public inbox for gcc-cvs@sourceware.org
help / color / mirror / Atom feed
* [gcc r12-2728] Rewrite more vector loads to scalar loads
@ 2021-08-04 11:32 Richard Biener
0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2021-08-04 11:32 UTC (permalink / raw)
To: gcc-cvs
https://gcc.gnu.org/g:2724d1bba6b36451404811fba3244f8897717ef3
commit r12-2728-g2724d1bba6b36451404811fba3244f8897717ef3
Author: Richard Biener <rguenther@suse.de>
Date: Fri Jul 30 11:06:50 2021 +0200
Rewrite more vector loads to scalar loads
This teaches forwprop to rewrite more vector loads that are only
used in BIT_FIELD_REFs as scalar loads. This provides the
remaining uplift to SPEC CPU 2017 510.parest_r on Zen 2 which
has CPU gathers disabled.
In particular vector load + vec_unpack + bit-field-ref is turned
into (extending) scalar loads which avoids costly XMM/GPR
transitions. To not conflict with vector load + bit-field-ref
+ vector constructor matching to vector load + shuffle the
extended transform is only done after vector lowering.
2021-07-30 Richard Biener <rguenther@suse.de>
* tree-ssa-forwprop.c (pass_forwprop::execute): Split
out code to decompose vector loads ...
(optimize_vector_load): ... here. Generalize it to
handle intermediate widening and TARGET_MEM_REF loads
and apply it to loads with a supported vector mode as well.
Diff:
---
gcc/tree-ssa-forwprop.c | 244 ++++++++++++++++++++++++++++++++++++------------
1 file changed, 182 insertions(+), 62 deletions(-)
diff --git a/gcc/tree-ssa-forwprop.c b/gcc/tree-ssa-forwprop.c
index db3b18b275c..bd64b8e46bc 100644
--- a/gcc/tree-ssa-forwprop.c
+++ b/gcc/tree-ssa-forwprop.c
@@ -2757,6 +2757,182 @@ simplify_vector_constructor (gimple_stmt_iterator *gsi)
}
+/* Rewrite the vector load at *GSI to component-wise loads if the load
+ is only used in BIT_FIELD_REF extractions with eventual intermediate
+ widening. */
+
+static void
+optimize_vector_load (gimple_stmt_iterator *gsi)
+{
+ gimple *stmt = gsi_stmt (*gsi);
+ tree lhs = gimple_assign_lhs (stmt);
+ tree rhs = gimple_assign_rhs1 (stmt);
+
+ /* Gather BIT_FIELD_REFs to rewrite, looking through
+ VEC_UNPACK_{LO,HI}_EXPR. */
+ use_operand_p use_p;
+ imm_use_iterator iter;
+ bool rewrite = true;
+ auto_vec<gimple *, 8> bf_stmts;
+ auto_vec<tree, 8> worklist;
+ worklist.quick_push (lhs);
+ do
+ {
+ tree def = worklist.pop ();
+ unsigned HOST_WIDE_INT def_eltsize
+ = TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (TREE_TYPE (def))));
+ FOR_EACH_IMM_USE_FAST (use_p, iter, def)
+ {
+ gimple *use_stmt = USE_STMT (use_p);
+ if (is_gimple_debug (use_stmt))
+ continue;
+ if (!is_gimple_assign (use_stmt))
+ {
+ rewrite = false;
+ break;
+ }
+ enum tree_code use_code = gimple_assign_rhs_code (use_stmt);
+ tree use_rhs = gimple_assign_rhs1 (use_stmt);
+ if (use_code == BIT_FIELD_REF
+ && TREE_OPERAND (use_rhs, 0) == def
+ /* If its on the VEC_UNPACK_{HI,LO}_EXPR
+ def need to verify it is element aligned. */
+ && (def == lhs
+ || (known_eq (bit_field_size (use_rhs), def_eltsize)
+ && constant_multiple_p (bit_field_offset (use_rhs),
+ def_eltsize))))
+ {
+ bf_stmts.safe_push (use_stmt);
+ continue;
+ }
+ /* Walk through one level of VEC_UNPACK_{LO,HI}_EXPR. */
+ if (def == lhs
+ && (use_code == VEC_UNPACK_HI_EXPR
+ || use_code == VEC_UNPACK_LO_EXPR)
+ && use_rhs == lhs)
+ {
+ worklist.safe_push (gimple_assign_lhs (use_stmt));
+ continue;
+ }
+ rewrite = false;
+ break;
+ }
+ if (!rewrite)
+ break;
+ }
+ while (!worklist.is_empty ());
+
+ if (!rewrite)
+ {
+ gsi_next (gsi);
+ return;
+ }
+ /* We now have all ultimate uses of the load to rewrite in bf_stmts. */
+
+ /* Prepare the original ref to be wrapped in adjusted BIT_FIELD_REFs.
