[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Richard Sandiford said on this PR, probably lost due to sourceware migration:

Sorry for the slow reply, been a bit of a hectic week.                          

I think both fixes would be valid.  Like you say, the address                   
parsing code isn't yet ready to handle addresses that apply                     
an offset *before* the address "mutations".  That's because                     
no target has yet wanted to support such an address.  So as                     
things stand, the current address is not valid and shouldn't                    
have been created.                                                              

In some ways this is similar to creating an invalid highpart                    
subreg for the upper word of a doubleword vector hard register,                 
or a subreg that falls foul of some simplify_subreg_regno rule.                 
The RA has code to avoid doing that, see init_subregs_of_mode                   
and its users.  We could do something similar here for                          
REG_EQUIV MEMs.  One option would be to key off whether                         
strip_address_mutations is a no-op on the address.                              
Another would be to check whether each required sub-MEM                         
is valid for its mode and offset.  The latter would be                          
more elaborate but might produce better code in general,                        
not just for cases like this.                                                   

Like you say in comment 4, even the zero-offset half isn't                      
actually a valid address for PowerPC, so either of the two                      
options should give better code as well as fixing the bug.                      

On the other hand, the idea was always that address_info                        
and its users could be extended if new targets have new                         
requirements.  So if we want to make this operation valid                       
then that would be OK too.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Any progress?

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #17 from Peter Bergner <bergner at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #16)
> Any progress?

I'm sorry, I've been swamped with other things.  Even so, given this (up to now
latent) bug has been there a while, and any patch here will affect all other
architectures, do we really want to patch this area so close to the release? 
Using -O2 is a workaround.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #18 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Yes, we want to fix it, unless you want to revert the PR93658 change even for
GCC 10 and reapply only to 11 once this bug is fixed too.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #19 from Peter Bergner <bergner at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #18)
> Yes, we want to fix it, unless you want to revert the PR93658 change even
> for GCC 10 and reapply only to 11 once this bug is fixed too.

Ok, let me take another stab at fixing this before we go the other route.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #20 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Have you managed to make some progress?  This is one of the last 10 P1 blockers
of the release.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #21 from Peter Bergner <bergner at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #20)
> Have you managed to make some progress?  This is one of the last 10 P1
> blockers of the release.

I'm still working on it.  I have a patch that fixes the issue, but I'm not sure
yet whether it's the "right" patch.  Hopefully I'll know soon.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #22 from Peter Bergner <bergner at gcc dot gnu.org> ---
To be more specific, I have implemented the hook cannot_substitute_mem_equiv_p
for rs6000 that rejects these and: altivec addresses.  The nice thing about the
patch is that it only affects rs6000, whereas a lra-constraints.c or rtlanal.c
change could affect everyone.

I have confirmed that the test case in Comment #1 compiles and we get the same
asm generated as a compiler built with the revision before my patch that caused
the ICE.  I also confirmed that test case pr93658.c which used to go into an
infinite loop in LRA before my previous patch still compiles fine.

I'm running the testsuite now to see whether there are any regressions.

Whether this is the best long term fix is unknown right now (it could be), but
it might be the best thing for GCC 10 with the release so close.  I'd rather
not revert the fix for PR93658 just to hide this bug again.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #23 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(cannot_substitute_mem_equiv_p,
 "A target hook which returns @code{true} if @var{subst} can't\n\
substitute safely pseudos with equivalent memory values during\n\
register allocation.\n\

I guess "ICEs the compiler" falls under "not safely" ;-)

Since AltiVec addresses are "legacy", this is a fine change (well
I haven't seen the actual patch yet, heh).

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

--- Comment #24 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Peter Bergner <bergner@gcc.gnu.org>:

https://gcc.gnu.org/g:dd75498db79675a1a0b73c25e5f110969ee72d9d

commit r10-7764-gdd75498db79675a1a0b73c25e5f110969ee72d9d
Author: Peter Bergner <bergner@linux.ibm.com>
Date:   Thu Apr 16 23:26:41 2020 -0500

    rs6000: Fix ICE in decompose_normal_address. [PR93974]

    Fix an ICE in decompose_normal_address(), which cannot handle Altivec AND:
    addresses, by disallowing them via implementing the target hook
    rs6000_cannot_substitute_mem_equiv_p.

    gcc/
            PR rtl-optimization/93974
            * config/rs6000/rs6000.c (TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P):
Define.
            (rs6000_cannot_substitute_mem_equiv_p): New function.

    gcc/testsuite/
            PR rtl-optimization/93974
            * g++.dg/pr93974.C: New test.

[Bug rtl-optimization/93974] [10 Regression] ICE in decompose_normal_address, at rtlanal.c:6403 on powerpc64le-linux-gnu since r10-6762

[bergner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93974

Peter Bergner <bergner at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #25 from Peter Bergner <bergner at gcc dot gnu.org> ---
Patch committed.  Fixed.