On 9/26/19 9:22 AM, Jakub Jelinek wrote: > On Thu, Sep 26, 2019 at 09:12:34AM -0600, Jeff Law wrote: >> On 9/26/19 12:49 AM, Jakub Jelinek wrote: >>> On Wed, Sep 25, 2019 at 10:06:13PM -0600, Jeff Law wrote: >>>> (insn 13 12 14 2 (set (reg:SI 124) >>>> (const_int -939524096 [0xffffffffc8000000])) "j.c":10:54 161 >>>> {*arm_movsi_insn} >>>> (nil)) >>>> >>>> (insn 14 13 16 2 (parallel [ >>>> (set (reg:SI 132) >>>> (plus:SI (mult:SI (zero_extend:DI (reg/v:SI 115 [ sec ])) >>>> (zero_extend:DI (reg:SI 124))) >>>> (reg:SI 130))) >>> >>> IMNSHO the bug is just in the backend, the above is not valid RTL. >>> SImode MULT has to have SImode operands, not DImode operands. >> Hmm, yea, I think you're right. I totally missed that last night. >> >> FWIW, we do allow the modes of the input operands to differ. But I >> don't think either is allowed to be wider than the result mode of the >> multiplication. > > Sure, there are some rtl codes that can have different operands, but most of > them don't. The exceptions are shifts/rotates (the second operand can have > different mode, but result/first operand needs to be the same), comparisons > (result mode can be different from operand modes, but operand mode needs to > be the same), zero_extract/sign_extract, some vec_* etc. > Most binary arithmetic rtxes need to have the same mode for the result and > both arguments though. Right. My point is that the multiplication patterns are an exception as well.
I don't like the exceptions for the multiplication patterns as I think the same effect could have been achieved with more zero/sign extractions. Jeff