+ For TARGET_MEM_REFs we have to separate the LEA from the reference. */
+ tree load_rhs = rhs;
+ if (TREE_CODE (load_rhs) == TARGET_MEM_REF)
+ {
+ if (TREE_CODE (TREE_OPERAND (load_rhs, 0)) == ADDR_EXPR)
+ mark_addressable (TREE_OPERAND (TREE_OPERAND (load_rhs, 0), 0));
+ tree tem = make_ssa_name (TREE_TYPE (TREE_OPERAND (load_rhs, 0)));
+ gimple *new_stmt
+ = gimple_build_assign (tem, build1 (ADDR_EXPR, TREE_TYPE (tem),
+ unshare_expr (load_rhs)));
+ gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+ load_rhs = build2_loc (EXPR_LOCATION (load_rhs),
+ MEM_REF, TREE_TYPE (load_rhs), tem,
+ build_int_cst
+ (TREE_TYPE (TREE_OPERAND (load_rhs, 1)), 0));
+ }
+
+ /* Rewrite the BIT_FIELD_REFs to be actual loads, re-emitting them at
+ the place of the original load. */
+ for (gimple *use_stmt : bf_stmts)
+ {
+ tree bfr = gimple_assign_rhs1 (use_stmt);
+ tree new_rhs = unshare_expr (load_rhs);
+ if (TREE_OPERAND (bfr, 0) != lhs)
+ {
+ /* When the BIT_FIELD_REF is on the promoted vector we have to
+ adjust it and emit a conversion afterwards. */
+ gimple *def_stmt
+ = SSA_NAME_DEF_STMT (TREE_OPERAND (bfr, 0));
+ enum tree_code def_code
+ = gimple_assign_rhs_code (def_stmt);
+
+ /* The adjusted BIT_FIELD_REF is of the promotion source
+ vector size and at half of the offset... */
+ new_rhs = fold_build3 (BIT_FIELD_REF,
+ TREE_TYPE (TREE_TYPE (lhs)),
+ new_rhs,
+ TYPE_SIZE (TREE_TYPE (TREE_TYPE (lhs))),
+ size_binop (EXACT_DIV_EXPR,
+ TREE_OPERAND (bfr, 2),
+ bitsize_int (2)));
+ /* ... and offsetted by half of the vector if VEC_UNPACK_HI_EXPR. */
+ if (def_code == (!BYTES_BIG_ENDIAN
+ ? VEC_UNPACK_HI_EXPR : VEC_UNPACK_LO_EXPR))
+ TREE_OPERAND (new_rhs, 2)
+ = size_binop (PLUS_EXPR, TREE_OPERAND (new_rhs, 2),
+ size_binop (EXACT_DIV_EXPR,
+ TYPE_SIZE (TREE_TYPE (lhs)),
+ bitsize_int (2)));
+ tree tem = make_ssa_name (TREE_TYPE (TREE_TYPE (lhs)));
+ gimple *new_stmt = gimple_build_assign (tem, new_rhs);
+ location_t loc = gimple_location (use_stmt);
+ gimple_set_location (new_stmt, loc);
+ gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+ /* Perform scalar promotion. */
+ new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
+ NOP_EXPR, tem);
+ gimple_set_location (new_stmt, loc);
+ gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+ }
+ else
+ {
+ /* When the BIT_FIELD_REF is on the original load result
+ we can just wrap that. */
+ tree new_rhs = fold_build3 (BIT_FIELD_REF, TREE_TYPE (bfr),
+ unshare_expr (load_rhs),
+ TREE_OPERAND (bfr, 1),
+ TREE_OPERAND (bfr, 2));
+ gimple *new_stmt = gimple_build_assign (gimple_assign_lhs (use_stmt),
+ new_rhs);
+ location_t loc = gimple_location (use_stmt);
+ gimple_set_location (new_stmt, loc);
+ gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
+ }
+ gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
+ unlink_stmt_vdef (use_stmt);
+ gsi_remove (&gsi2, true);
+ }
+
