On Fri, 19 Feb 2021, Tamar Christina wrote:
> Hi Richi,
>
> The attached testcase ICEs due to a couple of issues.
> In the testcase you have two SLP instances that share the majority of their
> definition with each other. One tree defines a COMPLEX_MUL sequence and the
> other tree a COMPLEX_FMA.
>
> The ice happens because:
>
> 1. the refcounts are wrong, in particular the FMA case doesn't correctly count
> the references for the COMPLEX_MUL that it consumes.
>
> 2. when the FMA is created it incorrectly assumes it can just tear apart the
> MUL
> node that it's consuming. This is wrong and should only be done when there is
> no more uses of the node, in which case the vector only pattern is no longer
> relevant.
>
> To fix the last part the SLP only pattern reset code was moved into
> vect_free_slp_tree which results in cleaner code. I also think it does belong
> there since that function knows when there are no more uses of the node and so
> the pattern should be unmarked, so when the the vectorizer is inspecting the
> BB
> it doesn't find the now invalid vector only patterns.
Note that if you rely on the pattern being reset when it's no longer used
then this won't work in case we end up in vect_slp_convert_to_external
because that does
/* Don't remove and free the child nodes here, since they could be
referenced by other structures. The analysis and scheduling phases
(need to) ignore child nodes of anything that isn't
vect_internal_def. */
unsigned int group_size = SLP_TREE_LANES (node);
SLP_TREE_DEF_TYPE (node) = vect_external_def;
SLP_TREE_SCALAR_OPS (node).safe_grow (group_size, true);
SLP_TREE_LOAD_PERMUTATION (node).release ();
FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
{
tree lhs = gimple_get_lhs (vect_orig_stmt (stmt_info)->stmt);
where I don't remember the reason for not freeing the children
(it's Richards code, so you'd have to ask him). OTOH the above only
triggers for non-loop vectorization where the pattern reset isn't
needed?
> This has the obvious problem in that, eventually after analysis is done, the
> entire SLP tree is dissolved before codegen. Where we get into trouble as we
> have now dissolved the patterns too...
Huh? Codegen needs the SLP nodes so they most definitely survive until
after vectorization has finished. So did you run into actual problems
with this approach?
> My initial thought was to add a parameter to vect_free_slp_tree, but I know
> you
> wouldn't like that. So I am sending this patch up as an RFC.
>
> PS. This testcase actually shows that the codegen we get in these cases is not
> optimal. Currently this won't vectorize as the compiler thinks the vector
> version is too expensive.
>
> My guess here is because the patterns now unshare the tree and it's likely
> costing the setup for the vector code twice?
I can see it figures there are LIVE lanes out of the SLP tree, likely
because the way vect_bb_slp_mark_live_stmts works out the stmt to
vectorize doesn't work for SLP patterns. It sees
t.ii:19:10: note: node 0x3b87dd0 (max_nunits=2, refcnt=1)
t.ii:19:10: note: op template: MEM <float> [(struct a *)_10] = _24;
t.ii:19:10: note: stmt 0 MEM <float> [(struct a *)_10] = _24;
t.ii:19:10: note: stmt 1 MEM <float> [(struct a *)_10 + 4B] = _25;
t.ii:19:10: note: children 0x3b87e58
t.ii:19:10: note: node 0x3b87e58 (max_nunits=2, refcnt=1)
t.ii:19:10: note: op template: slp_patt_34 = .COMPLEX_FMA (_24, _24, _24);
t.ii:19:10: note: stmt 0 _24 = l$b_11 + _20;
t.ii:19:10: note: stmt 1 _25 = l$c_12 + _23;
t.ii:19:10: note: children 0x3b87ee0 0x3b88100 0x3b88100
t.ii:19:10: note: node 0x3b87ee0 (max_nunits=2, refcnt=1)
t.ii:19:10: note: op template: l$b_11 = MEM[(const struct a &)_10].b;
t.ii:19:10: note: stmt 0 l$b_11 = MEM[(const struct a &)_10].b;
t.ii:19:10: note: stmt 1 l$c_12 = MEM[(const struct a &)_10].c;
t.ii:19:10: note: node 0x3b88100 (max_nunits=2, refcnt=4)
t.ii:19:10: note: op template: k$b_4 = MEM[(const struct a &)_3].b;
t.ii:19:10: note: stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
t.ii:19:10: note: stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
but the scalar stmts for the .COMPLEX_FMA pattern have uses that
are not marked as PURE_SLP (_20 and _23).
