https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120941
Filip Kastl <pheeck at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |NEW
--- Comment #17 from Filip Kastl <pheeck at gcc dot gnu.org> ---
This is the replacement that causes the slowdown (well, two replacements):
----------------------
Replace:
(insn 2224 2222 2228 20 (set (reg:V4DF 1604)
(vec_duplicate:V4DF (mem/u/c:DF (symbol_ref/u:DI ("*.LC3") [flags 0x2])
[0 S8 A64]))) 9260 {vec_dupv4df}
(expr_list:REG_EQUAL (const_vector:V4DF [
(const_double:DF
2.7777777777777776235801354687282582744956016540527344e-2
[0x0.e38e38e38e38ep-5]) repeated x4
])
(nil)))
with:
(insn 2224 2222 2228 20 (set (reg:V4DF 1604)
(reg:V4DF 1655)) 2428 {movv4df_internal}
(expr_list:REG_EQUAL (const_vector:V4DF [
(const_double:DF
2.7777777777777776235801354687282582744956016540527344e-2
[0x0.e38e38e38e38ep-5]) repeated x4
])
(nil)))
deferring rescan insn with uid = 2224.
Replace:
(insn 2228 2224 377 20 (set (reg:V2DF 1603)
(vec_duplicate:V2DF (mem/u/c:DF (symbol_ref/u:DI ("*.LC3") [flags 0x2])
[0 S8 A64]))) 7168 {vec_dupv2df}
(expr_list:REG_EQUAL (const_vector:V2DF [
(const_double:DF
2.7777777777777776235801354687282582744956016540527344e-2
[0x0.e38e38e38e38ep-5]) repeated x2
])
(nil)))
with:
(insn 2228 2224 377 20 (set (reg:V2DF 1603)
(subreg:V2DF (reg:V4DF 1655) 0)) 2429 {movv2df_internal}
(expr_list:REG_EQUAL (const_vector:V2DF [
(const_double:DF
2.7777777777777776235801354687282582744956016540527344e-2
[0x0.e38e38e38e38ep-5]) repeated x2
])
(nil)))
deferring rescan insn with uid = 2228.
----------------------
These instructions are inside function "main". Though, the last RTL debug
instruction is
(debug_insn 272 271 273 19 (debug_marker) "lbm.c":275:2 discrim 1 -1
(nil))
so I expect that function "LBM_performStreamCollideTRT" was inlined into main
and is the original source of these vector instructions.
Hopefully this helps. If you meant something else by "testcase", do tell me.
What I did in more detail:
I used a custom debug counter. If I set the 9-th call of
ix86_broadcast_inner() to return null (I adapted what hjl's patch does), I get
rid of the slowdown.
On r16-1644-gaba3b9d3a48a07 I added the debug counter and did:
/home/fkastl/gcc/inst/bin/gcc -std=gnu99 -m64 -c -o lbm.o -DSPEC -DNDEBUG
-DSPEC_AUTO_SUPPRESS_OPENMP -Ofast -march=native -mtune=native -g -flto=32
-fpermissive -std=gnu17 -DSPEC_LP64 lbm.c
-fdbg-cnt=foo_counter:1000000000-1000000000
/home/fkastl/gcc/inst/bin/gcc -std=gnu99 -m64 -c -o main.o -DSPEC -DNDEBUG
-DSPEC_AUTO_SUPPRESS_OPENMP -Ofast -march=native -mtune=native -g -flto=32
-fpermissive -std=gnu17 -DSPEC_LP64 main.c
-fdbg-cnt=foo_counter:1000000000-1000000000
/home/fkastl/gcc/inst/bin/gcc -std=gnu99 -m64
-Wl,-rpath,/home/fkastl/gcc/inst/lib64 -Ofast -march=native -mtune=native -g
-flto=32 -fpermissive -std=gnu17 lbm.o main.o -lm -o lbm_r
-fdbg-cnt=foo_counter:9-9 -fdump-rtl-all
-> 3m43s
/home/fkastl/gcc/inst/bin/gcc -std=gnu99 -m64
-Wl,-rpath,/home/fkastl/gcc/inst/lib64 -Ofast -march=native -mtune=native -g
-flto=32 -fpermissive -std=gnu17 lbm.o main.o -lm -o lbm_r
-fdbg-cnt=foo_counter:1000000000-1000000000 -fdump-rtl-all
-> 2m50s
Then I compared the *.rrvl rtl dumps. Btw I had to "backport" the
Replace:
...
with:
and
Add:
...
dumping from a newer commit.