+ /* Finally get rid of the intermediate stmts. */
+ gimple *use_stmt;
+ FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
+ {
+ if (is_gimple_debug (use_stmt))
+ {
+ if (gimple_debug_bind_p (use_stmt))
+ {
+ gimple_debug_bind_reset_value (use_stmt);
+ update_stmt (use_stmt);
+ }
+ continue;
+ }
+ gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
+ unlink_stmt_vdef (use_stmt);
+ release_defs (use_stmt);
+ gsi_remove (&gsi2, true);
+ }
+ /* And the original load. */
+ release_defs (stmt);
+ gsi_remove (gsi, true);
+}
+
+
/* Primitive "lattice" function for gimple_simplify. */
static tree
@@ -3007,71 +3183,15 @@ pass_forwprop::execute (function *fun)
gsi_next (&gsi);
}
else if (TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE
- && TYPE_MODE (TREE_TYPE (lhs)) == BLKmode
+ && (TYPE_MODE (TREE_TYPE (lhs)) == BLKmode
+ /* After vector lowering rewrite all loads, but
+ initially do not since this conflicts with
+ vector CONSTRUCTOR to shuffle optimization. */
+ || (fun->curr_properties & PROP_gimple_lvec))
&& gimple_assign_load_p (stmt)
&& !gimple_has_volatile_ops (stmt)
- && (TREE_CODE (gimple_assign_rhs1 (stmt))
- != TARGET_MEM_REF)
&& !stmt_can_throw_internal (cfun, stmt))
- {
- /* Rewrite loads used only in BIT_FIELD_REF extractions to
- component-wise loads. */
- use_operand_p use_p;
- imm_use_iterator iter;
- bool rewrite = true;
- FOR_EACH_IMM_USE_FAST (use_p, iter, lhs)
- {
- gimple *use_stmt = USE_STMT (use_p);
- if (is_gimple_debug (use_stmt))
- continue;
- if (!is_gimple_assign (use_stmt)
- || gimple_assign_rhs_code (use_stmt) != BIT_FIELD_REF
- || TREE_OPERAND (gimple_assign_rhs1 (use_stmt), 0) != lhs)
- {
- rewrite = false;
- break;
- }
- }
- if (rewrite)
- {
- gimple *use_stmt;
- FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
- {
- if (is_gimple_debug (use_stmt))
- {
- if (gimple_debug_bind_p (use_stmt))
- {
- gimple_debug_bind_reset_value (use_stmt);
- update_stmt (use_stmt);
- }
- continue;
- }
-
- tree bfr = gimple_assign_rhs1 (use_stmt);
- tree new_rhs = fold_build3 (BIT_FIELD_REF,
- TREE_TYPE (bfr),
- unshare_expr (rhs),
- TREE_OPERAND (bfr, 1),
- TREE_OPERAND (bfr, 2));
- gimple *new_stmt
- = gimple_build_assign (gimple_assign_lhs (use_stmt),
- new_rhs);
-
- location_t loc = gimple_location (use_stmt);
- gimple_set_location (new_stmt, loc);
- gimple_stmt_iterator gsi2 = gsi_for_stmt (use_stmt);
- unlink_stmt_vdef (use_stmt);
- gsi_remove (&gsi2, true);
-
- gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
- }
-
- release_defs (stmt);
- gsi_remove (&gsi, true);
- }
- else
- gsi_next (&gsi);
- }
+ optimize_vector_load (&gsi);
else if (code == COMPLEX_EXPR)
{
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2021-08-04 11:32 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-04 11:32 [gcc r12-2728] Rewrite more vector loads to scalar loads Richard Biener
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).