There's the other issue that we have
t.ii:19:10: note: op template: slp_patt_33 = .COMPLEX_MUL (_20, _20);
t.ii:19:10: note: stmt 0 _20 = _18 - _19;
t.ii:19:10: note: stmt 1 _23 = _21 + _22;
t.ii:19:10: note: children 0x3b88180 0x3b88538
t.ii:19:10: note: node 0x3b88538 (max_nunits=1, refcnt=1)
t.ii:19:10: note: op: VEC_PERM_EXPR
t.ii:19:10: note: { }
t.ii:19:10: note: lane permutation { 0[0] 1[1] }
t.ii:19:10: note: children 0x3b880f8 0x3b88290
t.ii:19:10: note: node 0x3b880f8 (max_nunits=2, refcnt=1)
t.ii:19:10: note: op template: k$b_4 = MEM[(const struct a &)_3].b;
t.ii:19:10: note: stmt 0 k$b_4 = MEM[(const struct a &)_3].b;
t.ii:19:10: note: stmt 1 k$b_4 = MEM[(const struct a &)_3].b;
t.ii:19:10: note: load permutation { 0 0 }
t.ii:19:10: note: node 0x3b88290 (max_nunits=2, refcnt=1)
t.ii:19:10: note: op template: k$c_5 = MEM[(const struct a &)_3].c;
t.ii:19:10: note: stmt 0 k$c_5 = MEM[(const struct a &)_3].c;
t.ii:19:10: note: stmt 1 k$c_5 = MEM[(const struct a &)_3].c;
t.ii:19:10: note: load permutation { 1 1 }
which is a load that wasn't converted somehow. Note I see
t.ii:19:10: note: converting stmts on permute node 0x3b88538
t.ii:19:10: note: re-using SLP tree 0x3b88180
t.ii:19:10: note: Pattern matched SLP tree
but eventually that didn't propagate to the use in the .COMPLEX_MUL
node.
So in the end your patch looks like a reasonable fix if the commented
lines are uncommented and that doesn't cause any problems.
Thanks,
Richard.
> Even with the shared code (without patterns, so same as GCC 10, or turning off
> the patterns) it won't vectorize it.
>
> The scalar code is
>
> mov w0, 0
> ldr x4, [x1, #:lo12:.LANCHOR0]
> ldrsw x2, [x3, 20]
> ldr x1, [x3, 8]
> lsl x2, x2, 3
> ldrsw x3, [x3, 16]
> ldp s3, s2, [x4]
> add x5, x1, x2
> ldr s0, [x1, x2]
> lsl x3, x3, 3
> add x4, x1, x3
> fmul s1, s2, s2
> fnmsub s1, s3, s3, s1
> fmul s0, s2, s0
> fmadd s0, s3, s2, s0
> ldr s3, [x1, x3]
> ldr s2, [x4, 4]
> fadd s3, s3, s1
> fadd s2, s0, s2
> str s3, [x1, x3]
> str s2, [x4, 4]
> str s1, [x1, x2]
> str s0, [x5, 4]
> ret
>
> and turning off the cost model we get
>
> movi v1.2s, 0
> mov w0, 0
> ldr x4, [x1, #:lo12:.LANCHOR0]
> ldrsw x3, [x2, 16]
> ldr x1, [x2, 8]
> ldrsw x2, [x2, 20]
> ldr d0, [x4]
> fcmla v1.2s, v0.2s, v0.2s, #0
> ldr d2, [x1, x3, lsl 3]
> fcmla v2.2s, v0.2s, v0.2s, #0
> fcmla v2.2s, v0.2s, v0.2s, #90
> str d2, [x1, x3, lsl 3]
> fcmla v1.2s, v0.2s, v0.2s, #90
> str d1, [x1, x2, lsl 3]
>
> however, if the pattern matcher doesn't create the FMA node because it would
> unshare the tree (which I think is a general heuristic that would work out to
> better code wouldn't it?) then we would get (with the cost model enabled even)
>
> movi v0.2s, 0
> mov w0, 0
> ldr x4, [x1, #:lo12:.LANCHOR0]
> ldrsw x3, [x2, 16]
> ldr x1, [x2, 8]
> ldr d1, [x4]
> fcmla v0.2s, v1.2s, v1.2s, #0
> fcmla v0.2s, v1.2s, v1.2s, #90
> ldrsw x2, [x2, 20]
> ldr d1, [x1, x3, lsl 3]
> fadd v1.2s, v1.2s, v0.2s
> str d1, [x1, x3, lsl 3]
> str d0, [x1, x2, lsl 3]
> ret
>
> Which is the most optimal form. So I think this should perhaps be handled in
> GCC 12 if there's a way to detect when you're going to unshare a sub-tree.
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> PR tree-optimization/99149
> * tree-vect-slp-patterns.c (vect_detect_pair_op): Don't recreate the
> buffer.
