Prathamesh Kulkarni <prathamesh.kulka...@linaro.org> writes:
> On Thu, 22 Aug 2019 at 16:44, Richard Biener <richard.guent...@gmail.com>
> wrote:
>>
>> On Wed, Aug 21, 2019 at 8:24 PM Prathamesh Kulkarni
>> <prathamesh.kulka...@linaro.org> wrote:
>> >
>> > On Thu, 15 Aug 2019 at 01:50, Richard Sandiford
>> > <richard.sandif...@arm.com> wrote:
>> > >
>> > > Richard Biener <richard.guent...@gmail.com> writes:
>> > > > On Wed, Aug 14, 2019 at 6:49 PM Richard Biener
>> > > > <richard.guent...@gmail.com> wrote:
>> > > >>
>> > > >> On Wed, Aug 14, 2019 at 5:06 PM Prathamesh Kulkarni
>> > > >> <prathamesh.kulka...@linaro.org> wrote:
>> > > >> >
>> > > >> > Hi,
>> > > >> > The attached patch tries to fix PR86753.
>> > > >> >
>> > > >> > For following test:
>> > > >> > void
>> > > >> > f1 (int *restrict x, int *restrict y, int *restrict z)
>> > > >> > {
>> > > >> > for (int i = 0; i < 100; ++i)
>> > > >> > x[i] = y[i] ? z[i] : 10;
>> > > >> > }
>> > > >> >
>> > > >> > vect dump shows:
>> > > >> > vect_cst__42 = { 0, ... };
>> > > >> > vect_cst__48 = { 0, ... };
>> > > >> >
>> > > >> > vect__4.7_41 = .MASK_LOAD (vectp_y.5_38, 4B, loop_mask_40);
>> > > >> > _4 = *_3;
>> > > >> > _5 = z_12(D) + _2;
>> > > >> > mask__35.8_43 = vect__4.7_41 != vect_cst__42;
>> > > >> > _35 = _4 != 0;
>> > > >> > vec_mask_and_46 = mask__35.8_43 & loop_mask_40;
>> > > >> > vect_iftmp.11_47 = .MASK_LOAD (vectp_z.9_44, 4B, vec_mask_and_46);
>> > > >> > iftmp.0_13 = 0;
>> > > >> > vect_iftmp.12_50 = VEC_COND_EXPR <vect__4.7_41 != vect_cst__48,
>> > > >> > vect_iftmp.11_47, vect_cst__49>;
>> > > >> >
>> > > >> > and following code-gen:
>> > > >> > L2:
>> > > >> > ld1w z0.s, p2/z, [x1, x3, lsl 2]
>> > > >> > cmpne p1.s, p3/z, z0.s, #0
>> > > >> > cmpne p0.s, p2/z, z0.s, #0
>> > > >> > ld1w z0.s, p0/z, [x2, x3, lsl 2]
>> > > >> > sel z0.s, p1, z0.s, z1.s
>> > > >> >
>> > > >> > We could reuse vec_mask_and_46 in vec_cond_expr since the conditions
>> > > >> > vect__4.7_41 != vect_cst__48 and vect__4.7_41 != vect_cst__42
>> > > >> > are equivalent, and vect_iftmp.11_47 depends on vect__4.7_41 !=
>> > > >> > vect_cst__48.
>> > > >> >
>> > > >> > I suppose in general for vec_cond_expr <C, T, E> if T comes from
>> > > >> > masked load,
>> > > >> > which is conditional on C, then we could reuse the mask used in
>> > > >> > load,
>> > > >> > in vec_cond_expr ?
>> > > >> >
>> > > >> > The patch maintains a hash_map cond_to_vec_mask
>> > > >> > from <cond, loop_mask -> vec_mask (with loop predicate applied).
>> > > >> > In prepare_load_store_mask, we record <cond, loop_mask> -> vec_mask
>> > > >> > & loop_mask,
>> > > >> > and in vectorizable_condition, we check if <cond, loop_mask> exists
>> > > >> > in
>> > > >> > cond_to_vec_mask
>> > > >> > and if found, the corresponding vec_mask is used as 1st operand of
>> > > >> > vec_cond_expr.
>> > > >> >
>> > > >> > <cond, loop_mask> is represented with cond_vmask_key, and the patch
>> > > >> > adds tree_cond_ops to represent condition operator and operands
>> > > >> > coming
>> > > >> > either from cond_expr
>> > > >> > or a gimple comparison stmt. If the stmt is not comparison, it
>> > > >> > returns
>> > > >> > <ne_expr, lhs, 0> and inserts that into cond_to_vec_mask.
>> > > >> >
>> > > >> > With patch, the redundant p1 is eliminated and sel uses p0 for
>> > > >> > above test.
