For this code:
void f(int *__restrict t, const int *__restrict s) {
unsigned long i;
for (i = 0; i < 1024; i++)
t[i] = s[i];
}
gcc-3.3 produces with -O3 -funroll-loops:
0000000000000000 <f>:
0: 17 04 ff 47 clr t9
4: ff 03 1f 23 lda t10,1023
8: 1f 04 ff 47 nop
c: 00 00 fe 2f unop
10: 00 00 51 a0 ldl t1,0(a1)
14: 04 00 91 a0 ldl t3,4(a1)
18: 08 00 b1 a0 ldl t4,8(a1)
1c: 0c 00 d1 a0 ldl t5,12(a1)
20: 10 00 f1 a0 ldl t6,16(a1)
24: 14 00 11 a1 ldl t7,20(a1)
28: 18 00 71 a0 ldl t2,24(a1)
2c: 08 00 f7 22 lda t9,8(t9)
30: 1c 00 51 a2 ldl a2,28(a1)
34: a0 07 f8 42 cmpule t9,t10,v0
38: 00 00 50 b0 stl t1,0(a0)
3c: 04 00 90 b0 stl t3,4(a0)
40: 20 00 31 22 lda a1,32(a1)
44: 08 00 b0 b0 stl t4,8(a0)
48: 0c 00 d0 b0 stl t5,12(a0)
4c: 10 00 f0 b0 stl t6,16(a0)
50: 14 00 10 b1 stl t7,20(a0)
54: 18 00 70 b0 stl t2,24(a0)
58: 1c 00 50 b2 stl a2,28(a0)
5c: 20 00 10 22 lda a0,32(a0)
60: eb ff 1f f4 bne v0,10 <f+0x10>
64: 01 80 fa 6b ret
however mainline produces:
0000000000000000 <f>:
0: 06 04 ff 47 clr t5
4: 00 00 fe 2f unop
8: 1f 04 ff 47 nop
c: 00 00 fe 2f unop
10: 05 04 26 42 addq a1,t5,t4
14: 04 04 06 42 addq a0,t5,t3
18: 04 00 c6 20 lda t5,4(t5)
1c: 00 00 65 a0 ldl t2,0(t4)
20: 05 04 26 42 addq a1,t5,t4
24: 01 04 e6 47 mov t5,t0
28: 1c f0 46 20 lda t1,-4068(t5)
2c: 00 00 64 b0 stl t2,0(t3)
30: 04 04 06 42 addq a0,t5,t3
34: 04 00 c6 20 lda t5,4(t5)
38: 00 00 65 a0 ldl t2,0(t4)
3c: 05 04 26 42 addq a1,t5,t4
40: 00 00 64 b0 stl t2,0(t3)
44: 04 04 06 42 addq a0,t5,t3
48: 08 00 c1 20 lda t5,8(t0)
4c: 00 00 65 a0 ldl t2,0(t4)
50: 05 04 26 42 addq a1,t5,t4
54: 00 00 64 b0 stl t2,0(t3)
58: 04 04 06 42 addq a0,t5,t3
5c: 0c 00 c1 20 lda t5,12(t0)
60: 00 00 65 a0 ldl t2,0(t4)
64: 05 04 26 42 addq a1,t5,t4
68: 00 00 64 b0 stl t2,0(t3)
6c: 04 04 06 42 addq a0,t5,t3
70: 10 00 c1 20 lda t5,16(t0)
74: 00 00 65 a0 ldl t2,0(t4)
78: 05 04 26 42 addq a1,t5,t4
7c: 00 00 64 b0 stl t2,0(t3)
80: 04 04 06 42 addq a0,t5,t3
84: 14 00 c1 20 lda t5,20(t0)
88: 00 00 65 a0 ldl t2,0(t4)
8c: 05 04 26 42 addq a1,t5,t4
90: 00 00 64 b0 stl t2,0(t3)
94: 04 04 06 42 addq a0,t5,t3
98: 18 00 c1 20 lda t5,24(t0)
9c: 00 00 65 a0 ldl t2,0(t4)
a0: 05 04 26 42 addq a1,t5,t4
a4: 00 00 64 b0 stl t2,0(t3)
a8: 04 04 06 42 addq a0,t5,t3
ac: 1c 00 c1 20 lda t5,28(t0)
b0: 00 00 65 a0 ldl t2,0(t4)
b4: 00 00 64 b0 stl t2,0(t3)
b8: d5 ff 5f f4 bne t1,10 <f+0x10>
bc: 01 80 fa 6b ret
that is, it doesn't reorder loads and stores anymore, and apparently introduces
lots of silly IVs.
--
Summary: Poor code from unrolled simple loop
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P2
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: falk at debian dot org
CC: gcc-bugs at gcc dot gnu dot org
GCC build triplet: alphaev68-unknown-linux-gnu
GCC host triplet: alphaev68-unknown-linux-gnu
GCC target triplet: alphaev68-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22031
