enable auto-vectorization

[chansarav <chansarav at gmail dot com>]

Hi,

I am studying on enabling auto-vectorization for a port.

Auto-vectorization is done for the command line option "-O2" for flag
"-ftree-vectorize". For command line option "-O3" auto-vectorization
is enabled by default.

There is a gimple pass "pass_vectorize" that would do the job of
auto-vectorization at gimple IR level. So it is enough to define the
machine descriptions (for vectorization) to generate the rtls for
vector instructions during the rtl generation phase.

On net, I found the following references for the machine descriptions,
macros to be defined for enabling the auto-vectorization.

1. "Extending GCC with new operations: RTL, SIMD and treecodes" by
Zbigniew Chamski

Adding SIMD vectorisation support
  - distinction between scalar and vector ops: operand modes
  - availability of vector ops: deducted from MD file
  - specify supported vector length: <target>.h
         #define UNITS_PER_SIMD_WORD max_SIMD_bytes
  - specify supported vector modes: <target>-modes.def
    /* Vector modes. */
    VECTOR_MODES (INT, 8);     /*       V8QI V4HI V2SI */
    VECTOR_MODES (INT, 16);    /* V16QI V8HI V4SI V2DI */
    VECTOR_MODE (INT, DI, 1);
    VECTOR_MODES (FLOAT, 8);   /*            V4HF V2SF */
    VECTOR_MODES (FLOAT, 16);  /*       V8HF V4SF V2DF */
  - build the compiler and invoke it with -O2 -ftree-vectorize

2. "GCC Tutorial – The compilation flow of the auto-vectorizer" by Dorit Nuzman.

Enabling vectorization for a new port
 Basic features:
  <target.md>
  - distinction between scalar and vector ops: operand modes
  - availability of vector ops: deduced from MD file
  <target>.h
  - specify supported vector length in bytes:
  #define UNITS_PER_SIMD_WORD 16
  <target>-modes.def
  - specify supported vector modes:
     /* Vector modes. */
     VECTOR_MODES (INT, 8); /* V8QI V4HI V2SI */
     VECTOR_MODES (INT, 16); /* V16QI V8HI V4SI V2DI */
     VECTOR_MODE (INT, DI, 1);
     VECTOR_MODES (FLOAT, 8); /* V4HF V2SF */
     VECTOR_MODES (FLOAT, 16); /* V8HF V4SF V2DF */


I experimented as said above for a port. But auto-vectorization is not enabled.

Is my current understanding correct?
Is there some other references I should refer that would help to
identify the required macros, machine descriptions to enable
vectorization?

Also I searched the "gcc-help mailing list archives" regarding this. I
didn't find any references there.


Thank you.