On 22/01/2026 16:51, Artemiy Volkov wrote:
On Thu, Jan 22, 2026 at 04:05:36PM +0000, Richard Earnshaw wrote:
On 14/01/2026 14:20, Artemiy Volkov wrote:
This is a v2 of
https://gcc.gnu.org/pipermail/gcc-patches/2026-January/705195.html with
feedback from Richard and Christophe addressed.
The check_effective_target_arm_v8_3a_fp16_complex_neon_ok_nocache
procedure in target-supports.exp should return a set of flags providing
all of AdvSIMD, complex numbers, and fp16 capabilities. However, to
achieve this we have to be able to overwrite the current -mfpu setting.
This means that the current empty string alternative for $flags does not
work for us. This patch splits sets of options to iterate over between
the arm and the aarch64 backend and prohibits an empty string for the
former. (This split of option sets allows us to remove the redundant
"+simd" from -march on aarch64, as simd is present on all base
architectures all the way from armv8-a.) Moreover, it changes
check_effective_target_arm_v8_3a_complex_neon_ok_nocache in a similar way,
except offering an empty string as the first alternative to try testing
with the flags that gcc was configured with.
This (together with patch 1/2) fixes the
advsimd-intrinsics/vector-complex_f16.c testcase, which was previously
being compiled with the -mfpu=neon-fp16 flag added by the .exp file
itself.
Re-tested on aarch64 by me and on arm by Christophe.
Changes from v1:
- Add "" and "-mfpu=auto" as the first alternative
in check_effective_target_arm_v8_3a{,_fp16}_complex_neon_ok,
respectively.
- Clarify the changes to -march values in the aarch64 case in the commit
message.
Co-authored-by: Richard Earnshaw <[email protected]>
gcc/testsuite/ChangeLog:
* lib/target-supports.exp:
(check_effective_target_arm_v8_3a_complex_neon_ok_nocache):
Split and fill in arm and aarch64 compile options. Remove the
cpu_unset variable.
(check_effective_target_arm_v8_3a_fp16_complex_neon_ok_nocache):
Likewise.
---
gcc/testsuite/lib/target-supports.exp | 33 ++++++++++++++++++---------
1 file changed, 22 insertions(+), 11 deletions(-)
diff --git a/gcc/testsuite/lib/target-supports.exp
b/gcc/testsuite/lib/target-supports.exp
index dbcba42629f..77399a04bc5 100644
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -13717,27 +13717,33 @@ proc check_effective_target_inff { } {
proc check_effective_target_arm_v8_3a_complex_neon_ok_nocache { } {
global et_arm_v8_3a_complex_neon_flags
set et_arm_v8_3a_complex_neon_flags ""
- set cpu_unset ""
if { ![istarget arm*-*-*] && ![istarget aarch64*-*-*] } {
return 0;
}
if { [istarget arm*-*-*] } {
- set cpu_unset "-mcpu=unset"
+ set flag_opts {
+ ""
I don't see how it can be right that we don't need -mcpu=auto here...
+ "-mcpu=unset -march=armv8.3-a+simd"
+ "-mfloat-abi=softfp -mfpu=auto -mcpu=unset -march=armv8.3-a+simd"
+ "-mfloat-abi=hard -mfpu=auto -mcpu=unset -march=armv8.3-a+simd"
+ }
+ } else {
+ set flag_opts { "" "-march=armv8.3-a" }
}
# Iterate through sets of options to find the compiler flags that
# need to be added to the -march option.
- foreach flags {"" "-mfloat-abi=softfp -mfpu=auto" "-mfloat-abi=hard
-mfpu=auto"} {
+ foreach flags $flag_opts {
if { [check_no_compiler_messages_nocache \
arm_v8_3a_complex_neon_ok assembly {
#if !defined (__ARM_FEATURE_COMPLEX)
#error "__ARM_FEATURE_COMPLEX not defined"
#endif
#include <complex.h>
- } "$flags $cpu_unset -march=armv8.3-a+simd"] } {
- set et_arm_v8_3a_complex_neon_flags "$flags $cpu_unset
-march=armv8.3-a+simd"
+ } "$flags"] } {
+ set et_arm_v8_3a_complex_neon_flags "$flags"
return 1;
}
}
@@ -13765,19 +13771,25 @@ proc add_options_for_arm_v8_3a_complex_neon { flags }
{
proc check_effective_target_arm_v8_3a_fp16_complex_neon_ok_nocache { } {
global et_arm_v8_3a_fp16_complex_neon_flags
set et_arm_v8_3a_fp16_complex_neon_flags ""
- set cpu_unset ""
if { ![istarget arm*-*-*] && ![istarget aarch64*-*-*] } {
return 0;
}
if { [istarget arm*-*-*] } {
- set cpu_unset "-mcpu=unset"
+ set flag_opts {
+ "-mfpu=auto"
... but do here. This implies something else is wrong. FWIW, I think the
first case is what we really want, though there may be other bugs in some
tests that need fixing as well.
