https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103571
Bug ID: 103571
Summary: ABI: V2HF, V4HF and V8HFmode argument passing issues
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ubizjak at gmail dot com
Target Milestone: ---
Following testcase:
--cut here--
typedef _Float16 v2hf __attribute__((vector_size(4)));
typedef _Float16 v4hf __attribute__((vector_size(8)));
typedef _Float16 v8hf __attribute__((vector_size(16)));
v2hf foo (v2hf a, v2hf b)
{
return b;
}
v4hf bar (v4hf a, v4hf b)
{
return b;
}
v8hf baz (v8hf a, v8hf b)
{
return b;
}
--cut here--
compiles with -O2 -msse2 -m64 to:
foo:
movl 16(%rsp), %edx # 6 [c=9 l=4] *movsi_internal/0
movq %rdi, %rax # 2 [c=4 l=3] *movdi_internal/3
movl %edx, (%rdi) # 7 [c=4 l=2] *movsi_internal/1
ret # 18 [c=0 l=1] simple_return_internal
and with -O2 -msse2 -m32 to:
foo:
movl 4(%esp), %eax # 2 [c=9 l=4] *movsi_internal/0
movl 12(%esp), %edx # 6 [c=9 l=4] *movsi_internal/0
movl %edx, (%eax) # 7 [c=4 l=2] *movsi_internal/1
ret $4 # 17 [c=0 l=3] simple_return_pop_internal
bar:
movq %mm1, %mm0 # 14 [c=4 l=3] *movv4hf_internal/6
ret # 18 [c=0 l=1] simple_return_internal
baz:
pushl %esi # 53 [c=4 l=1] *pushsi2/0
pushl %ebx # 54 [c=4 l=1] *pushsi2/0
subl $52, %esp # 55 [c=4 l=3]
movaps %xmm1, 16(%esp) # 5 [c=4 l=5] movv8hf_internal/3
movl 20(%esp), %ecx # 34 [c=9 l=4] *movsi_internal/0
movl 24(%esp), %edx # 35 [c=9 l=4] *movsi_internal/0
movl 28(%esp), %eax # 36 [c=9 l=4] *movsi_internal/0
movd %xmm1, (%esp) # 46 [c=4 l=5] *movsi_internal/11
movl %ecx, 4(%esp) # 47 [c=4 l=4] *movsi_internal/1
movl %edx, 8(%esp) # 48 [c=4 l=4] *movsi_internal/1
movl %eax, 12(%esp) # 49 [c=4 l=4] *movsi_internal/1
movdqa (%esp), %xmm0 # 50 [c=17 l=5] *movti_internal/4
addl $52, %esp # 58 [c=4 l=3]
popl %ebx # 59 [c=9 l=1] *popsi1
popl %esi # 60 [c=9 l=1] *popsi1
ret # 61 [c=0 l=1] simple_return_internal
Does ABI specify how to handle V2HF arguments and returns? foo looks a bit
suspicious to me, corresponding V2HI arguments are simply returned in %eax
register.
Also, baz iz highly un-optimal for 32bit targets.
