Re: [Patch, Vectorizer, SVE] fmin/fmax builtin reduction support

[Richard Biener <richard.guenther@gmail.com>]

On Wed, Dec 19, 2018 at 10:33 AM Alejandro Martinez Vicente
<Alejandro.MartinezVicente@arm.com> wrote:
>
> Hi all,
>
> Loops that use the fmin/fmax builtins can be vectorized even without
> -ffast-math using SVE's FMINNM/FMAXNM instructions. This is an example:
>
> double
> f (double *x, int n)
> {
>   double res = 100.0;
>   for (int i = 0; i < n; ++i)
>     res = __builtin_fmin (res, x[i]);
>   return res;
> }
>
> Before this patch, the compiler would generate this code (-march=armv8.2-a+sve
> -O2 -ftree-vectorize):
>
> 0000000000000000 <f>:
>    0:   7100003f        cmp     w1, #0x0
>    4:   5400018d        b.le    34 <f+0x34>
>    8:   51000422        sub     w2, w1, #0x1
>    c:   91002003        add     x3, x0, #0x8
>   10:   d2e80b21        mov     x1, #0x4059000000000000
>   14:   9e670020        fmov    d0, x1
>   18:   8b224c62        add     x2, x3, w2, uxtw #3
>   1c:   d503201f        nop
>   20:   fc408401        ldr     d1, [x0],#8
>   24:   1e617800        fminnm  d0, d0, d1
>   28:   eb02001f        cmp     x0, x2
>   2c:   54ffffa1        b.ne    20 <f+0x20>
>   30:   d65f03c0        ret
>   34:   d2e80b20        mov     x0, #0x4059000000000000
>   38:   9e670000        fmov    d0, x0
>   3c:   d65f03c0        ret
>
> After this patch, this is the code that gets generated:
>
> 0000000000000000 <f>:
>    0:   7100003f        cmp     w1, #0x0
>    4:   5400020d        b.le    44 <f+0x44>
>    8:   d2800002        mov     x2, #0x0
>    c:   25d8e3e0        ptrue   p0.d
>   10:   93407c21        sxtw    x1, w1
>   14:   90000003        adrp    x3, 0 <f>
>   18:   25804001        mov     p1.b, p0.b
>   1c:   91000063        add     x3, x3, #0x0
>   20:   85c0e060        ld1rd   {z0.d}, p0/z, [x3]
>   24:   25e11fe0        whilelo p0.d, xzr, x1
>   28:   a5e24001        ld1d    {z1.d}, p0/z, [x0, x2, lsl #3]
>   2c:   04f0e3e2        incd    x2
>   30:   65c58020        fminnm  z0.d, p0/m, z0.d, z1.d
>   34:   25e11c40        whilelo p0.d, x2, x1
>   38:   54ffff81        b.ne    28 <f+0x28>  // b.any
>   3c:   65c52400        fminnmv d0, p1, z0.d
>   40:   d65f03c0        ret
>   44:   d2e80b20        mov     x0, #0x4059000000000000
>   48:   9e670000        fmov    d0, x0
>   4c:   d65f03c0        ret
>
> This patch extends the support for reductions to include calls to internal
> functions, in addition to assign statements. For this purpose, in most places
> where a tree_code would be used, a code_helper is used instead. The code_helper
> allows to hold either a tree_code or combined_fn.
>
> This patch implements these tasks:
>
> - Detect a reduction candidate based on a call to an internal function
>   (currently only fmin or fmax).
> - Process the reduction using code_helper. This means that at several places
>   we have to check whether this is as assign-based reduction or a call-based
>   reduction.
> - Add new internal functions for the fmin/fmax reductions and for conditional
>   fmin/fmax. In architectures where ieee fmin/fmax reductions are available, it
>   is still possible to vectorize the loop using unconditional instructions.
> - Update SVE's md to support these new reductions.
> - Add new SVE tests to check that the optimal code is being generated.
>
> I tested this patch in an aarch64 machine bootstrapping the compiler and
> running the checks.

Just some quick comments based on the above and the changelog.
Using code_helper is reasonable I guess.

> Alejandro
>
> gcc/Changelog:
>
> 2018-12-18  Alejandro Martinez  <alejandro.martinezvicente@arm.com>
>
>         * gimple-match.h (code_helper_for_stmnt): New function to get a

code_helper_for_stmt I hope.

>         code_helper from an statement.
>         * internal-fn.def: New reduc_fmax_scal and reduc_fmin_scal optabs for
>         ieee fp max/min reductions

Aren't they necessarily fold_left reductions then?  Thus, should the optabs
be named accordingly fold_left_fmax_optab?

>         * optabs.def: Likewise.
>         * tree-vect-loop.c (reduction_fn_for_scalar_code): Changed function
>         signature to accept code_helper instead of tree_code. Handle the
>         fmax/fmin builtins.
>         (needs_fold_left_reduction_p): Likewise.
>         (check_reduction_path): Likewise.
>         (vect_is_simple_reduction): Use code_helper instead of tree_code. Check
>         for supported call-based reductions. Extend support for both
>         assignment-based and call-based reductions.
>         (vect_model_reduction_cost): Extend cost-model support to call-based
>         reductions (just use MAX expression).
>         (get_initial_def_for_reduction): Use code_helper instead of tree_code.
>         Extend support for both assignment-based and call-based reductions.
>         (vect_create_epilog_for_reduction): Likewise.
>         (vectorizable_reduction): Likewise.
>         * tree-vectorizer.h: include gimple-match.h for code_helper. Use
>         code_helper in check_reduction_path signature.
>         * config/aarch64/aarch64-sve.md: Added define_expand to capture new
>         reduc_fmax_scal and reduc_fmin_scal optabs.
>         * config/aarch64/iterators.md: New FMAXMINNMV and fmaxmin_uns iterators
>         to support the new define_expand.
>
> gcc/testsuite/Changelog:
>
> 2018-12-18  Alejandro Martinez  <alejandro.martinezvicente@arm.com>
>
>         * gcc.target/aarch64/sve/reduc_9.c: New test to check
>         SVE-vectorized reductions without -ffast-math.
>         * gcc.target/aarch64/sve/reduc_10.c: New test to check
>         SVE-vectorized builtin reductions without -ffast-math.