Re: New regression on ARM Linux

[Richard Biener <rguenther@suse.de>]

On Mon, 30 Mar 2015, Richard Biener wrote:

> On Mon, 30 Mar 2015, Alan Lawrence wrote:
> 
> > ...actually attach the testcase...
> 
> What compile options?

Just tried -O2.  The GIMPLE IL assumes 64bit alignment of .LC0 but
I can't see anything not guaranteeing that:

        .section        .rodata
        .align  3
.LANCHOR0 = . + 0
.LC1:
        .ascii  "%d %g %d\012\000"
        .space  6
.LC0:
        .word   7
        .space  4
        .word   0
        .word   1075838976
        .word   9
        .space  4

maybe there is some more generic code-gen bug for aligned aggregate
copy?  That is, the patch tells the backend that the loads and
stores to the 'int' vars (which have padding followed) is aligned
to 8 bytes.

I don't see what is wrong in the final assembler, but maybe
some endian issue exists?  The code looks quite ugly though ;)

Richard.

> 
> > Alan Lawrence wrote:
> > > We've been seeing a bunch of new failures in the *libffi* testsuite on ARM
> > > Linux (arm-none-linux-gnueabi, arm-none-linux-gnueabihf), following this
> > > one-liner fix. I've reduced the testcase down to the attached (including
> > > removing any dependency on libffi); with gcc r221347, this prints the
> > > expected
> > > 7 8 9
> > > whereas with gcc r221348, instead it prints
> > > 0 8 0
> > > 
> > > The action of r221348 is to change the alignment of a mem_ref, and a
> > > var_decl of b1, from 32 to 64; both have type
> > >   type <record_type 0x2b9b8d428d20 cls_struct_16byte sizes-gimplified type_0
> > > BLK
> > >          size <integer_cst 0x2b9b8d3720a8 constant 192>
> > >          unit size <integer_cst 0x2b9b8d372078 constant 24>
> > >          align 64 symtab 0 alias set 1 canonical type 0x2b9b8d428d20
> > >          fields <field_decl 0x2b9b8d42b098 a type <integer_type
> > > 0x2b9b8d092690 int>
> > >              SI file reduced.c line 12 col 7
> > >              size <integer_cst 0x2b9b8d08eeb8 constant 32>
> > >              unit size <integer_cst 0x2b9b8d08eed0 constant 4>
> > >              align 32 offset_align 64
> > >              offset <integer_cst 0x2b9b8d08eee8 constant 0>
> > >              bit offset <integer_cst 0x2b9b8d08ef48 constant 0> context
> > > <record_type 0x2b9b8d428d20 cls_struct_16byte> chain <field_decl
> > > 0x2b9b8d42b130 b>> context <translation_unit_decl 0x2b9b8d4232d0 D.6070>
> > >          pointer_to_this <pointer_type 0x2b9b8d42d0a8> chain <type_decl
> > > 0x2b9b8d42b000 D.6044>>
> > > 
> > > The tree-optimized output is the same with both compilers (as this does not
> > > mention alignment); the expand output differs.
> > > 
> > > Still investigating...
> > > 
> > > --Alan
> > > 
> > > 
> > > Richard Biener wrote:
> > > > This fixes a vectorizer testcase regression on powerpc where SRA
> > > > drops alignment info unnecessarily.
> > > > 
> > > > Bootstrapped and tested on x86_64-unknown-linux-gnu, applied.
> > > > 
> > > > Richard.
> > > > 
> > > > 2015-03-11  Richard Biener  <rguenther@suse.de>
> > > > 
> > > > 	PR tree-optimization/65310
> > > > 	* tree-sra.c (build_ref_for_offset): Also preserve larger
> > > > 	alignment.
> > > > 
> > > > Index: gcc/tree-sra.c
> > > > ===================================================================
> > > > --- gcc/tree-sra.c	(revision 221324)
> > > > +++ gcc/tree-sra.c	(working copy)
> > > > @@ -1597,7 +1597,7 @@ build_ref_for_offset (location_t loc, tr
> > > >    misalign = (misalign + offset) & (align - 1);
> > > >    if (misalign != 0)
> > > >      align = (misalign & -misalign);
> > > > -  if (align < TYPE_ALIGN (exp_type))
> > > > +  if (align != TYPE_ALIGN (exp_type))
> > > >      exp_type = build_aligned_type (exp_type, align);
> > > >     mem_ref = fold_build2_loc (loc, MEM_REF, exp_type, base, off);
> > > > 
> > > > 
> > > 
> > > 
> > 
> > 
> > 
> > 
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)