https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54089
--- Comment #96 from Alexander Klepikov <klepikov.alex+bugs at gmail dot com>
---
(In reply to Oleg Endo from comment #95)
> The infinite loop is in splitting of the 'ashrsi3_n_call' pattern with the
> constant 1. This is because 'ashrsi3_n_call' match overlaps with the 'shar'
> pattern.
Yes, I confirm that. 'operands[2] != const1_rtx' works well, thank you!
> Another point is that you had the
> 'cfun->machine->ashrsi3_libcall_expanded++;' in the wrong place. It was
> counting up even if it wouldn't have emitted the libcall.
Not really. 'expand_ashiftrt' could emit
mov #imm, r1
shad r1, r0
and 'mov' instruction would be loop invariant if there's a loop. It looks like
'ashrsi3_libcall_expanded' is a bad name. I think name
'ashrsi3_n_call_expanded' would be more appropriate.
> However, there is one case from your previous posts in PR 49263:
>
> int f_rshift(char v){
> return v >> S;
> }
>
> This will not work on SH2 and expand to a libcall.
I think you mean SH2A, because the same is going on with SH4.
> On SH4 the combine pass
> converts it into:
>
> Successfully matched this instruction:
> (set (reg:SI 166)
> (ashiftrt:SI (reg/v:SI 164 [ v ])
> (const_int 31 [0x1f])))
>
> But it doesn't even try to do so with the 'ashrsi3_n_call' pattern. Not
> sure why ...
Well, the same thing is going on when using vanilla GCC. It looks like it's
happening due to char sign extension. Then instruction is catched by
'ashrsi2_31' pattern. In another words, it looks to me like an optimization.