>> > > >> >
>> > > >> > For following test:
>> > > >> > void
>> > > >> > f2 (int *restrict x, int *restrict y, int *restrict z, int fallback)
>> > > >> > {
>> > > >> > for (int i = 0; i < 100; ++i)
>> > > >> > x[i] = y[i] ? z[i] : fallback;
>> > > >> > }
>> > > >> >
>> > > >> > input to vectorizer has operands swapped in cond_expr:
>> > > >> > _36 = _4 != 0;
>> > > >> > iftmp.0_14 = .MASK_LOAD (_5, 32B, _36);
>> > > >> > iftmp.0_8 = _4 == 0 ? fallback_12(D) : iftmp.0_14;
>> > > >> >
>> > > >> > So we need to check for inverted condition in cond_to_vec_mask,
>> > > >> > and swap the operands.
>> > > >> > Does the patch look OK so far ?
>> > > >> >
>> > > >> > One major issue remaining with the patch is value numbering.
>> > > >> > Currently, it does value numbering for entire function using sccvn
>> > > >> > during start of vect pass, which is too expensive since we only need
>> > > >> > block based VN. I am looking into that.
>> > > >>
>> > > >> Why do you need it at all? We run VN on the if-converted loop bodies
>> > > >> btw.
>> > >
>> > > This was my suggestion, but with the idea being to do the numbering
>> > > per-statement as we vectorise. We'll then see pattern statements too.
>> > >
>> > > That's important because we use pattern statements to set the right
>> > > vector boolean type (e.g. vect_recog_mask_conversion_pattern).
>> > > So some of the masks we care about don't exist after if converison.
>> > >
>> > > > Also I can't trivially see the equality of the masks and probably so
>> > > > can't VN. Is it that we just don't bother to apply loop_mask to
>> > > > VEC_COND but there's no harm if we do?
>> > >
>> > > Yeah. The idea of the optimisation is to decide when it's more
>> > > profitable
>> > > to apply the loop mask, even though doing so isn't necessary. It would
>> > > be hard to do after vectorisation because the masks aren't equivalent.
>> > > We're relying on knowledge of how the vectoriser uses the result.
>> > Hi,
>> > Sorry for late response. This is an updated patch, that integrates
>> > block-based VN into vect pass.
>> > The patch
>> > (a) Exports visit_stmt (renamed to vn_visit_stmt), vn_bb_init to
>> > initialize VN state, and vn_bb_free to free it.
>> > (b) Calls vn_visit_stmt in vect_transform_stmt for value numbering
>> > stmts. We're only interested in obtaining
>> > value numbers, not eliminating redundancies.
>> > Does it look in the right direction ?
>>
>> It looks a bit odd to me. I'd have expected it to work by generating
>> the stmts as before in the vectorizer and then on the stmts we care
>> invoke vn_visit_stmt that does both value-numbering and elimination.
>> Alternatively you could ask the VN state to generate the stmt for
>> you via vn_nary_build_or_lookup () (certainly that needs a bit more
>> work). One complication might be availability if you don't value-number
>> all stmts in the block, but well. I'm not sure constraining to a single
>> block is necessary - I've thought of having a "CSE"ing gimple_build
>> for some time (add & CSE new stmts onto a sequence), so one
>> should keep this mode in mind when designing the one working on
>> an existing BB. Note as you write it it depends on visiting the
>> stmts in proper order - is that guaranteed when for example
>> vectorizing SLP?
> Hi,
> Indeed, I wrote the function with assumption that, stmts would be
> visited in proper order.
> This doesn't affect SLP currently, because call to vn_visit_stmt in
> vect_transform_stmt is
> conditional on cond_to_vec_mask, which is only allocated inside
> vect_transform_loop.
> But I agree we could make it more general.
> AFAIU, the idea of constraining VN to single block was to avoid using defs
> from
> non-dominating scalar stmts during outer-loop vectorization.
Maybe we could do the numbering in a separate walk immediately before
the transform phase instead.
> * fmla_2.c regression with patch:
> This happens because with patch, forwprop4 is able to convert all 3
> vec_cond_expr's
> to .cond_fma(), which results in 3 calls to fmla, regressing the
> test-case. If matching with
> inverted condition is disabled in patch in vectorizable_condition,
> then the old behavior gets preserved.
Ugh, yeah. This all stems from the fact that we don't get rid of
the first redundant store (before or after the patch). I think it's
just luck that we managed to CSE the FMLAs feeding the two stores.
(That wasn't what the test was added for; it was added for a
suboptimal .md choice.)
I think it'd be OK to XFAIL the current scan-assembler-times and add a
new one for 3 times instead of 2 (with a comment of course). I've filed
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91532 for the redundant
stores.
Thanks,
Richard
