Bernd Schmidt schrieb: > On 06/01/2011 04:00 PM, Georg-Johann Lay wrote: >> Eric Botcazou schrieb: >>>> You are right, I was staring at the wrong place. subreg of hardreg >>>> should not be there. >>> You can take a look at PR target/48830, this is a related problem for the >>> SPARC where reload generates: >>> >>> (set (reg:SI 708 [ D.2989+4 ]) >>> (subreg:SI (reg:DI 72 %f40) 4)) >>> >>> and (subreg:SI (reg:DI 72 %f40) 4) isn't simplifiable either. H.P. wrote a >>> tentative patch for the subreg machinery to forbid this. Other references >>> are: >>> http://gcc.gnu.org/ml/gcc-patches/2008-07/msg01688.html >>> http://gcc.gnu.org/ml/gcc-patches/2008-08/msg01743.html >> The problems is that find_reloads generates the wrong reload type. >> >> It sets the reload_type of "=" operands to RELOAD_WRITE. But for >> subregs as above, there are bits between inner mode and outer mode >> that already contain meaningful data and that data isn't wiped out by >> writing to the subreg (because outer mode is > word_mode). >> >> Thus, a patch could look like that: > [...] >> Is this a reasonable fix? > > I don't really think so. You'll need to find out where the subreg of a > hard register was made and stop it from doing that. > > Bernd
The reason for why a subreg of hardreg is there during reload is that on avr, r29:r28 is the frame pointer (word_mode is QI and Pmode is HI). Because in many places of the compiler, there are tests like "if (regno == FRAME_POINTER_REGNUM)", the high byte of the frame pointer gets garbaged. The avr backend tries to work around that by allowing HI in the frame pointer but disallowing QI; this happens in avr_hard regno_mode_ok. There are other odd places, for example, there is an elimination rule for the low-part of fp/sp andn elimination rule for the high part. For the test program, allowing QI in r29:r28 yield correct code. But I do not know enough of frame pointer elimination or spilling of its parts as to be sure that allowing QI won't pop up problems in other places. Actually, disallowing QI was introduced because there /were/ problems. So if anyone could give the statement that mult-word_mode frame pointers and its parts and elimination are treated correctly throughout gcc today, I will make a patch to fix avr_hard_regno_mode_ok. The old lreg/greg allocator didn't allocate the inner HI reg to r29:r28, so that the code was correct and a potential defiecency in reload made no problems. But IRA generates (subreg:QI (reg:HI 28)) because it only looks at the inner part. Johann
