* [PATCH] Improve PHI handling in DSE
@ 2021-05-03 13:10 Richard Biener
0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2021-05-03 13:10 UTC (permalink / raw)
To: gcc-patches
This improves handling of PHI defs when walking uses in
dse_classify_store to track two PHI defs. This happens
when there are CFG merges and one PHI feeds into another.
If we decide to want more then using a sbitmap for this might be
the way to go.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2021-05-03 Richard Biener <rguenther@suse.de>
* tree-ssa-dse.c (dse_classify_store): Track two PHI defs.
* gcc.dg/tree-ssa/ssa-dse-42.c: New testcase.
* gcc.dg/pr81192.c: Disable DSE.
---
gcc/testsuite/gcc.dg/pr81192.c | 4 +++-
gcc/testsuite/gcc.dg/tree-ssa/ssa-dse-42.c | 20 +++++++++++++++++++
gcc/tree-ssa-dse.c | 23 +++++++++++++---------
3 files changed, 37 insertions(+), 10 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-dse-42.c
diff --git a/gcc/testsuite/gcc.dg/pr81192.c b/gcc/testsuite/gcc.dg/pr81192.c
index 71bbc13a0e9..6cab6056558 100644
--- a/gcc/testsuite/gcc.dg/pr81192.c
+++ b/gcc/testsuite/gcc.dg/pr81192.c
@@ -1,4 +1,4 @@
-/* { dg-options "-Os -fdump-tree-pre-details -fdisable-tree-evrp" } */
+/* { dg-options "-Os -fdump-tree-pre-details -fdisable-tree-evrp -fno-tree-dse" } */
/* Disable tree-evrp because the new version of evrp sees
<bb 3> :
@@ -16,6 +16,8 @@ produces
# iftmp.2_12 = PHI <2147483647(3), iftmp.2_11(4)>
which causes the situation being tested to dissapear before we get to PRE. */
+/* Likewise disable DSE which also elides the tail merging "opportunity". */
+
#if __SIZEOF_INT__ == 2
#define unsigned __UINT32_TYPE__
#define int __INT32_TYPE__
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dse-42.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dse-42.c
new file mode 100644
index 00000000000..34108c83828
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dse-42.c
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-options "-O -fdump-tree-dse1" } */
+
+int a[2];
+void foo(int i, int k, int j)
+{
+ a[0] = i;
+ if (k)
+ a[0] = a[i] + k;
+ else
+ {
+ if (j)
+ a[1] = 1;
+ a[0] = a[i] + 3;
+ }
+ a[0] = 0;
+}
+
+/* The last stores to a[0] and a[1] remain. */
+/* { dg-final { scan-tree-dump-times " = " 2 "dse1" } } */
diff --git a/gcc/tree-ssa-dse.c b/gcc/tree-ssa-dse.c
index e0a944c704a..dfa6d314727 100644
--- a/gcc/tree-ssa-dse.c
+++ b/gcc/tree-ssa-dse.c
@@ -799,7 +799,8 @@ dse_classify_store (ao_ref *ref, gimple *stmt,
return DSE_STORE_LIVE;
auto_vec<gimple *, 10> defs;
- gimple *phi_def = NULL;
+ gimple *first_phi_def = NULL;
+ gimple *last_phi_def = NULL;
FOR_EACH_IMM_USE_STMT (use_stmt, ui, defvar)
{
/* Limit stmt walking. */
@@ -825,7 +826,9 @@ dse_classify_store (ao_ref *ref, gimple *stmt,
SSA_NAME_VERSION (PHI_RESULT (use_stmt))))
{
defs.safe_push (use_stmt);
- phi_def = use_stmt;
+ if (!first_phi_def)
+ first_phi_def = use_stmt;
+ last_phi_def = use_stmt;
}
}
/* If the statement is a use the store is not dead. */
@@ -889,6 +892,8 @@ dse_classify_store (ao_ref *ref, gimple *stmt,
gimple *def = defs[i];
gimple *use_stmt;
use_operand_p use_p;
+ tree vdef = (gimple_code (def) == GIMPLE_PHI
+ ? gimple_phi_result (def) : gimple_vdef (def));
/* If the path to check starts with a kill we do not need to
process it further.
??? With byte tracking we need only kill the bytes currently
@@ -901,8 +906,7 @@ dse_classify_store (ao_ref *ref, gimple *stmt,
}
/* If the path ends here we do not need to process it further.
This for example happens with calls to noreturn functions. */
- else if (gimple_code (def) != GIMPLE_PHI
- && has_zero_uses (gimple_vdef (def)))
+ else if (has_zero_uses (vdef))
{
/* But if the store is to global memory it is definitely
not dead. */
@@ -912,12 +916,13 @@ dse_classify_store (ao_ref *ref, gimple *stmt,
}
/* In addition to kills we can remove defs whose only use
is another def in defs. That can only ever be PHIs of which
- we track a single for simplicity reasons (we fail for multiple
- PHIs anyways). We can also ignore defs that feed only into
+ we track two for simplicity reasons, the first and last in
+ {first,last}_phi_def (we fail for multiple PHIs anyways).
+ We can also ignore defs that feed only into
already visited PHIs. */
- else if (gimple_code (def) != GIMPLE_PHI
- && single_imm_use (gimple_vdef (def), &use_p, &use_stmt)
- && (use_stmt == phi_def
+ else if (single_imm_use (vdef, &use_p, &use_stmt)
+ && (use_stmt == first_phi_def
+ || use_stmt == last_phi_def
|| (gimple_code (use_stmt) == GIMPLE_PHI
&& bitmap_bit_p (visited,
SSA_NAME_VERSION
--
2.26.2
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2021-05-03 13:10 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-05-03 13:10 [PATCH] Improve PHI handling in DSE 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).