http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55623
Bug #: 55623
Summary: [ARM] GCC should not prefer long dependency chains,
they inhibit performance on superscalar processors
Classification: Unclassified
Product: gcc
Version: 4.7.2
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: tree-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: siarhei.siamas...@gmail.com
This is a missing optimization. Or in this particular case, it's more like GCC
is reversing an attempt of a programmer to optimize the code for superscalar
dual-issue processors.
$ arm-none-linux-gnueabi-gcc -O2 -mcpu=cortex-a8 -o badsched badsched.c
$ objdump -d badsched
00000000 <f1>:
0: e1a03120 lsr r3, r0, #2
4: e08330a0 add r3, r3, r0, lsr #1
8: e08331a0 add r3, r3, r0, lsr #3
c: e0833220 add r3, r3, r0, lsr #4
10: e08332a0 add r3, r3, r0, lsr #5
14: e0833320 add r3, r3, r0, lsr #6
18: e08333a0 add r3, r3, r0, lsr #7
1c: e0833420 add r3, r3, r0, lsr #8
20: e08334a0 add r3, r3, r0, lsr #9
24: e0833520 add r3, r3, r0, lsr #10
28: e08335a0 add r3, r3, r0, lsr #11
2c: e0833620 add r3, r3, r0, lsr #12
30: e08336a0 add r3, r3, r0, lsr #13
34: e0833720 add r3, r3, r0, lsr #14
38: e08337a0 add r3, r3, r0, lsr #15
3c: e0833820 add r3, r3, r0, lsr #16
40: e08338a0 add r3, r3, r0, lsr #17
44: e0833920 add r3, r3, r0, lsr #18
48: e08339a0 add r3, r3, r0, lsr #19
4c: e0833a20 add r3, r3, r0, lsr #20
50: e0833aa0 add r3, r3, r0, lsr #21
54: e0833b20 add r3, r3, r0, lsr #22
58: e0833ba0 add r3, r3, r0, lsr #23
5c: e0830c20 add r0, r3, r0, lsr #24
60: e12fff1e bx lr
00000064 <f2>:
64: e1a031a0 lsr r3, r0, #3
68: e1a02220 lsr r2, r0, #4
6c: e08330a0 add r3, r3, r0, lsr #1
70: e0822120 add r2, r2, r0, lsr #2
74: e08332a0 add r3, r3, r0, lsr #5
78: e0822320 add r2, r2, r0, lsr #6
7c: e08333a0 add r3, r3, r0, lsr #7
80: e0822420 add r2, r2, r0, lsr #8
84: e08334a0 add r3, r3, r0, lsr #9
88: e0822520 add r2, r2, r0, lsr #10
8c: e08335a0 add r3, r3, r0, lsr #11
90: e0822620 add r2, r2, r0, lsr #12
94: e08336a0 add r3, r3, r0, lsr #13
98: e0822720 add r2, r2, r0, lsr #14
9c: e08337a0 add r3, r3, r0, lsr #15
a0: e0822820 add r2, r2, r0, lsr #16
a4: e08338a0 add r3, r3, r0, lsr #17
a8: e0822920 add r2, r2, r0, lsr #18
ac: e08339a0 add r3, r3, r0, lsr #19
b0: e0822a20 add r2, r2, r0, lsr #20
b4: e0833aa0 add r3, r3, r0, lsr #21
b8: e0822b20 add r2, r2, r0, lsr #22
bc: e0833ba0 add r3, r3, r0, lsr #23
c0: e0820c20 add r0, r2, r0, lsr #24
c4: e0800003 add r0, r0, r3
c8: e12fff1e bx lr
Guess which one of these two functions will be faster?
=== Cortex-A8 @1000MHz ===
$ time ./badsched 1
real 0m2.512s
user 0m2.500s
sys 0m0.000s
$ time ./badsched 2
real 0m2.064s
user 0m2.008s
sys 0m0.008s
=== Cortex-A15 @1700MHz ===
real 0m2.786s
user 0m2.770s
sys 0m0.005s
real 0m1.451s
user 0m1.440s
sys 0m0.005s
There is a function call and loop overhead which prevents Cortex-A8 from
showing ~2x better performance in the case of using "f2" function. We can try
to mark these function as static in order to get them inlined, but in this case
the asm workaround becomes ineffective in a rather interesting way, which also
demonstrates instructions scheduling issues.
