* [PATCH v4 1/3] [RFC] ifcvt: handle sequences that clobber flags in noce_convert_multiple_sets
2024-04-23 10:47 [PATCH v4 0/3] ifcvt: Allow if conversion of arithmetic in basic blocks with multiple sets Manolis Tsamis
@ 2024-04-23 10:47 ` Manolis Tsamis
2024-04-23 10:47 ` [PATCH v4 2/3] [RFC] ifcvt: Allow more operations in multiple set if conversion Manolis Tsamis
` (2 subsequent siblings)
3 siblings, 0 replies; 7+ messages in thread
From: Manolis Tsamis @ 2024-04-23 10:47 UTC (permalink / raw)
To: gcc-patches
Cc: Robin Dapp, Jiangning Liu, Philipp Tomsich, Richard Sandiford,
Jakub Jelinek, Manolis Tsamis
This is an extension of what was done in PR106590.
Currently if a sequence generated in noce_convert_multiple_sets clobbers the
condition rtx (cc_cmp or rev_cc_cmp) then only seq1 is used afterwards
(sequences that emit the comparison itself). Since this applies only from the
next iteration it assumes that the sequences generated (in particular seq2)
doesn't clobber the condition rtx itself before using it in the if_then_else,
which is only true in specific cases (currently only register/subregister moves
are allowed).
This patch changes this so it also tests if seq2 clobbers cc_cmp/rev_cc_cmp in
the current iteration. This makes it possible to include arithmetic operations
in noce_convert_multiple_sets.
It also makes the code that checks whether the condition is used outside of the
if_then_else emitted more robust.
gcc/ChangeLog:
* ifcvt.cc (check_for_cc_cmp_clobbers): Use modified_in_p instead.
(noce_convert_multiple_sets_1): Don't use seq2 if it clobbers cc_cmp.
Refactor the code that sets read_comparison.
Signed-off-by: Manolis Tsamis <manolis.tsamis@vrull.eu>
---
(no changes since v1)
gcc/ifcvt.cc | 106 ++++++++++++++++++++++++++++-----------------------
1 file changed, 59 insertions(+), 47 deletions(-)
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index 58ed42673e5..763a25f816e 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -3592,20 +3592,6 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
return true;
}
-/* Helper function for noce_convert_multiple_sets_1. If store to
- DEST can affect P[0] or P[1], clear P[0]. Called via note_stores. */
-
-static void
-check_for_cc_cmp_clobbers (rtx dest, const_rtx, void *p0)
-{
- rtx *p = (rtx *) p0;
- if (p[0] == NULL_RTX)
- return;
- if (reg_overlap_mentioned_p (dest, p[0])
- || (p[1] && reg_overlap_mentioned_p (dest, p[1])))
- p[0] = NULL_RTX;
-}
-
/* This goes through all relevant insns of IF_INFO->then_bb and tries to
create conditional moves. In case a simple move sufficis the insn
should be listed in NEED_NO_CMOV. The rewired-src cases should be
@@ -3731,36 +3717,67 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
creating an additional compare for each. If successful, costing
is easier and this sequence is usually preferred. */
if (cc_cmp)
- seq2 = try_emit_cmove_seq (if_info, temp, cond,
