Adjust 'libgomp.c/declare-variant-{3,4}-[...]' for inter-procedural value range propagation (was: Propagate value ranges of return values)
Checks
Commit Message
Hi!
On 2023-11-19T16:05:42+0100, Jan Hubicka <hubicka@ucw.cz> wrote:
> this is updated version which also adds testuiste compensation
> I lost earlier while maintaining the patch in my testing tree.
> There are quite few testcases that use constant return values to hide
> something from optimizer.
One more: commit a53da3a213ee00866d132c228a0e89bd2f61d65c
"Adjust 'libgomp.c/declare-variant-{3,4}-[...]' for inter-procedural value range propagation"
pushed to master branch, see attached. (Those regressions are only
visible in GCC offloading configurations.) (And actually, all those test
cases have other issues; will install further patches later on.)
Jakub, Tobias, please let me know if it's not expected that *all* the
"variant" functions have to be tagged '__attribute__ ((noipa))' (as I've
done); just tagging the "dispatcher" function 'f' isn't sufficient.
Grüße
Thomas
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
From a53da3a213ee00866d132c228a0e89bd2f61d65c Mon Sep 17 00:00:00 2001
From: Thomas Schwinge <thomas@codesourcery.com>
Date: Tue, 21 Nov 2023 22:42:49 +0100
Subject: [PATCH] Adjust 'libgomp.c/declare-variant-{3,4}-[...]' for
inter-procedural value range propagation
..., that is, commit 53ba8d669550d3a1f809048428b97ca607f95cf5
"inter-procedural value range propagation", after which we see:
[-PASS:-]{+FAIL:+} libgomp.c/declare-variant-3-sm30.c scan-nvptx-none-offload-tree-dump optimized "= f30 \\(\\);"
Etc. That's due to:
@@ -144,13 +144,11 @@
__attribute__((omp target entrypoint, noclone))
void main._omp_fn.0 (const struct .omp_data_t.3 & restrict .omp_data_i)
{
- int _3;
int * _5;
<bb 2> [local count: 1073741824]:
- _3 = f30 ();
_5 = *.omp_data_i_4(D).v;
- *_5 = _3;
+ *_5 = 30;
return;
It's nice to see this optimization work here, too, but it does interfere with
how we're currently testing OpenMP 'declare variant'.
libgomp/
* testsuite/libgomp.c/declare-variant-3.h (f30, f35, f53, f70)
(f75, f80, f): Add '__attribute__ ((noipa))'.
* testsuite/libgomp.c/declare-variant-4.h (gfx803, gfx900, gfx906)
(gfx908, gfx90a, f): Likewise.
---
libgomp/testsuite/libgomp.c/declare-variant-3.h | 8 ++++++++
libgomp/testsuite/libgomp.c/declare-variant-4.h | 7 +++++++
2 files changed, 15 insertions(+)
@@ -1,34 +1,41 @@
#pragma omp declare target
+
+__attribute__ ((noipa))
int
f30 (void)
{
return 30;
}
+__attribute__ ((noipa))
int
f35 (void)
{
return 35;
}
+__attribute__ ((noipa))
int
f53 (void)
{
return 53;
}
+__attribute__ ((noipa))
int
f70 (void)
{
return 70;
}
+__attribute__ ((noipa))
int
f75 (void)
{
return 75;
}
+__attribute__ ((noipa))
int
f80 (void)
{
@@ -41,6 +48,7 @@ f80 (void)
#pragma omp declare variant (f70) match (device={isa("sm_70")})
#pragma omp declare variant (f75) match (device={isa("sm_75")})
#pragma omp declare variant (f80) match (device={isa("sm_80")})
+__attribute__ ((noipa))
int
f (void)
{
@@ -1,28 +1,34 @@
#pragma omp declare target
+
+__attribute__ ((noipa))
int
gfx803 (void)
{
return 0x803;
}
+__attribute__ ((noipa))
int
gfx900 (void)
{
return 0x900;
}
+__attribute__ ((noipa))
int
gfx906 (void)
{
return 0x906;
}
+__attribute__ ((noipa))
int
gfx908 (void)
{
return 0x908;
}
+__attribute__ ((noipa))
int
gfx90a (void)
{
@@ -38,6 +44,7 @@ gfx90a (void)
#pragma omp declare variant(gfx906) match(device = {isa("gfx906")})
#pragma omp declare variant(gfx908) match(device = {isa("gfx908")})
#pragma omp declare variant(gfx90a) match(device = {isa("gfx90a")})
+__attribute__ ((noipa))
int
f (void)
{
--
2.34.1