https://gcc.gnu.org/g:497e567c806af40f3ec3b135fae84de8d8b23818
commit 497e567c806af40f3ec3b135fae84de8d8b23818
Author: Alexandre Oliva <ol...@gnu.org>
Date: Fri Nov 15 23:18:20 2024 -0300
drop expensive mergeable tests in favor of gimple_vuse compares
Diff:
---
gcc/gimple-fold.cc | 179 +++++++++-------------------------
gcc/testsuite/gcc.dg/field-merge-10.c | 36 +++++++
gcc/tree-ssa-ifcombine.cc | 3 +-
3 files changed, 86 insertions(+), 132 deletions(-)
diff --git a/gcc/gimple-fold.cc b/gcc/gimple-fold.cc
index b7b4558ee2b4..6cb03ee09d19 100644
--- a/gcc/gimple-fold.cc
+++ b/gcc/gimple-fold.cc
@@ -7721,7 +7721,7 @@ build_split_load (tree /* out */ ln_arg[2],
tree type = lang_hooks.types.type_for_size (bitsiz[i], 1);
bitpos[i] = bit_pos;
ln_arg[i] = make_bit_field_load (loc, inner, orig_inner,
- type, bitsiz[i],
+ type, bitsiz[i],
bit_pos, 1, reversep, point[i]);
bit_pos += bitsiz[i];
}
@@ -7781,110 +7781,58 @@ reuse_split_load (tree /* in[0] out[1] */ ln_arg[2],
}
}
-/* Return nonzero if LSTMT and RSTMT, assumed to load from the same
- words, can be safely merged into a single load. Return zero if
- there are intervening stores, or if neither stmt dominates the
- other. If they are mergeable, return -1 if the merged load should
- be inserted before LSTMT to dominate both, otherwise return +1. */
-
-static int
-mergeable_loads_p (gimple *lstmt, gimple *rstmt)
-{
- gimple *stmts[2] = { lstmt, rstmt };
- int ret;
-
- if (stmt_dominates_stmt_p (lstmt, rstmt))
- {
- ret = -1;
- stmts[0] = lstmt;
- stmts[1] = rstmt;
- }
- else if (stmt_dominates_stmt_p (rstmt, lstmt))
- {
- ret = 1;
- stmts[1] = lstmt;
- stmts[0] = rstmt;
- }
- else
- return 0;
-
- basic_block bbs[2] = { gimple_bb (stmts[0]), gimple_bb (stmts[1]) };
- if (bbs[0] == bbs[1])
- {
- for (gimple_stmt_iterator gsi = gsi_for_stmt (stmts[0]);
- gsi_stmt (gsi) != stmts[1]; gsi_next (&gsi))
- if (gimple_vdef (gsi_stmt (gsi)))
- return 0;
- return ret;
- }
-
- for (gimple_stmt_iterator gsi = gsi_for_stmt (stmts[0]);
- !gsi_end_p (gsi); gsi_next (&gsi))
- if (gimple_vdef (gsi_stmt (gsi)))
- return 0;
+/* Find ways of folding logical expressions of LHS and RHS:
- for (gphi_iterator gpi = gsi_start_phis (bbs[1]);
- !gsi_end_p (gpi); gsi_next (&gpi))
- if (gimple_vdef (gsi_stmt (gpi)))
- return 0;
+ Try to merge two comparisons to nearby fields.
- for (gimple_stmt_iterator gsi = gsi_for_stmt (stmts[1]);
- !gsi_end_p (gsi); gsi_prev (&gsi))
- if (gimple_vdef (gsi_stmt (gsi)))
- return 0;
-
- for (basic_block bb = bbs[1]; bb != bbs[0]; )
- {
- /* ??? For now, stop if any visited block has more than one
- predecessor. */
- if (!single_pred_p (bb))
- return 0;
+ For example, if we have p->a == 2 && p->b == 4 and we can load both A and B
+ at once, we can do this with a comparison against the object ANDed with the
+ a mask.
- bb = single_pred (bb);
+ If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
+ operations to do this with one comparison, loading both fields from P at
+ once, and likewise from Q.
- for (gphi_iterator gpi = gsi_start_phis (bbs[1]);
- !gsi_end_p (gpi); gsi_next (&gpi))
- if (gimple_vdef (gsi_stmt (gpi)))
- return 0;
+ Herein, loading at once means loading from within the same alignment
+ boundary for the enclosing object. If (packed) fields cross such alignment
+ boundaries, we may still recombine the compares, so that loads do not cross
+ the boundaries.