- new_val, old_val, need_cmov,
- &cost2, &temp_dest2, cc_cmp, rev_cc_cmp);
+ {
+ seq2 = try_emit_cmove_seq (if_info, temp, cond,
+ new_val, old_val, need_cmov,
+ &cost2, &temp_dest2, cc_cmp, rev_cc_cmp);
+
+ /* The if_then_else in SEQ2 may be affected when cc_cmp/rev_cc_cmp is
+ clobbered. We can't safely use the sequence in this case. */
+ if (seq2 && (modified_in_p (cc_cmp, seq2)
+ || (rev_cc_cmp && modified_in_p (rev_cc_cmp, seq2))))
+ seq2 = NULL;
+ }
/* The backend might have created a sequence that uses the
- condition. Check this. */
+ condition as a value. Check this. */
+
+ /* We cannot handle anything more complex than a reg or constant. */
+ if (!REG_P (XEXP (cond, 0)) && !CONSTANT_P (XEXP (cond, 0)))
+ read_comparison = true;
+
+ if (!REG_P (XEXP (cond, 1)) && !CONSTANT_P (XEXP (cond, 1)))
+ read_comparison = true;
+
rtx_insn *walk = seq2;
- while (walk)
+ int if_then_else_count = 0;
+ while (walk && !read_comparison)
{
- rtx set = single_set (walk);
+ rtx exprs_to_check[2];
+ unsigned int exprs_count = 0;
- if (!set || !SET_SRC (set))
+ rtx set = single_set (walk);
+ if (set && XEXP (set, 1)
+ && GET_CODE (XEXP (set, 1)) == IF_THEN_ELSE)
{
- walk = NEXT_INSN (walk);
- continue;
+ /* We assume that this is the cmove created by the backend that
+ naturally uses the condition. */
+ exprs_to_check[exprs_count++] = XEXP (XEXP (set, 1), 1);
+ exprs_to_check[exprs_count++] = XEXP (XEXP (set, 1), 2);
+ if_then_else_count++;
}
+ else if (NONDEBUG_INSN_P (walk))
+ exprs_to_check[exprs_count++] = PATTERN (walk);
- rtx src = SET_SRC (set);
+ /* Bail if we get more than one if_then_else because the assumption
+ above may be incorrect. */
+ if (if_then_else_count > 1)
+ {
+ read_comparison = true;
+ break;
+ }
- if (XEXP (set, 1) && GET_CODE (XEXP (set, 1)) == IF_THEN_ELSE)
- ; /* We assume that this is the cmove created by the backend that
- naturally uses the condition. Therefore we ignore it. */
- else
+ for (unsigned int i = 0; i < exprs_count; i++)
{
- if (reg_mentioned_p (XEXP (cond, 0), src)
- || reg_mentioned_p (XEXP (cond, 1), src))
- {
- read_comparison = true;
- break;
- }
+ subrtx_iterator::array_type array;
+ FOR_EACH_SUBRTX (iter, array, exprs_to_check[i], NONCONST)
+ if (*iter != NULL_RTX
+ && (reg_overlap_mentioned_p (XEXP (cond, 0), *iter)
+ || reg_overlap_mentioned_p (XEXP (cond, 1), *iter)))
+ {
+ read_comparison = true;
+ break;
+ }
}
walk = NEXT_INSN (walk);
@@ -3788,21 +3805,16 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
return false;
}
- if (cc_cmp)
+ if (cc_cmp && seq == seq1)
{
- /* Check if SEQ can clobber registers mentioned in
- cc_cmp and/or rev_cc_cmp. If yes, we need to use
- only seq1 from that point on. */
- rtx cc_cmp_pair[2] = { cc_cmp, rev_cc_cmp };
- for (walk = seq; walk; walk = NEXT_INSN (walk))
+ /* Check if SEQ can clobber registers mentioned in cc_cmp/rev_cc_cmp.
+ If yes, we need to use only seq1 from that point on.
