[PATCH] tree-ssa-loop-prefetch.cc: Honour -fno-unroll-loops

[PATCH] tree-ssa-loop-prefetch.cc: Honour -fno-unroll-loops

[Stefan Schulze Frielinghaus]

On s390 the following tests fail

FAIL: gcc.dg/vect/pr109011-1.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CLZ \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-1.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-1.c scan-tree-dump-times optimized " = .CLZ \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-1.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-2.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
FAIL: gcc.dg/vect/pr109011-2.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-2.c scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
FAIL: gcc.dg/vect/pr109011-2.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-4.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
FAIL: gcc.dg/vect/pr109011-4.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
FAIL: gcc.dg/vect/pr109011-4.c scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
FAIL: gcc.dg/vect/pr109011-4.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1

because aprefetch unrolls loops even if -fno-unroll-loops is used.
Accordingly, the scan patterns match more than one time.

Could also be fixed by using -fno-prefetch-loop-arrays for the tests.
Though, I tend to prefer if aprefetch honours -fno-unroll-loops.  Any
preferences?

Bootstrapped and regtested on x86_64 and s390.  Ok for mainline?

gcc/ChangeLog:

	* tree-ssa-loop-prefetch.cc (determine_unroll_factor): Honour
	-fno-unroll-loops.
---
 gcc/tree-ssa-loop-prefetch.cc | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/gcc/tree-ssa-loop-prefetch.cc b/gcc/tree-ssa-loop-prefetch.cc
index 70073cc4fe4..bb5d5dec779 100644
--- a/gcc/tree-ssa-loop-prefetch.cc
+++ b/gcc/tree-ssa-loop-prefetch.cc
@@ -1401,6 +1401,10 @@ determine_unroll_factor (class loop *loop, struct mem_ref_group *refs,
   struct mem_ref_group *agp;
   struct mem_ref *ref;
 
+  /* Bail out early in case we must not unroll loops.  */
+  if (!flag_unroll_loops)
+    return 1;
+
   /* First check whether the loop is not too large to unroll.  We ignore
      PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
      from unrolling them enough to make exactly one cache line covered by each
-- 
2.44.0

Re: [PATCH] tree-ssa-loop-prefetch.cc: Honour -fno-unroll-loops

[Richard Biener]

On Wed, May 8, 2024 at 9:56 AM Stefan Schulze Frielinghaus
<stefansf@linux.ibm.com> wrote:
>
> On s390 the following tests fail
>
> FAIL: gcc.dg/vect/pr109011-1.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CLZ \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-1.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-1.c scan-tree-dump-times optimized " = .CLZ \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-1.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-2.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
> FAIL: gcc.dg/vect/pr109011-2.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-2.c scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
> FAIL: gcc.dg/vect/pr109011-2.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-4.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
> FAIL: gcc.dg/vect/pr109011-4.c -flto -ffat-lto-objects  scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
> FAIL: gcc.dg/vect/pr109011-4.c scan-tree-dump-times optimized " = .CTZ \\\\(vect" 2
> FAIL: gcc.dg/vect/pr109011-4.c scan-tree-dump-times optimized " = .POPCOUNT \\\\(vect" 1
>
> because aprefetch unrolls loops even if -fno-unroll-loops is used.
> Accordingly, the scan patterns match more than one time.
>
> Could also be fixed by using -fno-prefetch-loop-arrays for the tests.
> Though, I tend to prefer if aprefetch honours -fno-unroll-loops.  Any
> preferences?
>
> Bootstrapped and regtested on x86_64 and s390.  Ok for mainline?

OK.

Richard.

> gcc/ChangeLog:
>
>         * tree-ssa-loop-prefetch.cc (determine_unroll_factor): Honour
>         -fno-unroll-loops.
> ---
>  gcc/tree-ssa-loop-prefetch.cc | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/gcc/tree-ssa-loop-prefetch.cc b/gcc/tree-ssa-loop-prefetch.cc
> index 70073cc4fe4..bb5d5dec779 100644
> --- a/gcc/tree-ssa-loop-prefetch.cc
> +++ b/gcc/tree-ssa-loop-prefetch.cc
> @@ -1401,6 +1401,10 @@ determine_unroll_factor (class loop *loop, struct mem_ref_group *refs,
>    struct mem_ref_group *agp;
>    struct mem_ref *ref;
>
> +  /* Bail out early in case we must not unroll loops.  */
> +  if (!flag_unroll_loops)
> +    return 1;
> +
>    /* First check whether the loop is not too large to unroll.  We ignore
>       PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
>       from unrolling them enough to make exactly one cache line covered by each
> --
> 2.44.0
>