When we have a pointer relational compare we have stronger guarantees
about overflow, in particular rewriting BASE0 + STEP0 cmp BASE1 + STEP1
as BASE0 + STEP0 - STEP1 cmp BASE1 is always valid and the new IV0
does not overflow. The patch basically reverts the previous change
when pointers are involved, keeping only the more conservative handling
for equality compares which can involve comparing different object
addresses.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2022-01-25 Richard Biener <rguent...@suse.de>
PR tree-optimization/104214
* tree-ssa-loop-niter.cc (number_of_iterations_cond): Use
stronger guarantees for relational pointer compares when
rewriting BASE0 + STEP0 cmp BASE1 + STEP1 as
BASE0 + STEP0 - STEP1 cmp BASE1.
* gcc.dg/vect/pr81196-2.c: New variant testcase only
requiring vect_int.
---
gcc/testsuite/gcc.dg/vect/pr81196-2.c | 16 ++++++++++++++++
gcc/tree-ssa-loop-niter.cc | 15 ++++++++++++---
2 files changed, 28 insertions(+), 3 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/vect/pr81196-2.c
diff --git a/gcc/testsuite/gcc.dg/vect/pr81196-2.c
b/gcc/testsuite/gcc.dg/vect/pr81196-2.c
new file mode 100644
index 00000000000..8d5ce6bad53
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr81196-2.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target vect_int } */
+
+void b (int *p)
+{
+ p = (int *)__builtin_assume_aligned(p, __BIGGEST_ALIGNMENT__);
+ int *q = p + 255;
+ for(; p < q; ++p, --q)
+ {
+ int t = *p;
+ *p = *q;
+ *q = t;
+ }
+}
+
+/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
diff --git a/gcc/tree-ssa-loop-niter.cc b/gcc/tree-ssa-loop-niter.cc
index 04c209561d8..d33095b8e03 100644
--- a/gcc/tree-ssa-loop-niter.cc
+++ b/gcc/tree-ssa-loop-niter.cc
@@ -1915,14 +1915,23 @@ number_of_iterations_cond (class loop *loop,
}
/* If the new step of IV0 has changed sign or is of greater
magnitude then we do not know whether IV0 does overflow
- and thus the transform is not valid for code other than NE_EXPR */
+ and thus the transform is not valid for code other than NE_EXPR. */
else if (tree_int_cst_sign_bit (step) != tree_int_cst_sign_bit
(iv0->step)
|| wi::gtu_p (wi::abs (wi::to_widest (step)),
wi::abs (wi::to_widest (iv0->step))))
{
- if (code != NE_EXPR)
+ if (POINTER_TYPE_P (type) && code != NE_EXPR)
+ /* For relational pointer compares we have further guarantees
+ that the pointers always point to the same object (or one
+ after it) and that objects do not cross the zero page. So
+ not only is the transform always valid for relational
+ pointer compares, we also know the resulting IV does not
+ overflow. */
+ ;
+ else if (code != NE_EXPR)
return false;
- iv0->no_overflow = false;
+ else
+ iv0->no_overflow = false;
}
iv0->step = step;
--
2.31.1