+ Only check when we use seq1 since we have already tested seq2. */
+ if (modified_in_p (cc_cmp, seq)
+ || (rev_cc_cmp && modified_in_p (rev_cc_cmp, seq)))
{
- note_stores (walk, check_for_cc_cmp_clobbers, cc_cmp_pair);
- if (cc_cmp_pair[0] == NULL_RTX)
- {
- cc_cmp = NULL_RTX;
- rev_cc_cmp = NULL_RTX;
- break;
- }
+ cc_cmp = NULL_RTX;
+ rev_cc_cmp = NULL_RTX;
}
}
--
2.34.1
^ permalink raw reply [flat|nested] 7+ messages in thread* [PATCH v4 2/3] [RFC] ifcvt: Allow more operations in multiple set if conversion
2024-04-23 10:47 [PATCH v4 0/3] ifcvt: Allow if conversion of arithmetic in basic blocks with multiple sets Manolis Tsamis
2024-04-23 10:47 ` [PATCH v4 1/3] [RFC] ifcvt: handle sequences that clobber flags in noce_convert_multiple_sets Manolis Tsamis
@ 2024-04-23 10:47 ` Manolis Tsamis
2024-04-24 23:40 ` Hans-Peter Nilsson
2024-04-23 10:47 ` [PATCH v4 3/3] [RFC] ifcvt: Handle multiple rewired regs and refactor noce_convert_multiple_sets Manolis Tsamis
2024-05-14 12:06 ` [PATCH v4 0/3] ifcvt: Allow if conversion of arithmetic in basic blocks with multiple sets Manolis Tsamis
3 siblings, 1 reply; 7+ messages in thread
From: Manolis Tsamis @ 2024-04-23 10:47 UTC (permalink / raw)
To: gcc-patches
Cc: Robin Dapp, Jiangning Liu, Philipp Tomsich, Richard Sandiford,
Jakub Jelinek, Manolis Tsamis
Currently the operations allowed for if conversion of a basic block with
multiple sets are few, namely REG, SUBREG and CONST_INT (as controlled by
bb_ok_for_noce_convert_multiple_sets).
This commit allows more operations (arithmetic, compare, etc) to participate
in if conversion. The target's profitability hook and ifcvt's costing is
expected to reject sequences that are unprofitable.
This is especially useful for targets which provide a rich selection of
conditional instructions (like aarch64 which has cinc, csneg, csinv, ccmp, ...)
which are currently not used in basic blocks with more than a single set.
gcc/ChangeLog:
* ifcvt.cc (try_emit_cmove_seq): Modify comments.
(noce_convert_multiple_sets_1): Modify comments.
(bb_ok_for_noce_convert_multiple_sets): Allow more operations.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/ifcvt_multiple_sets_arithm.c: New test.
Signed-off-by: Manolis Tsamis <manolis.tsamis@vrull.eu>
---
Changes in v4:
- Remove unnecessary hardcoded list of allowed ops in
bb_ok_for_noce_convert_multiple_sets.
- Set need_cmov based on BB live_out instead of REG_DEAD notes.
- Fix preexisting issues and improve the code that sets read_comparison.
gcc/ifcvt.cc | 34 +++-----
.../aarch64/ifcvt_multiple_sets_arithm.c | 79 +++++++++++++++++++
2 files changed, 92 insertions(+), 21 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_arithm.c
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index 763a25f816e..dc00042be81 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -3432,13 +3432,13 @@ try_emit_cmove_seq (struct noce_if_info *if_info, rtx temp,
/* We have something like:
if (x > y)
- { i = a; j = b; k = c; }
+ { i = EXPR_A; j = EXPR_B; k = EXPR_C; }
Make it:
- tmp_i = (x > y) ? a : i;
- tmp_j = (x > y) ? b : j;
- tmp_k = (x > y) ? c : k;
+ tmp_i = (x > y) ? EXPR_A : i;
+ tmp_j = (x > y) ? EXPR_B : j;
+ tmp_k = (x > y) ? EXPR_C : k;
i = tmp_i;
j = tmp_j;
k = tmp_k;
@@ -3839,11 +3839,10 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
-/* Return true iff basic block TEST_BB is comprised of only
- (SET (REG) (REG)) insns suitable for conversion to a series
- of conditional moves. Also check that we have more than one set
- (other routines can handle a single set better than we would), and
- fewer than PARAM_MAX_RTL_IF_CONVERSION_INSNS sets. While going
+/* Return true iff basic block TEST_BB is suitable for conversion to a
