RE: [PATCH]middle-end Add optimized float addsub without needing VEC_PERM_EXPR.

[Richard Biener <rguenther@suse.de>]

On Mon, 20 Jun 2022, Tamar Christina wrote:

> > -----Original Message-----
> > From: Richard Biener <rguenther@suse.de>
> > Sent: Saturday, June 18, 2022 11:49 AM
> > To: Andrew Pinski via Gcc-patches <gcc-patches@gcc.gnu.org>
> > Cc: Tamar Christina <Tamar.Christina@arm.com>; nd <nd@arm.com>
> > Subject: Re: [PATCH]middle-end Add optimized float addsub without
> > needing VEC_PERM_EXPR.
> > 
> > 
> > 
> > > Am 17.06.2022 um 22:34 schrieb Andrew Pinski via Gcc-patches <gcc-
> > patches@gcc.gnu.org>:
> > >
> > > On Thu, Jun 16, 2022 at 3:59 AM Tamar Christina via Gcc-patches
> > > <gcc-patches@gcc.gnu.org> wrote:
> > >>
> > >> Hi All,
> > >>
> > >> For IEEE 754 floating point formats we can replace a sequence of
> > >> alternative
> > >> +/- with fneg of a wider type followed by an fadd.  This eliminated
> > >> +the need for
> > >> using a permutation.  This patch adds a math.pd rule to recognize and
> > >> do this rewriting.
> > >
> > > I don't think this is correct. You don't check the format of the
> > > floating point to make sure this is valid (e.g. REAL_MODE_FORMAT's
> > > signbit_rw/signbit_ro field).
> 
> Yes I originally had this check, but I wondered whether it would be needed.
> I'm not aware of any vector ISA where the 32-bit and 16-bit floats don't follow
> the IEEE data layout and semantics here.
> 
> My preference would be to ask the target about the data format of its vector
> Floating points because I don't think there needs to be a direct correlation between
> The scalar and vector formats strictly speaking.   But I know Richi won't like that so
> the check is probably most likely.
> 
> > > Also would just be better if you do the xor in integer mode (using
> > > signbit_rw field for the correct bit)?
> > > And then making sure the target optimizes the xor to the neg
> > > instruction when needed?
> 
> I don't really see the advantage of this one. It's not removing an instruction
> and it's assuming the vector ISA can do integer ops on a floating point vector
> cheaply.  Since match.pd doesn't have the ability to do costing I'd rather not
> do this.
> 
> > I’m also worried about using FP operations for the negate here.  When @1 is
> > constant do we still constant fold this correctly?
> 
> We never did constant folding for this case, the folding infrastructure doesn't
> know how to fold the VEC_PERM_EXPR.  So even with @0 and @1 constant no
> folding takes place even today if we vectorize.
> 
> > 
> > For costing purposes it would be nice to make this visible to the vectorizer.
> > 
> 
> I initially wanted to use VEC_ADDSUB for this, but noticed it didn't trigger in a number of
> place I had expected it to. While looking into it I noticed it's because this follows the x86
> instruction semantics so left it alone.
> 
> It felt like adding a third pattern here might be confusing. However I can also use the SLP
> pattern matcher to rewrite it without an optab if you prefer that?
> 
> The operations will then be costed normally.
> 
> > Also is this really good for all targets?  Can there be issues with reformatting
> > when using FP ops as in your patch or with using integer XOR as suggested
> > making this more expensive than the blend?
> 
> I don't think with the fp ops alone,  since it's using two fp ops already and after the change 2 fp ops.
> and I can't image that a target would have a slow -a.

Wouldn't a target need to re-check if lanes are NaN or denormal if after
a SFmode lane operation a DFmode lane operation follows?  IIRC that is
what usually makes punning "integer" vectors as FP vectors costly.

Note one option would be to emit a multiply with { 1, -1, 1, -1 } on
GIMPLE where then targets could opt-in to handle this via a DFmode
negate via a combine pattern?  Not sure if this can be even done
starting from the vec-perm RTL IL.

I fear whether (neg:V2DF (subreg:V2DF (reg:V4SF))) is a good idea
will heavily depend on the target CPU (not only the ISA).  For RISC-V
for example I think the DF lanes do not overlap with two SF lanes
(so same with gcn I think).

Richard.

