]> git.ipfire.org Git - thirdparty/gcc.git/commit
slp: remove unneeded permute calculation (PR99656)
authorTamar Christina <tamar.christina@arm.com>
Fri, 19 Mar 2021 14:29:36 +0000 (14:29 +0000)
committerTamar Christina <tamar.christina@arm.com>
Fri, 19 Mar 2021 14:29:36 +0000 (14:29 +0000)
commitc3a2bc6daaa2d278cb5f323e2df4b8c2af4198ac
tree4e45e808c1e9b8eac8fee0a55e2c273fb9d03a14
parent5e2eabe1eed1e53d39923517122d3c7de2013ad4
slp: remove unneeded permute calculation (PR99656)

The attach testcase ICEs because as you showed on the PR we have one child
which is an internal with a PERM of EVENEVEN and one with TOP.

The problem is while we can conceptually merge the permute itself into EVENEVEN,
merging the lanes don't really make sense.

That said, we no longer even require the merged lanes as we create the permutes
based on the KIND directly.

This patch just removes all of that code.

Unfortunately it still won't vectorize with the cost model enabled due to the
blend that's created combining the load and the external

note: node 0x51f2ce8 (max_nunits=1, refcnt=1)
note: op: VEC_PERM_EXPR
note:       { }
note:       lane permutation { 0[0] 1[1] }
note:       children 0x51f23e0 0x51f2578
note: node 0x51f23e0 (max_nunits=2, refcnt=1)
note: op template: _16 = REALPART_EXPR <*t1_9(D)>;
note:       stmt 0 _16 = REALPART_EXPR <*t1_9(D)>;
note:       stmt 1 _16 = REALPART_EXPR <*t1_9(D)>;
note:       load permutation { 0 0 }
note: node (external) 0x51f2578 (max_nunits=1, refcnt=1)
note:       { _18, _18 }

which costs the cost for the load-and-split and the cost of the external splat,
and the one for blending them while in reality it's just a scalar load and
insert.

The compiler (with the cost model disabled) generates

ldr     q1, [x19]
dup     v1.2d, v1.d[0]
ldr     d0, [x19, 8]
fneg    d0, d0
ins     v1.d[1], v0.d[0]

while really it should be

ldp     d1, d0, [x19]
fneg    d0, d0
ins     v1.d[1], v0.d[0]

but that's for another time.

gcc/ChangeLog:

PR tree-optimization/99656
* tree-vect-slp-patterns.c (linear_loads_p,
complex_add_pattern::matches, is_eq_or_top,
vect_validate_multiplication, complex_mul_pattern::matches,
complex_fms_pattern::matches): Remove complex_perm_kinds_t.
* tree-vectorizer.h: (complex_load_perm_t): Removed.
(slp_tree_to_load_perm_map_t): Use complex_perm_kinds_t instead of
complex_load_perm_t.

gcc/testsuite/ChangeLog:

PR tree-optimization/99656
* gfortran.dg/vect/pr99656.f90: New test.
gcc/testsuite/gfortran.dg/vect/pr99656.f90 [new file with mode: 0644]
gcc/tree-vect-slp-patterns.c
gcc/tree-vectorizer.h