Hi!
On 10/4/26 19:14, Torbjörn Svensson via Sourceware Forge wrote:
This is a summary of discussions relative to the merge request created by Torbjörn
Svensson (azoff) <[email protected]> titled
arm: Set Cortex-M55 tuning for MVE FP tests
since its creation.
Description: Changes since v1:
- Dropped the generic change in lib/target-supports.exp
- Added the change to the dg-options/dg-additional-options for the affected
tests.
--
When running DejaGNU with thumb/arch=a/tune=b and a does not happen to be
armv8.1-m.main+mve.fp, the effective target check will switch target to that. If that
happens, the tune "b" will survive and impact what assembler that will be
generated.
This change resets the tune to the default one for armv8.1-m.main+mve.fp.
Tested on arm-none-eabi with both -mcpu and -mtune on r17-4804-gf7094fe5ff882f
without any regressions.
Ok for trunk, releases/gcc-16 and releases/gcc-15?
--
CC: Richard Earnshaw <[email protected]>
CC: Andre Simoes Dias Vieira <[email protected]>
CC: Christophe Lyon <[email protected]>
The full and up to date discussion can be found at
https://forge.sourceware.org/gcc/gcc/pulls/238
The merge request has been closed without being merged directly on the forge
repository.
On 2026-10-04 07:03:32+00:00, Torbjörn Svensson (azoff)
<[email protected]> requested that Christophe Lyon (clyon)
<[email protected]> review the code:
On 2026-10-04 07:31:07+00:00, Richard Earnshaw (rearnsha) <[email protected]>
commented on the code:
+++ gcc/testsuite/lib/target-supports.exp
@@ -7035,2 +7035,3 @@
# need to be added to the -march option.
- foreach flags {"" "-mfloat-abi=softfp -mfpu=auto -mcpu=unset
-march=armv8.1-m.main+mve.fp" "-mfloat-abi=hard -mfpu=auto -mcpu=unset
-march=armv8.1-m.main+mve.fp"} {
+ foreach flags {
+ ""
Why would this alternative not also override the tuning? What if we have a CPU
with the same architecture but different tuning costs?
Which tests are affected by this issue? Would it make sense to override the
tuning there as needed?
+++ gcc/testsuite/lib/target-supports.exp
@@ -7035,2 +7035,3 @@
# need to be added to the -march option.
- foreach flags {"" "-mfloat-abi=softfp -mfpu=auto -mcpu=unset
-march=armv8.1-m.main+mve.fp" "-mfloat-abi=hard -mfpu=auto -mcpu=unset
-march=armv8.1-m.main+mve.fp"} {
+ foreach flags {
+ ""
The tests fixed are:
- FAIL: gcc.target/arm/mve/intrinsics/vcmpeqq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpeqq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpgeq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpgeq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpgtq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpgtq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpleq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpleq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpltq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpltq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpneq_n_f16.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vcmpneq_n_f32.c check-function-bodies foo2
- FAIL: gcc.target/arm/mve/intrinsics/vdupq_n_f16.c check-function-bodies foo1
- FAIL: gcc.target/arm/mve/intrinsics/vdupq_n_f32.c check-function-bodies foo1
All of these were failing with
`thumb/arch=armv7ve+simd/tune=cortex-a7/float-abi=hard/fpu=auto` and
`thumb/arch=armv7ve+nofp/tune=cortex-a7/float-abi=soft/fpu=auto`, while passing
when replacing `-mtune` with `-mcpu`.
The reason why I did not add the `-mtune=cortex-m55` for the empty string is
that I thought it was only needed when switching target.
I can add it there too in case you think that is better (should have no impact
regardless).
Comparing -mtune with -mcpu for gcc.target/arm/mve/intrinsics/vcmpeqq_n_f16.c,
I see the following difference in generated assembler:
```
--- mcpu.s 2026-10-04 09:49:39.835214873 +0200
+++ mtune.s 2026-10-04 09:49:39.836214873 +0200
@@ -66,11 +66,23 @@
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
- vmov d6, r0, r1 @ v8hf
- vmov d7, r2, r3
- movw r3, #15462
- vcmp.f16 eq, q3, r3
+ vmov d4, r0, r1 @ v8hf
+ vmov d5, r2, r3
+ vldr.64 d6, .L5
+ vldr.64 d7, .L5+8
+ vcmp.f16 eq, q2, q3
vmrs r0, p0 @ movhi
uxth r0, r0
bx lr
+.L6:
+ .align 3
+.L5:
+ .short 15462
+ .short 15462
+ .short 15462
+ .short 15462
+ .short 15462
+ .short 15462
+ .short 15462
+ .short 15462
.size foo2, .-foo2
```
Sorry it's not clear to me which code is with -mtune and which one is
with -mcpu?
and do you mean -mtune=cortexx-a7 / -mcpu=cortex-a7?
Also I'm not sure I understand the fix ;-)
IIUC without your patch the test fails because it's compiled with
thumb/arch=armv7ve+simd/tune=cortex-a7/float-abi=hard/fpu=auto followed
by "-mfloat-abi=hard -mfpu=auto -mcpu=unset
-march=armv8.1-m.main+mve.fp" from arm_v8_1m_mve_fp_ok_nocache?
which gets simplified to
-mthumb -march=armv8.1-m.main+mve.fp -mtune=cortex-a7
-mfloat-abi=hard -mfpu=auto -mcpu=unset
and -mcpu does not cancel the effect of -mtune?
Thanks,
Christophe
If you have a suggestion how to fix the tests in a better way, I'd be happy to
try it.
On 2026-10-04 11:52:44+00:00, Torbjörn Svensson (azoff) wrote:
Updated the patch after live discussion with Richard E.
On 2026-10-04 11:54:30+00:00, Richard Earnshaw (rearnsha) <[email protected]>
approved the changes:
OK
On 2026-10-04 17:13:16+00:00, Torbjörn Svensson (azoff) wrote:
Pushed as r17-4873-g8a5c76ca173f25, r16-9734-ga239230fc45aa9 and
r15-11613-gb56c7709b773a2.
