[Bug tree-optimization/63411] [4.9/5 regression] ivopt produces wrong struct offset
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63411 --- Comment #5 from Oleg Endo --- (In reply to amker from comment #4) > > We saw similar problems on arm too. Problem is IVOPT's model doesn't know > the detail of target addressing mode. This is should be fixed by improving > function get_address_cost, but that's just difficult to expose every target > details to middle end, especially we need to keep it simple and fast. I'm trying to make some progress on PR 56590, that's how I noticed this issue here in the first place. These kind of issues could be fixed up by an AMS pass since it will have to know all the details about target address modes and costs. E.g. turning the first access in the loop into a post-inc would fix this particular case.
[Bug tree-optimization/63411] [4.9/5 regression] ivopt produces wrong struct offset
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63411 amker at gcc dot gnu.org changed: What|Removed |Added CC||amker at gcc dot gnu.org --- Comment #4 from amker at gcc dot gnu.org --- (In reply to Oleg Endo from comment #3) > (In reply to bin.cheng from comment #2) > > > bt/s.L3 > > > mov r1,r2 > > > add #3,r2 <<< wrong struct offset '3' > > But r2 is initialized with [.L14] which is _OAM3+3, so the offset actually > > is 6 here, No? > > Ouch. Yes. Sorry for the fuzz. > On SH this is a performance regression (it can do only displacements that > are scaled by the access size), but not a wrong-code regression. We saw similar problems on arm too. Problem is IVOPT's model doesn't know the detail of target addressing mode. This is should be fixed by improving function get_address_cost, but that's just difficult to expose every target details to middle end, especially we need to keep it simple and fast.
[Bug tree-optimization/63411] [4.9/5 regression] ivopt produces wrong struct offset
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63411 Oleg Endo changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |INVALID --- Comment #3 from Oleg Endo --- (In reply to bin.cheng from comment #2) > > bt/s.L3 > > mov r1,r2 > > add #3,r2 <<< wrong struct offset '3' > But r2 is initialized with [.L14] which is _OAM3+3, so the offset actually > is 6 here, No? Ouch. Yes. Sorry for the fuzz. On SH this is a performance regression (it can do only displacements that are scaled by the access size), but not a wrong-code regression.
[Bug tree-optimization/63411] [4.9/5 regression] ivopt produces wrong struct offset
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63411
--- Comment #2 from bin.cheng ---
(In reply to Oleg Endo from comment #0)
> Compiling the following as C or C++:
>
> typedef struct
> {
> unsigned char sprindex;
> unsigned char y;
> unsigned char index;
> unsigned char attr;
> unsigned char x;
> unsigned short pattern;
> } oam;
>
> extern oam OAM3[8];
>
> int test2 (unsigned c, int r)
> {
> for (unsigned i = 0; i < c; ++i)
> {
> oam* s = &(OAM3[i]);
> if (s->attr & 0x40)
> r += s->pattern;
> }
> return r;
> }
>
> on SH (big or little endian, any CPU type) with -O2 results in the following
> wrong code:
>
> _test2:
> tst r4,r4
> bt .L12
> mov.l .L14,r1
> .align 2
> .L4:
> mov.b @r1,r0
> tst #64,r0
> bt/s.L3
> mov r1,r2
> add #3,r2 <<< wrong struct offset '3'
But r2 is initialized with [.L14] which is _OAM3+3, so the offset actually is 6
here, No?
> mov.w @r2,r2<<< should be '3*2', i.e. '6'.
> extu.w r2,r2
> add r2,r5
> .L3:
> dt r4
> bf/s.L4
> add #8,r1
> .L12:
> rts
> mov r5,r0
> .L15:
> .align 2
> .L14:
> .long _OAM3+3
> .size _test2, .-_test2
> .ident "GCC: (GNU) 4.9.2 20140929 (prerelease)"
>
>
> Disabling ivopt with '-fno-ivopts' produces correct code:
>
> _test2:
> tst r4,r4
> bt .L12
> mov #0,r2
> mov.l .L14,r3
> .align 2
> .L4:
> mov r2,r1
> shll2 r1
> add r1,r1
> add r3,r1
> mov.b @(3,r1),r0
> tst #64,r0
> bt .L3
> mov.w @(6,r1),r0<<< correct struct offset '6'
> extu.w r0,r1
> add r1,r5
> .L3:
> dt r4
> bf/s.L4
> add #1,r2
> .L12:
> rts
> mov r5,r0
> .L15:
> .align 2
> .L14:
> .long _OAM3
> .size _test2, .-_test2
> .ident "GCC: (GNU) 4.9.2 20140929 (prerelease)"
>
>
> I haven't checked whether this happens on other targets, but I guess this is
> not SH specific.
>
> This happens on the current 4.9 branch and trunk. 4.8 branch is not
> affected.
[Bug tree-optimization/63411] [4.9/5 regression] ivopt produces wrong struct offset
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63411 Oleg Endo changed: What|Removed |Added CC||amker.cheng at gmail dot com, ||ebotcazou at gcc dot gnu.org, ||rguenth at gcc dot gnu.org --- Comment #1 from Oleg Endo --- CC'ing some people who have recently worked on tree-ssa-loop-ivopts.c according to the 4.9 changelog.
