On 09/13/2017 03:46 PM, Steve Ellcey wrote: > On Wed, 2017-09-13 at 14:46 -0500, Segher Boessenkool wrote: >> On Wed, Sep 13, 2017 at 06:13:50PM +0100, Kyrill Tkachov wrote: >>> >>> We are usually hesitant to add explicit subreg matching in the MD pattern >>> (though I don't remember if there's a hard rule against it). >>> In this case this looks like a missing simplification from combine >>> (simplify-rtx) so >>> I think adding it there would be better. > >> Yes, it probably belongs as a generic simplification in simplify-rtx.c; >> if there is a reason not to do that, it can be done in combine.c >> instead. > > Actually, now that I look at it some more and compare it to the arm32 > version (where we do not have this problem) I think the problem starts > well before combine. > > In arm32 rtl expansion, when reading the QI memory location, I see > these instructions get generated: > > (insn 10 3 11 2 (set (reg:SI 119) > (zero_extend:SI (mem:QI (reg/v/f:SI 117 [ string ]) [0 *string_9(D)+0 > S1 A8]))) "pr77729.c":4 -1 > (nil)) > (insn 11 10 12 2 (set (reg:QI 118) > (subreg:QI (reg:SI 119) 0)) "pr77729.c":4 -1 > (nil)) > > And in aarch64 rtl expansion I see: > > (insn 10 9 11 (set (reg:QI 81) > (mem:QI (reg/v/f:DI 80 [ string ]) [0 *string_9(D)+0 S1 A8])) > "pr77729.c":3 -1 > (nil)) > > Both of these sequences expand to ldrb but in the arm32 case I know > that I set all 32 bits of the register (even though I only want the > bottom 8 bits), but for aarch64 I only know that I set the bottom 8 > bits and I don't know anything about the higher bits, meaning I have to > keep the AND instruction to mask out the upper bits on aarch64. It's one of the reasons I discourage subregs -- the number of cases where we can optimize based on the "don't care" semantics are relatively small in my experience and I consistently see cases where the "don't care" property of the subreg turns into "don't know" and suppresses downstream optimizations.
It's always a judgment call, but more and more often I find myself pushing towards defining those bits using a zero/sign extension, bit operation or whatever rather than using subregs. > > I think we should change the movqi/movhi expansions on aarch64 to > recognize that the ldrb/ldrh instructions zero out the upper bits in > the register by generating rtl like arm32 does. Is LOAD_EXTEND_OP defined for aarch64? It may also be worth looking at ree.c -- my recollection is that it didn't handle subregs, but it could and probably should. jeff