+ series of conditional moves. Also check that we have more than one
+ set (other routines can handle a single set better than we would),
+ and fewer than PARAM_MAX_RTL_IF_CONVERSION_INSNS sets. While going
through the insns store the sum of their potential costs in COST. */
static bool
@@ -3869,20 +3868,13 @@ bb_ok_for_noce_convert_multiple_sets (basic_block test_bb, unsigned *cost)
rtx dest = SET_DEST (set);
rtx src = SET_SRC (set);
- /* We can possibly relax this, but for now only handle REG to REG
- (including subreg) moves. This avoids any issues that might come
- from introducing loads/stores that might violate data-race-freedom
- guarantees. */
- if (!REG_P (dest))
- return false;
-
- if (!((REG_P (src) || CONSTANT_P (src))
- || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
- && subreg_lowpart_p (src))))
+ /* Do not handle anything involving memory loads/stores since it might
+ violate data-race-freedom guarantees. */
+ if (!REG_P (dest) || contains_mem_rtx_p (src))
return false;
- /* Destination must be appropriate for a conditional write. */
- if (!noce_operand_ok (dest))
+ /* Destination and source must be appropriate. */
+ if (!noce_operand_ok (dest) || !noce_operand_ok (src))
return false;
/* We must be able to conditionally move in this mode. */
diff --git a/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_arithm.c b/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_arithm.c
new file mode 100644
index 00000000000..d977f4d62ec
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_arithm.c
@@ -0,0 +1,79 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-rtl-ce1" } */
+
+void sink2(int, int);
+void sink3(int, int, int);
+
+void cond1(int cond, int x, int y)
+{
+ if (cond)
+ {
+ x = x << 4;
+ y = 1;
+ }
+
+ sink2(x, y);
+}
+
+void cond2(int cond, int x, int y)
+{
+ if (cond)
+ {
+ x++;
+ y++;
+ }
+
+ sink2(x, y);
+}
+
+void cond3(int cond, int x1, int x2, int x3)
+{
+ if (cond)
+ {
+ x1++;
+ x2++;
+ x3++;
+ }
+
+ sink3(x1, x2, x3);
+}
+
+void cond4(int cond, int x, int y)
+{
+ if (cond)
+ {
+ x += 2;
+ y += 3;
+ }
+
+ sink2(x, y);
+}
+
+void cond5(int cond, int x, int y, int r1, int r2)
+{
+ if (cond)
+ {
+ x = r1 + 2;
+ y = r2 - 34;
+ }
+
+ sink2(x, y);
+}
+
+void cond6(int cond, int x, int y)
+{
+ if (cond)
+ {
+ x = -x;
+ y = ~y;
+ }
+
+ sink2(x, y);
+}
+
+/* { dg-final { scan-assembler-times "cinc\t" 5 } } */
+/* { dg-final { scan-assembler-times "csneg\t" 1 } } */
+/* { dg-final { scan-assembler-times "csinv\t" 1 } } */
+/* { dg-final { scan-assembler "csel\t" } } */
+
+/* { dg-final { scan-rtl-dump-times "if-conversion succeeded through noce_convert_multiple_sets" 6 "ce1" } } */
\ No newline at end of file
--
2.34.1
^ permalink raw reply [flat|nested] 7+ messages in thread* [PATCH v4 3/3] [RFC] ifcvt: Handle multiple rewired regs and refactor noce_convert_multiple_sets
2024-04-23 10:47 [PATCH v4 0/3] ifcvt: Allow if conversion of arithmetic in basic blocks with multiple sets Manolis Tsamis
2024-04-23 10:47 ` [PATCH v4 1/3] [RFC] ifcvt: handle sequences that clobber flags in noce_convert_multiple_sets Manolis Tsamis
2024-04-23 10:47 ` [PATCH v4 2/3] [RFC] ifcvt: Allow more operations in multiple set if conversion Manolis Tsamis
@ 2024-04-23 10:47 ` Manolis Tsamis
2024-05-14 12:06 ` [PATCH v4 0/3] ifcvt: Allow if conversion of arithmetic in basic blocks with multiple sets Manolis Tsamis
3 siblings, 0 replies; 7+ messages in thread
From: Manolis Tsamis @ 2024-04-23 10:47 UTC (permalink / raw)
To: gcc-patches
Cc: Robin Dapp, Jiangning Liu, Philipp Tomsich, Richard Sandiford,
Jakub Jelinek, Manolis Tsamis
The existing implementation of need_cmov_or_rewire and
noce_convert_multiple_sets_1 assumes that sets are either REG or SUBREG.