- for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
- !gsi_end_p (gsi); gsi_next (&gsi))
- if (gimple_vdef (gsi_stmt (gsi)))
- return 0;
- }
+ CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
+ TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
- return ret;
-}
+ TRUTH_TYPE is the type of the logical operand.
-/* Find ways of folding logical expressions of LHS and RHS:
- Try to merge two comparisons to the same innermost item.
- Look for range tests like "ch >= '0' && ch <= '9'".
- Look for combinations of simple terms on machines with expensive branches
- and evaluate the RHS unconditionally.
+ LHS is denoted as LL_ARG LCODE LR_ARG.
- For example, if we have p->a == 2 && p->b == 4 and we can make an
- object large enough to span both A and B, we can do this with a comparison
- against the object ANDed with the a mask.
+ RHS is denoted as RL_ARG RCODE RR_ARG.
- If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
- operations to do this with one comparison.
+ LHS is assumed to dominate RHS.
- We check for both normal comparisons and the BIT_AND_EXPRs made this by
- function and the one above.
+ Combined loads are inserted next to preexisting loads, once we determine
+ that the combination is viable, and the combined condition references new
+ SSA_NAMEs that hold the loaded values. Since the original loads are
+ verified to have the same gimple_vuse, the insertion point doesn't matter
+ for correctness. ??? The loads may be a lot earlier than the compares, and
+ it's conceivable that one or two loads for RHS appear before those for LHS.
+ It could be advantageous to try to place the loads optimally, taking
+ advantage of knowing whether RHS is accessed before LHS, or that both are
+ accessed before both compares, but we don't do that (yet?).
- CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
- TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
+ SEPARATEP should be NULL if the combined condition must be returned as a
+ single expression, even if it is a compound condition. This must only be
+ done if LHS and RHS are adjacent, without intervening conditions, and the
+ combined condition is to replace RHS, while LHS is dropped altogether.
- TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
- two operands.
+ Otherwise, SEPARATEP must be a non-NULL pointer to a NULL_TREE, that may be
+ replaced by a part of the compound condition that could replace RHS, while
+ the returned expression replaces LHS.
- SEPARATEP should be NULL if LHS and RHS are adjacent in
- CODE-chained compares, and a non-NULL pointer to NULL_TREE
- otherwise. If the "words" accessed by RHS are already accessed by
- LHS, this won't matter, but if RHS accesses "words" that LHS
- doesn't, then *SEPARATEP will be set to the compares that should
- take RHS's place. By "words" we mean contiguous bits that do not
- cross a an TYPE_ALIGN boundary of the accessed object's type.
+ If the "words" accessed by RHS are already accessed by LHS, this won't
+ matter, but if RHS accesses "words" that LHS doesn't, then *SEPARATEP will
+ be set to the compares that should take RHS's place. By "words" we mean
+ contiguous bits that do not cross a an TYPE_ALIGN boundary of the accessed
+ object's type.
We return the simplified tree or 0 if no optimization is possible. */
@@ -7908,7 +7856,6 @@ fold_truth_andor_maybe_separate (location_t loc,
enum tree_code wanted_code;
tree ll_inner, lr_inner, rl_inner, rr_inner;
gimple *ll_load, *lr_load, *rl_load, *rr_load;
- int l_mergeable = 0, r_mergeable = 0;
HOST_WIDE_INT ll_bitsize, ll_bitpos, lr_bitsize, lr_bitpos;
HOST_WIDE_INT rl_bitsize, rl_bitpos, rr_bitsize, rr_bitpos;