> The XOR one I wouldn't do, as the vector int and vector float could for instance be in different register
> files or FP be a co-processor etc.  Mixing FP and Integer ops in this case I can image can lead to something
> suboptimal.  Also for targets with masking/predication the VEC_PERM_EXP could potentially be lowered to
> a mask/predicate in the backend. Whereas the XOR approach is far less likely.
> 
> Thanks,
> Tamar
> 
> > 
> > Richard.
> > 
> > > Thanks,
> > > Andrew Pinski
> > >
> > >
> > >
> > >>
> > >> For
> > >>
> > >> void f (float *restrict a, float *restrict b, float *res, int n) {
> > >>   for (int i = 0; i < (n & -4); i+=2)
> > >>    {
> > >>      res[i+0] = a[i+0] + b[i+0];
> > >>      res[i+1] = a[i+1] - b[i+1];
> > >>    }
> > >> }
> > >>
> > >> we generate:
> > >>
> > >> .L3:
> > >>        ldr     q1, [x1, x3]
> > >>        ldr     q0, [x0, x3]
> > >>        fneg    v1.2d, v1.2d
> > >>        fadd    v0.4s, v0.4s, v1.4s
> > >>        str     q0, [x2, x3]
> > >>        add     x3, x3, 16
> > >>        cmp     x3, x4
> > >>        bne     .L3
> > >>
> > >> now instead of:
> > >>
> > >> .L3:
> > >>        ldr     q1, [x0, x3]
> > >>        ldr     q2, [x1, x3]
> > >>        fadd    v0.4s, v1.4s, v2.4s
> > >>        fsub    v1.4s, v1.4s, v2.4s
> > >>        tbl     v0.16b, {v0.16b - v1.16b}, v3.16b
> > >>        str     q0, [x2, x3]
> > >>        add     x3, x3, 16
> > >>        cmp     x3, x4
> > >>        bne     .L3
> > >>
> > >> Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
> > >>
> > >> Thanks to George Steed for the idea.
> > >>
> > >> Ok for master?
> > >>
> > >> Thanks,
> > >> Tamar
> > >>
> > >> gcc/ChangeLog:
> > >>
> > >>        * match.pd: Add fneg/fadd rule.
> > >>
> > >> gcc/testsuite/ChangeLog:
> > >>
> > >>        * gcc.target/aarch64/simd/addsub_1.c: New test.
> > >>        * gcc.target/aarch64/sve/addsub_1.c: New test.
> > >>
> > >> --- inline copy of patch --
> > >> diff --git a/gcc/match.pd b/gcc/match.pd index
> > >>
> > 51b0a1b562409af535e53828a10c30b8a3e1ae2e..af1c98d4a2831f38258d6fc1bb
> > e
> > >> 811c8ee6c7c6e 100644
> > >> --- a/gcc/match.pd
> > >> +++ b/gcc/match.pd
> > >> @@ -7612,6 +7612,49 @@ and,
> > >>   (simplify (reduc (op @0 VECTOR_CST@1))
> > >>     (op (reduc:type @0) (reduc:type @1))))
> > >>
> > >> +/* Simplify vector floating point operations of alternating sub/add pairs
> > >> +   into using an fneg of a wider element type followed by a normal add.
> > >> +   under IEEE 754 the fneg of the wider type will negate every even entry
> > >> +   and when doing an add we get a sub of the even and add of every odd
> > >> +   elements.  */
> > >> +(simplify
> > >> + (vec_perm (plus:c @0 @1) (minus @0 @1) VECTOR_CST@2)  (if
> > >> +(!VECTOR_INTEGER_TYPE_P (type) && !BYTES_BIG_ENDIAN)
> > >> +  (with
> > >> +   {
> > >> +     /* Build a vector of integers from the tree mask.  */
> > >> +     vec_perm_builder builder;
> > >> +     if (!tree_to_vec_perm_builder (&builder, @2))
> > >> +       return NULL_TREE;
> > >> +
> > >> +     /* Create a vec_perm_indices for the integer vector.  */
> > >> +     poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (type);
> > >> +     vec_perm_indices sel (builder, 2, nelts);
> > >> +   }
> > >> +   (if (sel.series_p (0, 2, 0, 2))
> > >> +    (with
> > >> +     {
> > >> +       machine_mode vec_mode = TYPE_MODE (type);
> > >> +       auto elem_mode = GET_MODE_INNER (vec_mode);
> > >> +       auto nunits = exact_div (GET_MODE_NUNITS (vec_mode), 2);
> > >> +       tree stype;
> > >> +       switch (elem_mode)
> > >> +        {
> > >> +        case E_HFmode:
> > >> +          stype = float_type_node;
> > >> +          break;
> > >> +        case E_SFmode:
> > >> +          stype = double_type_node;
> > >> +          break;
> > >> +        default:
> > >> +          return NULL_TREE;
> > >> +        }
> > >> +       tree ntype = build_vector_type (stype, nunits);
