OpenE2K/gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

tree-optimization/92645 - improve SLP with existing vectors

Browse Source 
This improves SLP discovery in the face of existing vectors allowing
punning of the vector shape (or even punning from an integer type).
For punning from integer types this does not yet handle lane zero
extraction being represented as conversion rather than BIT_FIELD_REF.

2021-01-13  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/92645
	* tree-vect-slp.c (vect_build_slp_tree_1): Relax supported
	BIT_FIELD_REF argument.
	(vect_build_slp_tree_2): Record the desired vector type
	on the external vector def.
	(vectorizable_slp_permutation): Handle required punning
	of existing vector defs.

	* gcc.target/i386/pr92645-6.c: New testcase.
This commit is contained in:
Richard Biener 2021-01-13 12:40:01 +01:00
parent 5ab67cdee6
commit 3ddc18251a
2 changed files with 63 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
gcc/testsuite/gcc.target/i386/pr92645-6.c Normal file
View File

@ -0,0 +1,34 @@
/* { dg-do compile } */
/* { dg-require-effective-target lp64 } */
/* { dg-options "-O3 -msse2" } */
typedef long v2di __attribute__((vector_size(16)));
typedef int v4si __attribute__((vector_size(16)));
void foo (v4si *p, v2di *q)
{
union { v2di a; v4si b; } u;
u.a = *q;
(*p)[0] = u.b[0];
(*p)[1] = u.b[2];
(*p)[2] = u.b[1];
(*p)[3] = u.b[3];
}
void bar (v4si *p, __int128_t *q)
{
union { __int128_t a; v4si b; } u;
u.a = *q;
(*p)[0] = u.b[1];
(*p)[1] = u.b[2];
(*p)[2] = u.b[1];
(*p)[3] = u.b[3];
}
/* Both functions should end up with sth like
[v]pshufd $val, (%esi), %xmm0
[v]movdqa %xmm0, (%edi)
ret
recognized by SLP vectorization involving an existing "vector". */
/* { dg-final { scan-assembler-not "punpck" } } */
/* { dg-final { scan-assembler-times "pshufd" 2 } } */

31
gcc/tree-vect-slp.c
View File

@ -1105,7 +1105,8 @@ vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
if (!is_a <bb_vec_info> (vinfo)
|| TREE_CODE (vec) != SSA_NAME
|| !types_compatible_p (vectype, TREE_TYPE (vec)))
|| !operand_equal_p (TYPE_SIZE (vectype),
TYPE_SIZE (TREE_TYPE (vec))))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
@ -1642,7 +1643,11 @@ vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
lperm.safe_push (std::make_pair (0, (unsigned)lane));
}
slp_tree vnode = vect_create_new_slp_node (vNULL);
SLP_TREE_VECTYPE (vnode) = TREE_TYPE (vec);
/* ??? We record vectype here but we hide eventually necessary
punning and instead rely on code generation to materialize
VIEW_CONVERT_EXPRs as necessary. We instead should make
this explicit somehow. */
SLP_TREE_VECTYPE (vnode) = vectype;
SLP_TREE_VEC_DEFS (vnode).safe_push (vec);
/* We are always building a permutation node even if it is an identity
permute to shield the rest of the vectorizer from the odd node
@ -5671,6 +5676,18 @@ vectorizable_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
slp_tree first_node = SLP_TREE_CHILDREN (node)[first_vec.first];
tree first_def
= vect_get_slp_vect_def (first_node, first_vec.second);
/* ??? We SLP match existing vector element extracts but
allow punning which we need to re-instantiate at uses
but have no good way of explicitely representing. */
if (!types_compatible_p (TREE_TYPE (first_def), vectype))
{
gassign *conv_stmt;
conv_stmt = gimple_build_assign (make_ssa_name (vectype),
build1 (VIEW_CONVERT_EXPR,
vectype, first_def));
vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
first_def = gimple_assign_lhs (conv_stmt);
}
gassign *perm_stmt;
tree perm_dest = make_ssa_name (vectype);
if (!identity_p)
@ -5679,6 +5696,16 @@ vectorizable_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
= SLP_TREE_CHILDREN (node)[second_vec.first];
tree second_def
= vect_get_slp_vect_def (second_node, second_vec.second);
if (!types_compatible_p (TREE_TYPE (second_def), vectype))
{
gassign *conv_stmt;
conv_stmt = gimple_build_assign (make_ssa_name (vectype),
build1
(VIEW_CONVERT_EXPR,
vectype, second_def));
vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
second_def = gimple_assign_lhs (conv_stmt);
}
tree mask_vec = vect_gen_perm_mask_checked (vectype, indices);
perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
first_def, second_def,