https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96403
Bug ID: 96403
Summary: [nvptx] Less optimal code in v2si-cvt.c after setting
TARGET_TRULY_NOOP_TRUNCATION to false
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
Target Milestone: ---
[ I've rewritten the v2si-cvt.c source to something more minimal:
...
__v2si __attribute__((unused))
vector_cvt (__v2si arg)
{
unsigned short *p = (unsigned short*)&arg;
volatile unsigned short s = p[0];
return arg;
}
__v2si __attribute__((unused))
vector_cvt_2 (__v2si arg)
{
unsigned char *p = (unsigned char*)&arg;
volatile unsigned char s = p[0];
return arg;
}
...
]
When changing TARGET_TRULY_NOOP_TRUNCATION to false, we have a regression in
v2si-cvt.c, this for vector_cvt:
...
- cvt.u16.u32 %r27, %r25.x;
+ mov.b64 %r26, %r25;
+ cvt.u16.u64 %r27, %r26;
...
and this for vector_cvt_2:
...
- cvt.u32.u32 %r27, %r25.x;
- st.u8 [%frame], %r27;
+ mov.b64 %r26, %r25;
+ cvt.u32.u64 %r27, %r26;
+ cvt.u16.u8 %r32, %r27;
+ mov.u16 %r29, %r32;
+ cvt.u32.u16 %r30, %r29;
+ st.u8 [%frame], %r30;
...
