On Wed, Jun 15, 2016 at 9:57 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> On the builtin-arith-overflow-p-1{2,3}.c testcases (posted earlier today)
> i?86 miscompiles e.g. t111_4mul function. Before peephole2 we have:
> (insn 9 6 50 2 (parallel [
> (set (reg:CCO 17 flags)
> (eq:CCO (mult:DI (sign_extend:DI (reg/v:SI 0 ax [orig:90 x ]
> [90]))
> (const_int -1073741824 [0xffffffffc0000000]))
> (sign_extend:DI (mult:SI (reg/v:SI 0 ax [orig:90 x ] [90])
> (const_int -1073741824 [0xffffffffc0000000])))))
> (set (reg:SI 1 dx [91])
> (mult:SI (reg/v:SI 0 ax [orig:90 x ] [90])
> (const_int -1073741824 [0xffffffffc0000000])))
> ]) builtin-arith-overflow-p-12.i:35 326 {*mulvsi4_1}
> (expr_list:REG_UNUSED (reg:SI 1 dx [91])
> (nil)))
>
> (insn 50 9 51 2 (set (reg:QI 1 dx [orig:89 _5+4 ] [89])
> (eq:QI (reg:CCO 17 flags)
> (const_int 0 [0]))) builtin-arith-overflow-p-12.i:35 631
> {*setcc_qi}
> (expr_list:REG_DEAD (reg:CCO 17 flags)
> (nil)))
>
> (insn 51 50 16 2 (set (reg:SI 1 dx [orig:89 _5+4 ] [89])
> (zero_extend:SI (reg:QI 1 dx [orig:89 _5+4 ] [89])))
> builtin-arith-overflow-p-12.i:35 130 {*zero_extendqisi2}
> (nil))
>
> and the setcc+movzbl peephole ignores the dx = ax * -1073741824
> SET in the PARALLEL, because it isn't a CLOBBER and thus initializes
> %edx before insn 9, then insn 9 overwrites it and later on we store
> the QImode part, assuming the rest is zero.
> Fixed by using reg_set_p, to actually test if the operands[3] REG
> is SET or CLOBBERed.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Please also change similar peephole2 pattern (that does a zext with an
and insn) a couple of patterns below the one you are changing.
The patch is OK for mainline and also release branches.
Thanks,
Uros.