> (vect_slp_reset_pattern): Remove.
> (complex_fma_pattern::matches): Remove call to vect_slp_reset_pattern.
> (complex_mul_pattern::build, complex_fma_pattern::build,
> complex_fms_pattern::build): Fix ref counts.
> * tree-vect-slp.c (vect_free_slp_tree): Undo SLP only pattern relevancy
> when node is being deleted.
> * tree-vectorizer.c (vec_info::new_stmt_vec_info): Initialize value.
>
> gcc/testsuite/ChangeLog:
>
> PR tree-optimization/99149
> * gcc.dg/vect/pr99149.C: New test.
>
> --- inline copy of patch --
> diff --git a/gcc/testsuite/gcc.dg/vect/pr99149.C
> b/gcc/testsuite/gcc.dg/vect/pr99149.C
> new file mode 100755
> index
> 0000000000000000000000000000000000000000..b12fe17e4ded148ce2bf67486e425dd65461a148
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/pr99149.C
> @@ -0,0 +1,25 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-w -O3 -march=armv8.3-a" { target { aarch64*-*-*
> } } } */
> +
> +class a {
> + float b;
> + float c;
> +
> +public:
> + a(float d, float e) : b(d), c(e) {}
> + a operator+(a d) { return a(b + d.b, c + d.c); }
> + a operator*(a d) { return a(b * b - c * c, b * c + c * d.b); }
> +};
> +int f, g;
> +class {
> + a *h;
> + a *i;
> +
> +public:
> + void j() {
> + a k = h[0], l = i[g], m = k * i[f];
> + i[g] = l + m;
> + i[f] = m;
> + }
> +} n;
> +main() { n.j(); }
> diff --git a/gcc/tree-vect-slp-patterns.c b/gcc/tree-vect-slp-patterns.c
> index
> f0817da9f622d22e3df2e30410d1cf610b4ffa1d..1e2769662a54229ab8e24390f97dfe206f17ab57
> 100644
> --- a/gcc/tree-vect-slp-patterns.c
> +++ b/gcc/tree-vect-slp-patterns.c
> @@ -407,9 +407,8 @@ vect_detect_pair_op (slp_tree node1, slp_tree node2,
> lane_permutation_t &lanes,
>
> if (result != CMPLX_NONE && ops != NULL)
> {
> - ops->create (2);
> - ops->quick_push (node1);
> - ops->quick_push (node2);
> + ops->safe_push (node1);
> + ops->safe_push (node2);
> }
> return result;
> }
> @@ -1090,15 +1089,17 @@ complex_mul_pattern::build (vec_info *vinfo)
> {
> slp_tree node;
> unsigned i;
> + slp_tree newnode
> + = vect_build_combine_node (this->m_ops[0], this->m_ops[1],
> *this->m_node);
> + SLP_TREE_REF_COUNT (this->m_ops[2])++;
> +
> FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (*this->m_node), i, node)
> vect_free_slp_tree (node);
>
> /* First re-arrange the children. */
> SLP_TREE_CHILDREN (*this->m_node).reserve_exact (2);
> SLP_TREE_CHILDREN (*this->m_node)[0] = this->m_ops[2];
> - SLP_TREE_CHILDREN (*this->m_node)[1] =
> - vect_build_combine_node (this->m_ops[0], this->m_ops[1], *this->m_node);
> - SLP_TREE_REF_COUNT (this->m_ops[2])++;
> + SLP_TREE_CHILDREN (*this->m_node)[1] = newnode;
>
> /* And then rewrite the node itself. */
> complex_pattern::build (vinfo);
> @@ -1133,18 +1134,6 @@ class complex_fma_pattern : public complex_pattern
> }
> };
>
> -/* Helper function to "reset" a previously matched node and undo the changes
> - made enough so that the node is treated as an irrelevant node. */
> -
> -static inline void
> -vect_slp_reset_pattern (slp_tree node)
> -{
> - stmt_vec_info stmt_info = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (node));
> - STMT_VINFO_IN_PATTERN_P (stmt_info) = false;
> - STMT_SLP_TYPE (stmt_info) = pure_slp;
> - SLP_TREE_REPRESENTATIVE (node) = stmt_info;
> -}
> -
> /* Pattern matcher for trying to match complex multiply and accumulate
> and multiply and subtract patterns in SLP tree.
