public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug target/110473] New: vec_convert for aarch64 seems to lower to something which should be improved
@ 2023-06-29 4:15 pinskia at gcc dot gnu.org
2023-06-29 7:26 ` [Bug target/110473] " rguenth at gcc dot gnu.org
0 siblings, 1 reply; 2+ messages in thread
From: pinskia at gcc dot gnu.org @ 2023-06-29 4:15 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110473
Bug ID: 110473
Summary: vec_convert for aarch64 seems to lower to something
which should be improved
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: pinskia at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
Take:
```
typedef unsigned int v4si __attribute__ ((vector_size (4*sizeof(int))));
typedef unsigned short v4hi __attribute__ ((vector_size (4*sizeof(short))));
v4si f(v4si a, v4si b)
{
v4hi t = __builtin_convertvector (a, v4hi);
v4si t1 = __builtin_convertvector (t, v4si);
return t1;
}
```
This gets lowered in veclower21 to
```
_6 = BIT_FIELD_REF <_5, 64, 0>;
t_2 = _6;
_7 = BIT_FIELD_REF <t_2, 16, 0>;
_8 = (unsigned int) _7;
_9 = BIT_FIELD_REF <t_2, 16, 16>;
_10 = (unsigned int) _9;
_11 = BIT_FIELD_REF <t_2, 16, 32>;
_12 = (unsigned int) _11;
_13 = BIT_FIELD_REF <t_2, 16, 48>;
_14 = (unsigned int) _13;
t1_3 = {_8, _10, _12, _14};
```
And then forwprop optimizes this to:
```
_6 = BIT_FIELD_REF <_5, 64, 0>;
t1_3 = (v4si) _6;
```
And then combine comes along and optimizes that to:
(insn 9 8 11 2 (set (reg:V8HI 98)
(vec_concat:V8HI (truncate:V4HI (reg:V4SI 102))
(const_vector:V4HI [
(const_int 0 [0]) repeated x4
]))) "/app/example.cpp":7:8 5467
{truncv4siv4hi2_vec_concatz_le}
(expr_list:REG_DEAD (reg:V4SI 102)
(nil)))
(note 11 9 16 2 NOTE_INSN_DELETED)
(insn 16 11 17 2 (set (reg/i:V4SI 32 v0)
(sign_extend:V4SI (subreg:V4HI (reg:V8HI 98) 0)))
"/app/example.cpp":10:1 5459 {extendv4hiv4si2}
(expr_list:REG_DEAD (reg:V8HI 98)
(nil)))
But the first one is basically just (truncate:V4HI (reg:V4SI 102)) (due to the
way the instruction works, the top parts is also zeros).
So we get in the end:
xtn v0.4h, v0.4s
sxtl v0.4s, v0.4h
Why couldn't vectlower could just do:
```
_6 = (v4hi)a_1(D);
t1_3 = (v4si) _6;
```
In the first place instead of depending on later optimizations (at least handle
the second part)?
note with the above lowering we might hit an issue in match.pd where it is
trying to turn that into t1_3 = t1_3 = {0xFFFF,0xFFFF,0xFFFF,0xFFFF} & _6;
(both because of TYPE_PRECISION and because it just uses wide_int_to_tree ...
^ permalink raw reply [flat|nested] 2+ messages in thread
* [Bug target/110473] vec_convert for aarch64 seems to lower to something which should be improved
2023-06-29 4:15 [Bug target/110473] New: vec_convert for aarch64 seems to lower to something which should be improved pinskia at gcc dot gnu.org
@ 2023-06-29 7:26 ` rguenth at gcc dot gnu.org
0 siblings, 0 replies; 2+ messages in thread
From: rguenth at gcc dot gnu.org @ 2023-06-29 7:26 UTC (permalink / raw)
To: gcc-bugs
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110473
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2023-06-29
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
probably nobody implemented this in veclower?
On x86-64 I get
_5 = VEC_PACK_TRUNC_EXPR <a_1(D), { 0, 0, 0, 0 }>;
_6 = BIT_FIELD_REF <_5, 64, 0>;
t_2 = _6;
_7 = [vec_unpack_lo_expr] t_2;
_8 = [vec_unpack_hi_expr] t_2;
t1_3 = {_7, _8};
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2023-06-29 7:26 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-06-29 4:15 [Bug target/110473] New: vec_convert for aarch64 seems to lower to something which should be improved pinskia at gcc dot gnu.org
2023-06-29 7:26 ` [Bug target/110473] " rguenth at gcc dot gnu.org
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).