> > >> +       if (!ntype)
> > >> +        return NULL_TREE;
> > >> +     }
> > >> +     (plus (view_convert:type (negate (view_convert:ntype @1)))
> > >> + @0))))))
> > >> +
> > >> (simplify
> > >>  (vec_perm @0 @1 VECTOR_CST@2)
> > >>  (with
> > >> diff --git a/gcc/testsuite/gcc.target/aarch64/simd/addsub_1.c
> > >> b/gcc/testsuite/gcc.target/aarch64/simd/addsub_1.c
> > >> new file mode 100644
> > >> index
> > >>
> > 0000000000000000000000000000000000000000..1fb91a34c421bbd2894faa0db
> > bf
> > >> 1b47ad43310c4
> > >> --- /dev/null
> > >> +++ b/gcc/testsuite/gcc.target/aarch64/simd/addsub_1.c
> > >> @@ -0,0 +1,56 @@
> > >> +/* { dg-do compile } */
> > >> +/* { dg-require-effective-target arm_v8_2a_fp16_neon_ok } */
> > >> +/* { dg-options "-Ofast" } */
> > >> +/* { dg-add-options arm_v8_2a_fp16_neon } */
> > >> +/* { dg-final { check-function-bodies "**" "" "" { target { le } } }
> > >> +} */
> > >> +
> > >> +#pragma GCC target "+nosve"
> > >> +
> > >> +/*
> > >> +** f1:
> > >> +** ...
> > >> +**     fneg    v[0-9]+.2d, v[0-9]+.2d
> > >> +**     fadd    v[0-9]+.4s, v[0-9]+.4s, v[0-9]+.4s
> > >> +** ...
> > >> +*/
> > >> +void f1 (float *restrict a, float *restrict b, float *res, int n) {
> > >> +   for (int i = 0; i < (n & -4); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >> +
> > >> +/*
> > >> +** d1:
> > >> +** ...
> > >> +**     fneg    v[0-9]+.4s, v[0-9]+.4s
> > >> +**     fadd    v[0-9]+.8h, v[0-9]+.8h, v[0-9]+.8h
> > >> +** ...
> > >> +*/
> > >> +void d1 (_Float16 *restrict a, _Float16 *restrict b, _Float16 *res,
> > >> +int n) {
> > >> +   for (int i = 0; i < (n & -8); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >> +
> > >> +/*
> > >> +** e1:
> > >> +** ...
> > >> +**     fadd    v[0-9]+.2d, v[0-9]+.2d, v[0-9]+.2d
> > >> +**     fsub    v[0-9]+.2d, v[0-9]+.2d, v[0-9]+.2d
> > >> +**     ins     v[0-9]+.d\[1\], v[0-9]+.d\[1\]
> > >> +** ...
> > >> +*/
> > >> +void e1 (double *restrict a, double *restrict b, double *res, int n)
> > >> +{
> > >> +   for (int i = 0; i < (n & -4); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/addsub_1.c
> > >> b/gcc/testsuite/gcc.target/aarch64/sve/addsub_1.c
> > >> new file mode 100644
> > >> index
> > >>
> > 0000000000000000000000000000000000000000..ea7f9d9db2c8c9a3efe5c7951a
> > 3
> > >> 14a29b7a7a922
> > >> --- /dev/null
> > >> +++ b/gcc/testsuite/gcc.target/aarch64/sve/addsub_1.c
> > >> @@ -0,0 +1,52 @@
> > >> +/* { dg-do compile } */
> > >> +/* { dg-options "-Ofast" } */
> > >> +/* { dg-final { check-function-bodies "**" "" "" { target { le } } }
> > >> +} */
> > >> +
> > >> +/*
> > >> +** f1:
> > >> +** ...
> > >> +**     fneg    z[0-9]+.d, p[0-9]+/m, z[0-9]+.d
> > >> +**     fadd    z[0-9]+.s, z[0-9]+.s, z[0-9]+.s
> > >> +** ...
> > >> +*/
> > >> +void f1 (float *restrict a, float *restrict b, float *res, int n) {
> > >> +   for (int i = 0; i < (n & -4); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >> +
> > >> +/*
> > >> +** d1:
> > >> +** ...
> > >> +**     fneg    z[0-9]+.s, p[0-9]+/m, z[0-9]+.s
> > >> +**     fadd    z[0-9]+.h, z[0-9]+.h, z[0-9]+.h
> > >> +** ...
> > >> +*/
> > >> +void d1 (_Float16 *restrict a, _Float16 *restrict b, _Float16 *res,
> > >> +int n) {
> > >> +   for (int i = 0; i < (n & -8); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >> +
> > >> +/*
> > >> +** e1:
> > >> +** ...
> > >> +**     fsub    z[0-9]+.d, z[0-9]+.d, z[0-9]+.d
> > >> +**     movprfx z[0-9]+.d, p[0-9]+/m, z[0-9]+.d
> > >> +**     fadd    z[0-9]+.d, p[0-9]+/m, z[0-9]+.d, z[0-9]+.d
> > >> +** ...
> > >> +*/
> > >> +void e1 (double *restrict a, double *restrict b, double *res, int n)
> > >> +{
> > >> +   for (int i = 0; i < (n & -4); i+=2)
> > >> +    {
> > >> +      res[i+0] = a[i+0] + b[i+0];
> > >> +      res[i+1] = a[i+1] - b[i+1];
> > >> +    }
> > >> +}
> > >>
> > >>
> > >>
> > >>
> > >> --
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstraße 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)