The issue that "-mfpu=auto" is working around occurs at the intersection of
vector, complex, and fp16 capabilities. This combination can not be
achieved for just any value of -mfpu, but rather only for those that
satisfy TARGET_VFP5. In particular, when advsimd-intrinsics.exp injects
-mfpu=neon-fp-armv8 via code on ll. 57-63, there is nothing an individual
test can do short of resetting the value by doing -mfpu=auto before
trying any other options. This is not a concern for the first check as
things work out fine for non-fp16 complex vectorization, and an empty
string is sufficient.
No, the issue is that the dg- headers in vector-complex_f16.c are
completely broken. You can't ask the framework to calculate the flags
needed for adding complex, then ask it to calculate the flags needed for
adding fp16 and then finally try to merge the two; it simply won't work
that way - these are two separate calculations that both start from the
base flags, and what's more the results are cached for use across other
tests. Even worse, you can't then start trying to override the flags
that have been calculated with something else - especially when you want
to make that work in both aarch32 and aarch64.
So, instead of
/* { dg-require-effective-target arm_v8_3a_complex_neon_ok } */
/* { dg-require-effective-target arm_v8_2a_fp16_neon_ok } */
/* { dg-add-options arm_v8_3a_complex_neon } */
/* { dg-add-options arm_v8_2a_fp16_neon } */
/* { dg-additional-options "-O2 -march=armv8.3-a+fp16 -save-temps" } */
that test should be doing:
/* { dg-require-effective-target arm_v8_3a_fp16_complex_neon_ok } */
/* { dg-add-options arm_v8_3a_fp16_complex_neon } */
/* { dg-additional-options "-O2 -save-temps" } */
I'd be inclined to push a change like that directly, but I'm not
entirely sure if Tamar had something else in mind when he originally
wrote the patch, so I'd like to give him a chance to comment first.
These things are bound to surface every once in a while until the pattern
"find a set of flags to satisfy a target requirement and cache it, then
add a bunch of other, sometimes conflicting, flags in the .exp file, then
append the cached result" persists in the testsuite. This patch does not
try to address this situation, but rather minimize the workaround for one
specific test, and as such is IMO a net improvement.
Thanks,
Artemiy
+ "-mfpu=auto -mcpu=unset -march=armv8.3-a+fp16+simd"
+ "-mfloat-abi=softfp -mfpu=auto -mcpu=unset
-march=armv8.3-a+fp16+simd"
+ "-mfloat-abi=hard -mfpu=auto -mcpu=unset -march=armv8.3-a+fp16+simd"
+ }
+ } else {
+ set flag_opts { "" "-march=armv8.3-a+fp16" }
}
# Iterate through sets of options to find the compiler flags that
# need to be added to the -march option.
- foreach flags {"" "-mfloat-abi=softfp -mfpu=auto" "-mfloat-abi=hard
-mfpu=auto"} {
+ foreach flags $flag_opts {
if { [check_no_compiler_messages_nocache \
arm_v8_3a_fp16_complex_neon_ok assembly {
#if !defined (__ARM_FEATURE_COMPLEX)
@@ -13787,9 +13799,8 @@ proc
check_effective_target_arm_v8_3a_fp16_complex_neon_ok_nocache { } {
#error "__ARM_FEATURE_FP16_VECTOR_ARITHMETIC not defined"
#endif
#include <complex.h>
- } "$flags $cpu_unset -march=armv8.3-a+fp16+simd"] } {
- set et_arm_v8_3a_fp16_complex_neon_flags \
- "$flags $cpu_unset -march=armv8.3-a+fp16+simd"
+ } "$flags"] } {
+ set et_arm_v8_3a_fp16_complex_neon_flags "$flags"
return 1;
}
}
R.
R.