This commit enchances them so they can handle/rewire arbitrary set statements.
To do that a new helper struct noce_multiple_sets_info is introduced which is
used by noce_convert_multiple_sets and its helper functions. This results in
cleaner function signatures, improved efficientcy (a number of vecs and hash
set/map are replaced with a single vec of struct) and simplicity.
gcc/ChangeLog:
* ifcvt.cc (need_cmov_or_rewire): Renamed init_noce_multiple_sets_info.
(init_noce_multiple_sets_info): Initialize noce_multiple_sets_info.
(noce_convert_multiple_sets_1): Use noce_multiple_sets_info and handle
rewiring of multiple registers.
(noce_convert_multiple_sets): Updated to use noce_multiple_sets_info.
* ifcvt.h (struct noce_multiple_sets_info): Introduce new struct
noce_multiple_sets_info to store info for noce_convert_multiple_sets.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/ifcvt_multiple_sets_rewire.c: New test.
Signed-off-by: Manolis Tsamis <manolis.tsamis@vrull.eu>
---
(no changes since v1)
gcc/ifcvt.cc | 243 ++++++++----------
gcc/ifcvt.h | 16 ++
.../aarch64/ifcvt_multiple_sets_rewire.c | 20 ++
3 files changed, 141 insertions(+), 138 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_rewire.c
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index dc00042be81..8e36c16ee57 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -98,14 +98,10 @@ static bool dead_or_predicable (basic_block, basic_block, basic_block,
edge, bool);
static void noce_emit_move_insn (rtx, rtx);
static rtx_insn *block_has_only_trap (basic_block);
-static void need_cmov_or_rewire (basic_block, hash_set<rtx_insn *> *,
- hash_map<rtx_insn *, int> *);
+static void init_noce_multiple_sets_info (basic_block,
+ auto_delete_vec<noce_multiple_sets_info> &);
static bool noce_convert_multiple_sets_1 (struct noce_if_info *,
- hash_set<rtx_insn *> *,
- hash_map<rtx_insn *, int> *,
- auto_vec<rtx> *,
- auto_vec<rtx> *,
- auto_vec<rtx_insn *> *, int *);
+ auto_delete_vec<noce_multiple_sets_info> &, int *);
\f
/* Count the number of non-jump active insns in BB. */
@@ -3487,24 +3483,13 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
rtx x = XEXP (cond, 0);
rtx y = XEXP (cond, 1);
- /* The true targets for a conditional move. */
- auto_vec<rtx> targets;
- /* The temporaries introduced to allow us to not consider register
- overlap. */
- auto_vec<rtx> temporaries;
- /* The insns we've emitted. */
- auto_vec<rtx_insn *> unmodified_insns;
-
- hash_set<rtx_insn *> need_no_cmov;
- hash_map<rtx_insn *, int> rewired_src;
-
- need_cmov_or_rewire (then_bb, &need_no_cmov, &rewired_src);
+ auto_delete_vec<noce_multiple_sets_info> insn_info;
+ init_noce_multiple_sets_info (then_bb, insn_info);
int last_needs_comparison = -1;
bool ok = noce_convert_multiple_sets_1
- (if_info, &need_no_cmov, &rewired_src, &targets, &temporaries,
- &unmodified_insns, &last_needs_comparison);
+ (if_info, insn_info, &last_needs_comparison);
if (!ok)
return false;
@@ -3519,8 +3504,7 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
end_sequence ();
start_sequence ();
ok = noce_convert_multiple_sets_1
- (if_info, &need_no_cmov, &rewired_src, &targets, &temporaries,
- &unmodified_insns, &last_needs_comparison);
+ (if_info, insn_info, &last_needs_comparison);
/* Actually we should not fail anymore if we reached here,
but better still check. */
if (!ok)
@@ -3529,12 +3513,12 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
/* We must have seen some sort of insn to insert, otherwise we were
given an empty BB to convert, and we can't handle that. */
- gcc_assert (!unmodified_insns.is_empty ());
+ gcc_assert (!insn_info.is_empty ());
/* Now fixup the assignments. */
- for (unsigned i = 0; i < targets.length (); i++)
- if (targets[i] != temporaries[i])
- noce_emit_move_insn (targets[i], temporaries[i]);
+ for (unsigned i = 0; i < insn_info.length (); i++)
+ if (insn_info[i]->target != insn_info[i]->temporary)
+ noce_emit_move_insn (insn_info[i]->target, insn_info[i]->temporary);
