Hi,
On Mon, 19 Mar 2018, Jakub Jelinek wrote:
> The inline-asm here has "1p" constraint and we end up with
> (plus (frame) (const_int ...))
Blaeh. Note that "1p" is actually invalid:
--------------
`0', `1', `2', ... `9'
An operand that matches the specified operand number is allowed.
If a digit is used together with letters within the same
alternative, the digit should come last.
--------------
Changing the order to "p1" would disable the transformation as well,
because match_asm_constraints_1() uses strtoul on the constraint start.
But let's say we don't want to go there and reject this form for good
(though I think we should), so ...
> as input; even when the matching output is a REG, I'm not really
> sure emit_move_insn can handle arbitrary expressions (consider e.g.
> "1X" constraint), and calling reg_overlap_mentioned_p on something other
> than REG/SUBREG/MEM/constant (and a couple of selected other cases)
> will ICE too. My understanding is that the match_asm_constraints mini-pass
> is an optimization, so we can try to handle cases which make sense, but if
> there is something too difficult to handle we can just punt and let reload
> do its job or error_for_asm if it can't reload it. Especially when I
> believe such input expressions can be there only when the numeric constraint
> is mixed with some other one, otherwise it should have been a REG or MEM
> or something similar.
... this makes sense. But I think you're too generous in supporting
strange inputs:
> if (! REG_P (output)
> || rtx_equal_p (output, input)
> || (GET_MODE (input) != VOIDmode
> - && GET_MODE (input) != GET_MODE (output)))
> + && GET_MODE (input) != GET_MODE (output))
> + || !(REG_P (input) || SUBREG_P (input)
> + || MEM_P (input) || CONSTANT_P (input)))
I'd only allow REG_P (input) as well, not any of the other forms.
Ciao,
Michael.
