REGNO checks work fine for pseudos as well. But, why does it do
> this at all, instead of using match_dup? That should be clearer.
Because with the hard regs it has different modes, so match_dup
wouldn't work.
We are talking here about:
Trying 7, 2 -> 8:
7: r98:SI=x0:SI<<0xb&0x7f800
REG_DEAD x0:QI
2: r96:SI=zero_extend(x0:QI)
8: r97:SI=r98:SI|r96:SI
REG_DEAD r98:SI
REG_DEAD r96:SI
Failed to match this instruction:
(set (reg:SI 97)
(ior:SI (and:SI (ashift:SI (reg:SI 0 x0 [ i ])
(const_int 11 [0xb]))
(const_int 522240 [0x7f800]))
(zero_extend:SI (reg:QI 0 x0 [ i ]))))
Failed to match this instruction:
(set (reg:SI 97)
(ior:SI (and:SI (ashift:SI (reg:SI 0 x0 [ i ])
(const_int 11 [0xb]))
(const_int 522240 [0x7f800]))
(and:SI (reg:SI 0 x0)
(const_int 255 [0xff]))))
Splitting with gen_split_28 (aarch64.md:4434)
Successfully matched this instruction:
(set (reg:SI 99)
(zero_extend:SI (reg:QI 0 x0 [ i ])))
Successfully matched this instruction:
(set (reg:SI 97)
(ior:SI (ashift:SI (reg:SI 99)
(const_int 11 [0xb]))
(reg:SI 99)))
match_dup means insn-recog.c calls rtx_equal_p and that returns false if the
mode is not the same.
Before combine the 3 insns are:
(insn 2 4 3 2 (set (reg/v:SI 96 [ i ])
(zero_extend:SI (reg:QI 0 x0 [ i ]))) "pr100056.c":10:1 114
{*zero_extendqisi2_aarch64}
(expr_list:REG_DEAD (reg:QI 0 x0 [ i ])
(nil)))
(note 3 2 7 2 NOTE_INSN_FUNCTION_BEG)
(insn 7 3 8 2 (set (reg:SI 98)
(ashift:SI (reg/v:SI 96 [ i ])
(const_int 11 [0xb]))) "pr100056.c":11:17 691
{*aarch64_ashl_sisd_or_int_si3}
(nil))
(insn 8 7 13 2 (set (reg:SI 97)
(ior:SI (reg:SI 98)
(reg/v:SI 96 [ i ]))) "pr100056.c":11:12 488 {iorsi3}
(expr_list:REG_DEAD (reg:SI 98)
(expr_list:REG_DEAD (reg/v:SI 96 [ i ])
(nil))))
and I must say I don't know if make_more_copies was meant to
split insn 2 into (set (reg:QI pseudo) (reg:QI 0 x0)) and
(set (reg/v:SI 96) (zero_extend:SI (reg:QI pseudo)))
or not.
> The point of make_more_copies is that the hard registers from function
> arguments are not pushed down by combine into actual instructions. This
> can be done by RA if it thinks that is a good idea, and not done if it
> thinks it is a bad idea. Having combine usurp part of the register
> allocators role is not a good idea.
>
> There are other reasons hard regs can still end up in RTL insns in
> earlier RTL passes of course, but the other changes that went together
> with make_more_copies stop combine from doing that a lot (the function
> itself makes sure every hard reg is copied to a new pseudo, because
> combining that trivial move (from that new pseudo to the pseudo it was
> copying it to already!) can still be beneficial for other reasons, all
> strange and pretty unhappy, but important on many targets).
Jakub