/* Actually emit the sequence if it isn't too expensive. */
rtx_insn *seq = get_insns ();
@@ -3549,10 +3533,10 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
set_used_flags (insn);
/* Mark all our temporaries and targets as used. */
- for (unsigned i = 0; i < targets.length (); i++)
+ for (unsigned i = 0; i < insn_info.length (); i++)
{
- set_used_flags (temporaries[i]);
- set_used_flags (targets[i]);
+ set_used_flags (insn_info[i]->temporary);
+ set_used_flags (insn_info[i]->target);
}
set_used_flags (cond);
@@ -3571,7 +3555,7 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
return false;
emit_insn_before_setloc (seq, if_info->jump,
- INSN_LOCATION (unmodified_insns.last ()));
+ INSN_LOCATION (insn_info.last ()->unmodified_insn));
/* Clean up THEN_BB and the edges in and out of it. */
remove_edge (find_edge (test_bb, join_bb));
@@ -3592,20 +3576,12 @@ noce_convert_multiple_sets (struct noce_if_info *if_info)
return true;
}
-/* This goes through all relevant insns of IF_INFO->then_bb and tries to
- create conditional moves. In case a simple move sufficis the insn
- should be listed in NEED_NO_CMOV. The rewired-src cases should be
- specified via REWIRED_SRC. TARGETS, TEMPORARIES and UNMODIFIED_INSNS
- are specified and used in noce_convert_multiple_sets and should be passed
- to this function.. */
+/* This goes through all relevant insns of IF_INFO->then_bb and tries to create
+ conditional moves. Information for the insns is kept in INSN_INFO. */
static bool
noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
- hash_set<rtx_insn *> *need_no_cmov,
- hash_map<rtx_insn *, int> *rewired_src,
- auto_vec<rtx> *targets,
- auto_vec<rtx> *temporaries,
- auto_vec<rtx_insn *> *unmodified_insns,
+ auto_delete_vec<noce_multiple_sets_info> &insn_info,
int *last_needs_comparison)
{
basic_block then_bb = if_info->then_bb;
@@ -3624,11 +3600,6 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
rtx_insn *insn;
int count = 0;
-
- targets->truncate (0);
- temporaries->truncate (0);
- unmodified_insns->truncate (0);
-
bool second_try = *last_needs_comparison != -1;
FOR_BB_INSNS (then_bb, insn)
@@ -3637,6 +3608,8 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
if (!active_insn_p (insn))
continue;
+ noce_multiple_sets_info *info = insn_info[count];
+
rtx set = single_set (insn);
gcc_checking_assert (set);
@@ -3644,9 +3617,12 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
rtx temp;
rtx new_val = SET_SRC (set);
- if (int *ii = rewired_src->get (insn))
- new_val = simplify_replace_rtx (new_val, (*targets)[*ii],
- (*temporaries)[*ii]);
+
+ int i, ii;
+ FOR_EACH_VEC_ELT (info->rewired_src, i, ii)
+ new_val = simplify_replace_rtx (new_val, insn_info[ii]->target,
+ insn_info[ii]->temporary);
+
rtx old_val = target;
/* As we are transforming
@@ -3687,7 +3663,7 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
/* We have identified swap-style idioms before. A normal
set will need to be a cmov while the first instruction of a swap-style
idiom can be a regular move. This helps with costing. */
- bool need_cmov = !need_no_cmov->contains (insn);
+ bool need_cmov = info->need_cmov;
/* If we had a non-canonical conditional jump (i.e. one where
the fallthrough is to the "else" case) we need to reverse
@@ -3823,9 +3799,10 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
/* Bookkeeping. */
count++;
- targets->safe_push (target);
- temporaries->safe_push (temp_dest);
- unmodified_insns->safe_push (insn);
+
+ info->target = target;
+ info->temporary = temp_dest;
+ info->unmodified_insn = insn;
}
/* Even if we did not actually need the comparison, we want to make sure
@@ -3833,11 +3810,84 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info,
if (*last_needs_comparison == -1)
*last_needs_comparison = 0;
-
return true;
}
+/* Find local swap-style idioms in BB and mark the first insn (1)
+ that is only a temporary as not needing a conditional move as
+ it is going to be dead afterwards anyway.
