[Bug rtl-optimization/66275] __attribute__((sysv_abi)) with x86_64-w64-mingw32-gcc generates incorrect code

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66275

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-05-25
          Component|c                           |rtl-optimization
   Target Milestone|---                         |4.9.3
     Ever confirmed|0                           |1

--- Comment #1 from Uroš Bizjak <ubizjak at gmail dot com> ---
This problem can also be seen on x86_64-linux with gcc-6.0 and

gcc -O3 -funroll-loops -mabi=ms

In _.ce3 dump, we have following sequence:

  ...
   73: L73:
   72: NOTE_INSN_BASIC_BLOCK 5
   44: dx:DI=[si:DI+ax:DI]
      REG_EQUIV [si:DI+ax:DI]
   45: r9:DI=zero_extend(dx:QI)
   46: NOTE_INSN_DELETED
   47: xmm0:V4SI=vec_merge(vec_duplicate([r9:DI*0x4+cx:DI]),const_vector,0x1)
      REG_DEAD r9:DI
   50: NOTE_INSN_DELETED
   51: NOTE_INSN_DELETED
   52: dx:DI=zero_extract(dx:DI,0x8,0x8)
   53: NOTE_INSN_DELETED
   54: NOTE_INSN_DELETED
   55:
xmm1:V4SI=vec_merge(vec_duplicate([dx:DI*0x4+cx:DI+0x400]),const_vector,0x1)
      REG_DEAD dx:DI
   56: xmm0:V2DI=xmm0:V2DI^xmm1:V2DI
      REG_DEAD xmm1:V2DI
   57: xmm0:V2DI=xmm0:V2DI^[di:DI+ax:DI]
      REG_EQUIV [di:DI+ax:DI]
   58: [di:DI+ax:DI]=xmm0:V2DI
      REG_DEAD xmm0:V2DI
  183: dx:DI=[si:DI+ax:DI+0x10]
      REG_EQUIV [si:DI+ax:DI+0x10]
  ...

Please note (insn 183), a consumer of SI register.

In _.rnreg dump, the above sequence is transformed to:

  ...
   73: L73:
   72: NOTE_INSN_BASIC_BLOCK 5
   44: dx:DI=[si:DI+r10:DI]
      REG_EQUIV [si:DI+ax:DI]
   45: si:DI=zero_extend(dx:QI)
   46: NOTE_INSN_DELETED
   47: xmm8:V4SI=vec_merge(vec_duplicate([si:DI*0x4+cx:DI]),const_vector,0x1)
      REG_DEAD r9:DI
   50: NOTE_INSN_DELETED
   51: NOTE_INSN_DELETED
   52: ax:DI=zero_extract(dx:DI,0x8,0x8)
   53: NOTE_INSN_DELETED
   54: NOTE_INSN_DELETED
   55:
xmm9:V4SI=vec_merge(vec_duplicate([ax:DI*0x4+cx:DI+0x400]),const_vector,0x1)
      REG_DEAD dx:DI
   56: xmm8:V2DI=xmm8:V2DI^xmm9:V2DI
      REG_DEAD xmm1:V2DI
   57: xmm8:V2DI=xmm8:V2DI^[di:DI+r10:DI]
      REG_EQUIV [di:DI+ax:DI]
   58: [di:DI+r10:DI]=xmm8:V2DI
      REG_DEAD xmm0:V2DI
  183: dx:DI=[si:DI+r10:DI+0x10]
      REG_EQUIV [si:DI+ax:DI+0x10]
  ...

So, (insn 45) now indeed clobbers SI register that is needed by (insn 183).

Confirmed as RTL optimization problem, specifically, a rnreg pass problem.