> If the operation matches then IFN is set to the operation it matched and
> @@ -1208,15 +1197,6 @@ complex_fma_pattern::matches (complex_operation_t op,
> if (!vect_pattern_validate_optab (ifn, vnode))
> return IFN_LAST;
>
> - /* FMA matched ADD + CMUL. During the matching of CMUL the
> - stmt that starts the pattern is marked as being in a pattern,
> - namely the CMUL. When replacing this with a CFMA we have to
> - unmark this statement as being in a pattern. This is because
> - vect_mark_pattern_stmts will only mark the current stmt as being
> - in a pattern. Later on when the scalar stmts are examined the
> - old statement which is supposed to be irrelevant will point to
> - CMUL unless we undo the pattern relationship here. */
> - vect_slp_reset_pattern (node);
> ops->truncate (0);
> ops->create (3);
>
> @@ -1259,10 +1239,17 @@ complex_fma_pattern::recognize
> (slp_tree_to_load_perm_map_t *perm_cache,
> void
> complex_fma_pattern::build (vec_info *vinfo)
> {
> + slp_tree node = SLP_TREE_CHILDREN (*this->m_node)[1];
> +
> SLP_TREE_CHILDREN (*this->m_node).release ();
> SLP_TREE_CHILDREN (*this->m_node).create (3);
> SLP_TREE_CHILDREN (*this->m_node).safe_splice (this->m_ops);
>
> + SLP_TREE_REF_COUNT (this->m_ops[1])++;
> + SLP_TREE_REF_COUNT (this->m_ops[2])++;
> +
> + vect_free_slp_tree (node);
> +
> complex_pattern::build (vinfo);
> }
>
> @@ -1427,6 +1414,11 @@ complex_fms_pattern::build (vec_info *vinfo)
> {
> slp_tree node;
> unsigned i;
> + slp_tree newnode =
> + vect_build_combine_node (this->m_ops[2], this->m_ops[3], *this->m_node);
> + SLP_TREE_REF_COUNT (this->m_ops[0])++;
> + SLP_TREE_REF_COUNT (this->m_ops[1])++;
> +
> FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (*this->m_node), i, node)
> vect_free_slp_tree (node);
>
> @@ -1436,10 +1428,7 @@ complex_fms_pattern::build (vec_info *vinfo)
> /* First re-arrange the children. */
> SLP_TREE_CHILDREN (*this->m_node).quick_push (this->m_ops[0]);
> SLP_TREE_CHILDREN (*this->m_node).quick_push (this->m_ops[1]);
> - SLP_TREE_CHILDREN (*this->m_node).quick_push (
> - vect_build_combine_node (this->m_ops[2], this->m_ops[3], *this->m_node));
> - SLP_TREE_REF_COUNT (this->m_ops[0])++;
> - SLP_TREE_REF_COUNT (this->m_ops[1])++;
> + SLP_TREE_CHILDREN (*this->m_node).quick_push (newnode);
>
> /* And then rewrite the node itself. */
> complex_pattern::build (vinfo);
> diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
> index
> ea8a97b01c6371791ac66de3e1dabfedee69cb67..65c2ff867ab41ea70367087dc26fb6eea1375ffb
> 100644
> --- a/gcc/tree-vect-slp.c
> +++ b/gcc/tree-vect-slp.c
> @@ -146,6 +146,16 @@ vect_free_slp_tree (slp_tree node)
> if (child)
> vect_free_slp_tree (child);
>
> + /* If the node defines any SLP only patterns then those patterns are no
> + longer valid and should be removed. */
> + stmt_vec_info rep_stmt_info = SLP_TREE_REPRESENTATIVE (node);
> + if (rep_stmt_info && STMT_VINFO_SLP_VECT_ONLY_PATTERN (rep_stmt_info))
> + {
> + stmt_vec_info stmt_info = vect_orig_stmt (rep_stmt_info);
> + //STMT_VINFO_IN_PATTERN_P (stmt_info) = false;
> + //STMT_SLP_TYPE (stmt_info) = STMT_SLP_TYPE (rep_stmt_info);
> + }
> +
> delete node;
> }
>
> diff --git a/gcc/tree-vectorizer.c b/gcc/tree-vectorizer.c
> index
> 5b45df3a4e00266b7530eb4da6985f0d940cb05b..63ba594f2276850a00fc372072d98326891f19e6
> 100644
> --- a/gcc/tree-vectorizer.c
> +++ b/gcc/tree-vectorizer.c
> @@ -695,6 +695,7 @@ vec_info::new_stmt_vec_info (gimple *stmt)
> STMT_VINFO_REDUC_FN (res) = IFN_LAST;
> STMT_VINFO_REDUC_IDX (res) = -1;
> STMT_VINFO_SLP_VECT_ONLY (res) = false;
> + STMT_VINFO_SLP_VECT_ONLY_PATTERN (res) = false;
> STMT_VINFO_VEC_STMTS (res) = vNULL;
>
> if (is_a <loop_vec_info> (this)
>
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)
