* Regarding Bug 85539 - x86_64: loads are not always narrowed [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85539]
@ 2019-05-06 6:57 navya deepika Garakapati
0 siblings, 0 replies; 2+ messages in thread
From: navya deepika Garakapati @ 2019-05-06 6:57 UTC (permalink / raw)
To: gcc; +Cc: rguenth
Hi All,
For the below testcase,
$cat test.c
int foo(long *p)
{
return *p;
}
when compile with clang -O2
$clang test.c -O2 -S
$cat test.s
foo:
movl (%rdi), %eax
retq
while gcc gives
$gcc test.c -O2 -S
$cat test.s
foo:
movq (%rdi), %rax
retq
As you can see the difference clang is loading value in eax(32 bit) while
gcc uses rax(64 bit).
While analysing it we found that ,
In gcc trunck with -O0 optimization,
we can notice that in ira pass, there is a conversion information from
64bit to 32bit (subreg:SI (reg:DI 82 [ _1 ]) 0)):
ira pass:
(insn 8 7 11 2 (set (reg:SI 83 [ _4 ])
(subreg:SI (reg:DI 82 [ _1 ]) 0)) "t.c":3:8 67 {*movsi_internal}
(expr_list:REG_DEAD (reg:DI 82 [ _1 ])
(nil)))
while reload pass is deleting this information and assigning 32bit to 32bit
i.e; SI to SI which results this as a dead code for the next pass and as a
result this insn 8 is getting deleted in next pass which is split2 pass.
reload pass:
(insn 8 7 15 2 (set (reg:SI 0 ax [orig:83 _4 ] [83])
(reg:SI 0 ax [orig:82 _1 ] [82])) "t.c":3:8 67 {*movsi_internal}
(nil))
We would like to know why conversion from DI to SI is converted to SI to
SI in reload pass (which uses any target hook).
Thanks & Regards
Navya.
^ permalink raw reply [flat|nested] 2+ messages in thread
* Regarding Bug 85539 - x86_64: loads are not always narrowed [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85539]
@ 2019-05-20 7:36 navya deepika Garakapati
0 siblings, 0 replies; 2+ messages in thread
From: navya deepika Garakapati @ 2019-05-20 7:36 UTC (permalink / raw)
To: rguenth, gcc; +Cc: tilkax
Hi All,
I am looking https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85539 bug and did
some analysis.
please help me on the below query.
For the below testcase,
$cat test.c
int foo(long *p)
{
return *p;
}
when compile with clang -O2
$clang test.c -O2 -S
$cat test.s
foo:
movl (%rdi), %eax
retq
while gcc gives
$gcc test.c -O2 -S
$cat test.s
foo:
movq (%rdi), %rax
retq
As you can see the difference clang is loading value in eax(32 bit) while
gcc uses rax(64 bit).
While analysing it we found that ,
In gcc trunck with -O0 optimization,
we can notice that in ira pass, there is a conversion information from
64bit to 32bit (subreg:SI (reg:DI 82 [ _1 ]) 0)):
ira pass:
(insn 8 7 11 2 (set (reg:SI 83 [ _4 ])
(subreg:SI (reg:DI 82 [ _1 ]) 0)) "t.c":3:8 67 {*movsi_internal}
(expr_list:REG_DEAD (reg:DI 82 [ _1 ])
(nil)))
while reload pass is deleting this information and assigning 32bit to 32bit
i.e; SI to SI which results this as a dead code for the next pass and as a
result this insn 8 is getting deleted in next pass which is split2 pass.
reload pass:
(insn 8 7 15 2 (set (reg:SI 0 ax [orig:83 _4 ] [83])
(reg:SI 0 ax [orig:82 _1 ] [82])) "t.c":3:8 67 {*movsi_internal}
(nil))
We would like to know why conversion from DI to SI is converted to SI to
SI in reload pass (which uses any target hook).
Thanks & Regards
Navya.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2019-05-20 7:36 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-05-06 6:57 Regarding Bug 85539 - x86_64: loads are not always narrowed [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85539] navya deepika Garakapati
2019-05-20 7:36 navya deepika Garakapati
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).