HOST_WIDE_INT xll_bitpos, xlr_bitpos, xrl_bitpos, xrr_bitpos;
@@ -8015,7 +7962,7 @@ fold_truth_andor_maybe_separate (location_t loc,
|| ll_inner == 0 || rl_inner == 0
|| ! operand_equal_p (ll_inner, rl_inner, 0)
|| (ll_load && rl_load
- && ! (l_mergeable = mergeable_loads_p (ll_load, rl_load))))
+ && gimple_vuse (ll_load) != gimple_vuse (rl_load)))
return 0;
if (TREE_CODE (lr_arg) == INTEGER_CST
@@ -8028,7 +7975,7 @@ fold_truth_andor_maybe_separate (location_t loc,
|| lr_inner == 0 || rr_inner == 0
|| ! operand_equal_p (lr_inner, rr_inner, 0)
|| (lr_load && rr_load
- && ! (r_mergeable = mergeable_loads_p (lr_load, rr_load))))
+ && gimple_vuse (lr_load) != gimple_vuse (rr_load)))
return 0;
if (lsignbit)
@@ -8340,12 +8287,7 @@ fold_truth_andor_maybe_separate (location_t loc,
ld_arg[1][0] = make_bit_field_load (loc, lr_inner, lr_arg,
rntype, rnbitsize, rnbitpos,
lr_unsignedp || rr_unsignedp,
- lr_reversep,
- r_mergeable == 0
- ? NULL
- : r_mergeable < 0
- ? lr_load
- : rr_load);
+ lr_reversep, lr_load);
}
if (parts > 1)
@@ -8353,18 +8295,8 @@ fold_truth_andor_maybe_separate (location_t loc,
if (r_split_load)
{
gimple *point[2];
- if (r_mergeable == 0)
- point[0] = point[1] = NULL;
- else if (r_mergeable < 0)
- {
- point[0] = lr_load;
- point[1] = rr_load;
- }
- else
- {
- point[0] = rr_load;
- point[1] = lr_load;
- }
+ point[0] = lr_load;
+ point[1] = rr_load;
build_split_load (ld_arg[1], bitpos[1], bitsiz[1], toshift[1],
shifted[1], loc, lr_inner, lr_arg,
rnmode, rnmode2, rnbitpos, lr_reversep, point);
@@ -8430,12 +8362,7 @@ fold_truth_andor_maybe_separate (location_t loc,
ld_arg[0][0] = make_bit_field_load (loc, ll_inner, ll_arg,
lntype, lnbitsize, lnbitpos,
ll_unsignedp || rl_unsignedp,
- ll_reversep,
- l_mergeable == 0
- ? NULL
- : l_mergeable < 0
- ? ll_load
- : rl_load);
+ ll_reversep, ll_load);
}
if (parts > 1)
@@ -8443,18 +8370,8 @@ fold_truth_andor_maybe_separate (location_t loc,
if (l_split_load)
{
gimple *point[2];
- if (l_mergeable == 0)
- point[0] = point[1] = NULL;
- else if (l_mergeable < 0)
- {
- point[0] = ll_load;
- point[1] = rl_load;
- }
- else
- {
- point[0] = rl_load;
- point[1] = ll_load;
- }
+ point[0] = ll_load;
+ point[1] = rl_load;
build_split_load (ld_arg[0], bitpos[0], bitsiz[0], toshift[0],
shifted[0], loc, ll_inner, ll_arg,
lnmode, lnmode2, lnbitpos, ll_reversep, point);
diff --git a/gcc/testsuite/gcc.dg/field-merge-10.c
b/gcc/testsuite/gcc.dg/field-merge-10.c
new file mode 100644
index 000000000000..dca6febb7585
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/field-merge-10.c
@@ -0,0 +1,36 @@
+/* { dg-do run } */
+/* { dg-options "-O" } */
+
+/* Check that we don't move loads across stores. */
+
+struct s {
+ short a;
+ short b;
+} __attribute__ ((aligned (4)));
+
+struct s p[2] = {
+ { 42, 0xfe01 },
+ { 42, 0xfe10 }
+};
+
+void f (void) {
+ short a0 = p[0].a;
+ short b0 = p[0].b;
+ short a1 = p[1].a;
+ short b1 = p[1].b;
+ __builtin_memset (p, 0, sizeof (p));
+ asm ("" : "+m" (p));
+ if (0
+ || p[0].a != p[1].a
+ || p[0].b != p[1].b
+ )
+ __builtin_abort ();
+ if (a0 == a1)
+ if (b0 == b1)
+ __builtin_abort ();
+}
+
+int main () {
+ f ();
+ return 0;
+}
diff --git a/gcc/tree-ssa-ifcombine.cc b/gcc/tree-ssa-ifcombine.cc
index d204209923a8..dff6e02faa31 100644
--- a/gcc/tree-ssa-ifcombine.cc
+++ b/gcc/tree-ssa-ifcombine.cc
@@ -952,7 +952,8 @@ ifcombine_ifandif (basic_block inner_cond_bb, bool
inner_inv,
inner_cond_code,
gimple_cond_lhs (inner_cond),
gimple_cond_rhs (inner_cond),
- &ts))))
+ single_pred (inner_cond_bb) == outer_cond_bb
+ ? &ts : 0))))
{
tree t1, t2;
bool logical_op_non_short_circuit = LOGICAL_OP_NON_SHORT_CIRCUIT;