+
+ (1) int tmp = a;
+ a = b;
+ b = tmp;
+
+ ifcvt
+ -->
+
+ tmp = a;
+ a = cond ? b : a_old;
+ b = cond ? tmp : b_old;
+
+ Additionally, store the index of insns like (2) when a subsequent
+ SET reads from their destination.
+
+ (2) int c = a;
+ int d = c;
+
+ ifcvt
+ -->
+ c = cond ? a : c_old;
+ d = cond ? d : c; // Need to use c rather than c_old here.
+*/
+
+static void
+init_noce_multiple_sets_info (basic_block bb,
+ auto_delete_vec<noce_multiple_sets_info> &insn_info)
+{
+ rtx_insn *insn;
+ int count = 0;
+ auto_vec<rtx> dests;
+ bitmap bb_live_out = df_get_live_out (bb);
+
+ /* Iterate over all SETs, storing the destinations in DEST.
+ - If we encounter a previously changed register,
+ rewire the read to the original source.
+ - If we encounter a SET that writes to a destination
+ that is not live after this block then the register
+ does not need to be moved conditionally. */
+ FOR_BB_INSNS (bb, insn)
+ {
+ if (!active_insn_p (insn))
+ continue;
+
+ noce_multiple_sets_info *info = new noce_multiple_sets_info;
+ info->target = NULL_RTX;
+ info->temporary = NULL_RTX;
+ info->unmodified_insn = NULL;
+ insn_info.safe_push (info);
+
+ rtx set = single_set (insn);
+ gcc_checking_assert (set);
+
+ rtx src = SET_SRC (set);
+ rtx dest = SET_DEST (set);
+
+ gcc_checking_assert (REG_P (dest));
+ info->need_cmov = bitmap_bit_p (bb_live_out, REGNO (dest));
+
+ /* Check if the current SET's source is the same
+ as any previously seen destination.
+ This is quadratic but the number of insns in BB
+ is bounded by PARAM_MAX_RTL_IF_CONVERSION_INSNS. */
+ for (int i = count - 1; i >= 0; --i)
+ if (reg_mentioned_p (dests[i], src))
+ insn_info[count]->rewired_src.safe_push (i);
+
+ dests.safe_push (dest);
+ count++;
+ }
+}
/* Return true iff basic block TEST_BB is suitable for conversion to a
series of conditional moves. Also check that we have more than one
@@ -4307,89 +4357,6 @@ check_cond_move_block (basic_block bb,
return true;
}
-/* Find local swap-style idioms in BB and mark the first insn (1)
- that is only a temporary as not needing a conditional move as
- it is going to be dead afterwards anyway.
-
- (1) int tmp = a;
- a = b;
- b = tmp;
-
- ifcvt
- -->
-
- tmp = a;
- a = cond ? b : a_old;
- b = cond ? tmp : b_old;
-
- Additionally, store the index of insns like (2) when a subsequent
- SET reads from their destination.
