https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109697
Bug ID: 109697
Summary: arm: lack of MVE instruction costing causing worse
codegen on a vec_duplicate
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: stammark at gcc dot gnu.org
Target Milestone: ---
Hi all,
In the arm backend, for MVE targets we previously had this bug on the vcmp
patterns: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107987
The fix is fine, but it resulted in some failing tests:
* gcc.target/arm/mve/intrinsics/vcmpcsq_n_u16.c
* gcc.target/arm/mve/intrinsics/vcmpcsq_n_u32.c
* gcc.target/arm/mve/intrinsics/vcmpcsq_n_u8.c
* gcc.target/arm/mve/intrinsics/vcmpeqq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpeqq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmpeqq_n_u16.c
* gcc.target/arm/mve/intrinsics/vcmpeqq_n_u32.c
* gcc.target/arm/mve/intrinsics/vcmpeqq_n_u8.c
* gcc.target/arm/mve/intrinsics/vcmpgeq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpgeq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmpgtq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpgtq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmphiq_n_u16.c
* gcc.target/arm/mve/intrinsics/vcmphiq_n_u32.c
* gcc.target/arm/mve/intrinsics/vcmphiq_n_u8.c
* gcc.target/arm/mve/intrinsics/vcmpleq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpleq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmpltq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpltq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmpneq_n_f16.c
* gcc.target/arm/mve/intrinsics/vcmpneq_n_f32.c
* gcc.target/arm/mve/intrinsics/vcmpneq_n_u16.c
* gcc.target/arm/mve/intrinsics/vcmpneq_n_u32.c
* gcc.target/arm/mve/intrinsics/vcmpneq_n_u8.c
(after Andrea improved these tests in GCC13)
The testcases that are failing are the ones that compare against a scalar
immediate (e.g. "vcmpeqq (a, 1.1)"), because the compiler prefers to do:
```
vldr.64 d6, .L5
vldr.64 d7, .L5+8
vcmp.f16 eq, q0, q3
```
When previously we would much more simply:
```
movs r3, #1
vcmp.u16 cs, q0, r3
```
The underlying reason for this change is a known deficiency of the MVE
implementation: the lack of proper instruction costing.
The compiler falls back to calculating costs based on the operands and the new
vec_duplicate in the patterns (mve_vcmp<mve_cmp_op>q_n_<mode>, etc) gets given
a cost of 32 (when instead it should know that the vec duplicate is free and
this is all just one instruction...), so the "literal load + vector-vector
compare" wins out against the "put the immediate in a GP reg + vector-scalar
compare".
For now, I plan on simply XFAIL-ing the tests.
