The DF framework provides us a way to run dataflow problems on
sub-graphs. Naturally a bitmap of interesting blocks is passed into
those routines. At a confluence point, the DF framework will not mark
a block for re-processing if it's not in that set of interesting blocks.
When ext-dce sets up that set of interesting blocks it's using the wrong
counter. ie, it's using n_basic_blocks rather than last_basic_block.
If there are holes in the block indices, some number of blocks won't get
marked as interesting.
In this case the block needing reprocessing has an index higher than
n_basic_blocks. It never gets reprocessed and the newly found live
chunks don't propagate further up the CFG -- ultimately resulting in a
pseudo appearing to have only the low 8 bits live, when in fact the low
32 bits are actually live.
Fixed in the obvious way, by using last_basic_block instead.
Bootstrapped and regression tested on x86_64. Pushing to the trunk.
jeff
commit c9377734b798d8d311dfd3a5618dc49407703b93
Author: Jeff Law <j...@ventanamicro.com>
Date: Thu Aug 22 12:48:49 2024 -0600
[PR rtl-optimization/116420] Fix interesting block bitmap DF dataflow
The DF framework provides us a way to run dataflow problems on sub-graphs.
Naturally a bitmap of interesting blocks is passed into those routines.
At a
confluence point, the DF framework will not mark a block for re-processing
if
it's not in that set of interesting blocks.
When ext-dce sets up that set of interesting blocks it's using the wrong
counter. ie, it's using n_basic_blocks rather than last_basic_block. If
there
are holes in the block indices, some number of blocks won't get marked as
interesting.
In this case the block needing reprocessing has an index higher than
n_basic_blocks. It never gets reprocessed and the newly found live chunks
don't propagate further up the CFG -- ultimately resulting in a pseudo
appearing to have only the low 8 bits live, when in fact the low 32 bits are
actually live.
Fixed in the obvious way, by using last_basic_block instead.
Bootstrapped and regression tested on x86_64. Pushing to the trunk.
PR rtl-optimization/116420
gcc/
* ext-dce.cc (ext_dce_init): Fix loop iteration when setting up the
interesting block for DF to analyze.
gcc/testsuite
* gcc.dg/torture/pr116420.c: New test.
diff --git a/gcc/ext-dce.cc b/gcc/ext-dce.cc
index 35c06469b82..4a2503f1831 100644
--- a/gcc/ext-dce.cc
+++ b/gcc/ext-dce.cc
@@ -988,7 +988,7 @@ ext_dce_init (void)
all_blocks = BITMAP_ALLOC (NULL);
changed_pseudos = BITMAP_ALLOC (NULL);
- for (int i = 0; i < n_basic_blocks_for_fn (cfun); i++)
+ for (int i = 0; i < last_basic_block_for_fn (cfun); i++)
if (i != ENTRY_BLOCK && i != EXIT_BLOCK)
bitmap_set_bit (all_blocks, i);
diff --git a/gcc/testsuite/gcc.dg/torture/pr116420.c
b/gcc/testsuite/gcc.dg/torture/pr116420.c
new file mode 100644
index 00000000000..9a784f59429
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr116420.c
@@ -0,0 +1,17 @@
+/* { dg-do run } */
+/* { dg-additional-options "-fno-forward-propagate -fno-tree-ch" } */
+int a, d, e;
+char b = -1, c, f;
+int main() {
+ int g;
+ for (; d < 1; d++) {
+ g = b;
+ for (; c; c = g)
+ ;
+ }
+ f = g;
+ for (; e < 1; e++)
+ if (g >= a)
+ __builtin_abort();
+ return 0;
+}
