On Wed, 27 Apr 2022, Richard Biener wrote:
> The following makes sure to take into account prologue peeling
> when trying to narrow down the maximum number of iterations
> computed for the epilogue of a vectorized epilogue.
>
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
>
> I did not verify this solves the original aarch64 testcase yet
> but it looks like a simpler fix and explains why I don't see
> the issue on the 11 branch which does otherwise the same transforms.
So the following also includes ->peeling_for_gaps which should have
a similar issue. We discussed being more precise here but I think
we should do this as followup during stage1.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Can you verify the original aarch64 case is fixed and maybe add
a testcase for that to gcc.target/aarch64?
OK for trunk?
Thanks,
Richard.
>From 03f437c9b2d5260888535e0dad570daee694e4b3 Mon Sep 17 00:00:00 2001
From: Richard Biener <rguent...@suse.de>
Date: Wed, 27 Apr 2022 14:06:12 +0200
Subject: [PATCH] tree-optimization/105219 - bogus max iters for vectorized
epilogue
To: gcc-patches@gcc.gnu.org
The following makes sure to take into account prologue peeling
when trying to narrow down the maximum number of iterations
computed for the vectorized epilogue. A similar issue exists when
peeling for gaps.
2022-04-27 Richard Biener <rguent...@suse.de>
PR tree-optimization/105219
* tree-vect-loop.cc (vect_transform_loop): Disable
special code narrowing the vectorized epilogue max
iterations when peeling for alignment or gaps was in effect.
* gcc.dg/vect/pr105219.c: New testcase.
---
gcc/testsuite/gcc.dg/vect/pr105219.c | 29 ++++++++++++++++++++++++++++
gcc/tree-vect-loop.cc | 6 +++++-
2 files changed, 34 insertions(+), 1 deletion(-)
create mode 100644 gcc/testsuite/gcc.dg/vect/pr105219.c
diff --git a/gcc/testsuite/gcc.dg/vect/pr105219.c
b/gcc/testsuite/gcc.dg/vect/pr105219.c
new file mode 100644
index 00000000000..0cb7ae2f4d6
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr105219.c
@@ -0,0 +1,29 @@
+/* { dg-do run } */
+/* { dg-additional-options "-O3" } */
+/* { dg-additional-options "-mtune=intel" { target x86_64-*-* i?86-*-* } } */
+
+#include "tree-vect.h"
+
+int data[128];
+
+void __attribute((noipa))
+foo (int *data, int n)
+{
+ for (int i = 0; i < n; ++i)
+ data[i] = i;
+}
+
+int main()
+{
+ check_vect ();
+ for (int start = 0; start < 16; ++start)
+ for (int n = 1; n < 3*16; ++n)
+ {
+ __builtin_memset (data, 0, sizeof (data));
+ foo (&data[start], n);
+ for (int j = 0; j < n; ++j)
+ if (data[start + j] != j)
+ __builtin_abort ();
+ }
+ return 0;
+}
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index d7bc34636bd..f53a634a390 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -9977,7 +9977,11 @@ vect_transform_loop (loop_vec_info loop_vinfo, gimple
*loop_vectorized_call)
lowest_vf) - 1
: wi::udiv_floor (loop->nb_iterations_upper_bound + bias_for_lowest,
lowest_vf) - 1);
- if (main_vinfo)
+ if (main_vinfo
+ /* Both peeling for alignment and peeling for gaps can end up
+ with the scalar epilogue running for more than VF-1 iterations. */
+ && !main_vinfo->peeling_for_alignment
+ && !main_vinfo->peeling_for_gaps)
{
unsigned int bound;
poly_uint64 main_iters
--
2.34.1
