* [PATCH] allow nul-over-nul elimination only for single-byte stores (PR 93213)
@ 2020-01-10 14:21 Martin Sebor
0 siblings, 0 replies; only message in thread
From: Martin Sebor @ 2020-01-10 14:21 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 589 bytes --]
The multi-byte store enhancement to the strlen optimization checked
sometime last summer didn't take care to prevent the nul-over-nul
store elimination of multi-byte assignments. This made it possible
for subsequent multi-byte stores of fewer nuls to cause prior larger
stores to be eliminated. The latent bug was exposed by a small
unrelated change last December that had masked it until then.
The attached patch makes sure that only single-byte nul stores are
considered as candidates for this elimination. I will commit it
next week if there are no concerns or suggestions.
Martin
[-- Attachment #2: gcc-93213.diff --]
[-- Type: text/x-patch, Size: 2408 bytes --]
PR tree-optimization/93213 - wrong code with -Og -foptimize-strlen
gcc/testsuite/ChangeLog:
PR tree-optimization/93213
* gcc.c-torture/execute/pr93213.c: New test.
gcc/ChangeLog:
PR tree-optimization/93213
* tree-ssa-strlen.c (handle_store): Only allow single-byte nul-over-nul
stores to be eliminated.
Index: gcc/testsuite/gcc.c-torture/execute/pr93213.c
===================================================================
--- gcc/testsuite/gcc.c-torture/execute/pr93213.c (nonexistent)
+++ gcc/testsuite/gcc.c-torture/execute/pr93213.c (working copy)
@@ -0,0 +1,45 @@
+/* PR tree-optimization/93213 - wrong code on a multibyte store with
+ -Og -foptimize-strlen
+ { dg-additional-options "-Og -foptimize-strlen" } */
+
+typedef unsigned short u16;
+typedef unsigned int u32;
+typedef unsigned __int128 u128;
+
+static inline u128
+foo (u16 u16_1, u32 u32_1, u128 u128_1)
+{
+ u128 u128_0 = 0;
+ u128_1 -= __builtin_mul_overflow (u32_1, u16_1, &u32_1);
+ __builtin_memmove (&u16_1, &u128_0, 2);
+ __builtin_memmove (&u16_1, &u128_1, 1);
+ return u16_1;
+}
+
+__attribute__ ((noipa)) void
+bar (void)
+{
+ char a[] = { 1, 2 };
+ const char b[] = { 0, 0 };
+ const char c[] = { 2 };
+ __builtin_memcpy (a, b, 2);
+ // The above is transformed into
+ // MEM <short unsigned int> [(char * {ref-all})&a] = 0;
+ // which was then dropped because of the non-nul store below.
+ __builtin_memcpy (a, c, 1);
+
+ volatile char *p = a;
+ if (p[0] != 2 || p[1] != 0)
+ __builtin_abort ();
+}
+
+int
+main (void)
+{
+ u16 x = foo (-1, -1, 0);
+ if (x != 0xff)
+ __builtin_abort ();
+
+ bar ();
+ return 0;
+}
Index: gcc/tree-ssa-strlen.c
===================================================================
--- gcc/tree-ssa-strlen.c (revision 280100)
+++ gcc/tree-ssa-strlen.c (working copy)
@@ -5239,10 +5239,10 @@ handle_store (gimple_stmt_iterator *gsi, bool *zer
}
}
- if (si != NULL && offset == 0 && storing_all_zeros_p)
+ if (si != NULL && offset == 0 && storing_all_zeros_p && lenrange[2] == 1)
{
- /* Allow adjust_last_stmt to remove it if the stored '\0'
- is immediately overwritten. */
+ /* For single-byte stores only, allow adjust_last_stmt to remove
+ the statement if the stored '\0' is immediately overwritten. */
laststmt.stmt = stmt;
laststmt.len = build_int_cst (size_type_node, 1);
laststmt.stridx = si->idx;
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2020-01-10 14:13 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-01-10 14:21 [PATCH] allow nul-over-nul elimination only for single-byte stores (PR 93213) Martin Sebor
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).