[Bug tree-optimization/66917] [4.9/5/6 regression] ARM: NEON: memcpy compiles to vst1 with incorrect alignment due to SRA

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66917

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.9.4
            Summary|[4.9/5/6 regression] ARM:   |[4.9/5/6 regression] ARM:
                   |NEON: memcpy compiles to    |NEON: memcpy compiles to
                   |vld1 and vst1 with          |vst1 with incorrect
                   |incorrect alignment         |alignment due to SRA

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
I see the following, 'a' having an alignment of 8 bytes:

 <array_ref 0x7ffff68201c0
    type <integer_type 0x7ffff67f5d20 uint64_t unsigned DI
        size <integer_cst 0x7ffff68ccf18 constant 64>
        unit size <integer_cst 0x7ffff68ccf30 constant 8>
        align 64 symtab 0 alias set -1 canonical type 0x7ffff68d19d8 precision
64 min <integer_cst 0x7ffff68de210 0> max <integer_cst 0x7ffff68d7500
18446744073709551615> context <translation_unit_decl 0x7ffff7ff10f0 D.5218>
        pointer_to_this <pointer_type 0x7ffff68215e8>>

    arg 0 <component_ref 0x7ffff681f570
        type <array_type 0x7ffff6821540 type <integer_type 0x7ffff67f5d20
uint64_t>
            TI
            size <integer_cst 0x7ffff68de228 constant 128>
            unit size <integer_cst 0x7ffff68de240 constant 16>
            align 64 symtab 0 alias set -1 canonical type 0x7ffff6821690 domain
<integer_type 0x7ffff6821498>>

        arg 0 <var_decl 0x7ffff68d6900 a type <union_type 0x7ffff68213f0>
            addressable used BLK file t.c line 12 col 5 size <integer_cst
0x7ffff68de228 128> unit size <integer_cst 0x7ffff68de240 16>
            align 64 context <function_decl 0x7ffff67f6e00
test_neon_load_store_alignment> chain <var_decl 0x7ffff68d6990 b>>
        arg 1 <field_decl 0x7ffff69b3558 u type <array_type 0x7ffff6821540>
            TI file t.c line 10 col 16 size <integer_cst 0x7ffff68de228 128>
unit size <integer_cst 0x7ffff68de240 16>
            align 64 offset_align 64
            offset <integer_cst 0x7ffff68ccee8 constant 0>
            bit offset <integer_cst 0x7ffff68ccf48 constant 0> context
<union_type 0x7ffff68213f0> chain <field_decl 0x7ffff69b35f0 c>>>
    arg 1 <integer_cst 0x7ffff68de2e8 type <integer_type 0x7ffff68d1690 int>
constant 1>>

so the smoking dump isn't the reads but the write:

(insn 11 10 0 (set (mem:V2DI (reg/v/f:SI 115 [ outp ]) [0 MEM[(char *
{ref-all})outp_7(D)]+0 S16 A64])
        (reg:V2DI 116 [ vect__4.13 ])) t.c:18 -1
     (nil))

which for some reason also got 8 byte alignment.  I suppose this is
SRA at work as -fno-tree-sra fixes this.

Thus confirmed as SRA problem.