Among other things, this testcase tests an addition of the four
values (n*4+[0:3])*9//2 for each n in [0:99]. The addition is
done in multiple integer and floating-point types and the test
is compiled with -ffast-math.
One of the floating-point types is _Float16, and as Andrew says
in the PR, _Float16's limited precision means that the order of the
additions begins to matter for higher n. Specifically, some orders
begin to give different results from others at n=38, and at many
higher n as well.
This patch uses 5/3 rather than 9/2. I tested locally that
all addition orders give the same result over the test range.
Tested on aarch64-linux-gnu & pushed.
Richard
gcc/testsuite/
PR testsuite/113965
* gcc.target/aarch64/sve/mask_struct_load_3_run.c: Use an
input range that is suitable for _Float16.
---
gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_3_run.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_3_run.c
b/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_3_run.c
index 8bc3b08fcf4..c0a7416cfaf 100644
--- a/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_3_run.c
+++ b/gcc/testsuite/gcc.target/aarch64/sve/mask_struct_load_3_run.c
@@ -18,7 +18,7 @@
asm volatile ("" ::: "memory"); \
} \
for (int i = 0; i < N * 4; ++i) \
- in[i] = i * 9 / 2; \
+ in[i] = i * 5 / 3; \
NAME##_4 (out, in, mask, N); \
for (int i = 0; i < N; ++i) \
{ \
--
2.25.1