-
- (2) int c = a;
- int d = c;
-
- ifcvt
- -->
-
- c = cond ? a : c_old;
- d = cond ? d : c; // Need to use c rather than c_old here.
-*/
-
-static void
-need_cmov_or_rewire (basic_block bb,
- hash_set<rtx_insn *> *need_no_cmov,
- hash_map<rtx_insn *, int> *rewired_src)
-{
- rtx_insn *insn;
- int count = 0;
- auto_vec<rtx_insn *> insns;
- auto_vec<rtx> dests;
-
- /* Iterate over all SETs, storing the destinations
- in DEST.
- - If we hit a SET that reads from a destination
- that we have seen before and the corresponding register
- is dead afterwards, the register does not need to be
- moved conditionally.
- - If we encounter a previously changed register,
- rewire the read to the original source. */
- FOR_BB_INSNS (bb, insn)
- {
- rtx set, src, dest;
-
- if (!active_insn_p (insn))
- continue;
-
- set = single_set (insn);
- if (set == NULL_RTX)
- continue;
-
- src = SET_SRC (set);
- if (SUBREG_P (src))
- src = SUBREG_REG (src);
- dest = SET_DEST (set);
-
- /* Check if the current SET's source is the same
- as any previously seen destination.
- This is quadratic but the number of insns in BB
- is bounded by PARAM_MAX_RTL_IF_CONVERSION_INSNS. */
- if (REG_P (src))
- for (int i = count - 1; i >= 0; --i)
- if (reg_overlap_mentioned_p (src, dests[i]))
- {
- if (find_reg_note (insn, REG_DEAD, src) != NULL_RTX)
- need_no_cmov->add (insns[i]);
- else
- rewired_src->put (insn, i);
- }
-
- insns.safe_push (insn);
- dests.safe_push (dest);
-
- count++;
- }
-}
-
/* Given a basic block BB suitable for conditional move conversion,
a condition COND, and pointer maps THEN_VALS and ELSE_VALS containing
the register values depending on COND, emit the insns in the block as
diff --git a/gcc/ifcvt.h b/gcc/ifcvt.h
index 37147f99129..204bcf6d18a 100644
--- a/gcc/ifcvt.h
+++ b/gcc/ifcvt.h
@@ -40,6 +40,22 @@ struct ce_if_block
int pass; /* Pass number. */
};
+struct noce_multiple_sets_info
+{
+ /* A list of indices to instructions that we need to rewire into this
+ instruction when we replace them with temporary conditional moves. */
+ auto_vec<int> rewired_src;
+ /* The true targets for a conditional move. */
+ rtx target;
+ /* The temporaries introduced to allow us to not consider register
+ overlap. */
+ rtx temporary;
+ /* The insns we've emitted. */
+ rtx_insn *unmodified_insn;
+ /* True if a simple move can be used instead of a conditional move. */
+ bool need_cmov;
+};
+
/* Used by noce_process_if_block to communicate with its subroutines.
The subroutines know that A and B may be evaluated freely. They
diff --git a/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_rewire.c b/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_rewire.c
new file mode 100644
index 00000000000..411874e96c2
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/ifcvt_multiple_sets_rewire.c
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-rtl-ce1" } */
+
+void sink2(int, int);
+
+void cond1(int cond, int x, int y, int z)
+{
+ if (x)
+ {
+ x = y + z;
+ y = z + x;
+ }
+
+ sink2(x, y);
+}
+
+/* { dg-final { scan-assembler-times "csel\tw0, w0, w1" 1 } } */
+/* { dg-final { scan-assembler-times "csel\tw1, w3, w2" 1 } } */
+
+/* { dg-final { scan-rtl-dump-times "if-conversion succeeded through noce_convert_multiple_sets" 1 "ce1" } } */
\ No newline at end of file
--
2.34.1
^ permalink raw reply [flat|nested] 7+ messages in thread