------- Comment #5 from ubizjak at gmail dot com 2009-02-09 08:28 -------
On alpha, we generate (-O3 -fno-tree-vectorize -funroll-loops):
$L2:
lda $19,4($1)
addq $17,$1,$21
lda $2,8($1)
cpys $f31,$f31,$f31
addq $17,$19,$20
ldl $22,0($21)
lda $0,12($1)
cpys $f31,$f31,$f31
addq $17,$2,$5
ldl $23,0($20)
addq $17,$0,$4
cpys $f31,$f31,$f31
lda $27,16($1)
ldl $25,0($5)
addq $16,$1,$8
ldl $24,0($4)
lda $20,20($1)
cpys $f31,$f31,$f31
addq $16,$19,$18
stl $22,0($8)
addq $17,$27,$7
cpys $f31,$f31,$f31
lda $19,24($1)
stl $23,0($18)
addq $16,$2,$6
ldl $22,0($7)
addq $17,$20,$3
cpys $f31,$f31,$f31
lda $18,28($1)
stl $25,0($6)
addq $16,$0,$21
ldl $8,0($3)
addq $17,$19,$28
cpys $f31,$f31,$f31
addq $17,$18,$23
stl $24,0($21)
addq $16,$27,$5
ldl $7,0($28)
addq $16,$20,$4
ldl $6,0($23)
stl $22,0($5)
addq $16,$19,$3
stl $8,0($4)
addq $16,$18,$0
stl $7,0($3)
lda $1,32($1)
stl $6,0($0)
lda $2,-4096($1)
bne $2,$L2
However, on x86_64 we compile the loop in similar way:
.L2:
leaq 1(%rax), %rdx
leaq 2(%rax), %r10
movl (%rsi,%rax,4), %r8d
movl (%rsi,%rdx,4), %ecx
movl %r8d, (%rdi,%rax,4)
movl (%rsi,%r10,4), %r11d
movl %ecx, (%rdi,%rdx,4)
leaq 3(%rax), %r8
movl %r11d, (%rdi,%r10,4)
leaq 4(%rax), %rdx
movl (%rsi,%r8,4), %r9d
movl (%rsi,%rdx,4), %ecx
movl %r9d, (%rdi,%r8,4)
leaq 5(%rax), %r10
movl %ecx, (%rdi,%rdx,4)
leaq 6(%rax), %r8
leaq 7(%rax), %rdx
movl (%rsi,%r10,4), %r11d
movl (%rsi,%r8,4), %r9d
movl %r11d, (%rdi,%r10,4)
movl (%rsi,%rdx,4), %ecx
movl %r9d, (%rdi,%r8,4)
addq $8, %rax
movl %ecx, (%rdi,%rdx,4)
cmpq $1024, %rax
jne .L2
It looks that RTL loop optimizer prefers calculated addresses instead of
offseted addressing.
--
ubizjak at gmail dot com changed:
What |Removed |Added
----------------------------------------------------------------------------
CC|uros at gcc dot gnu dot org |ubizjak at gmail dot com
Last reconfirmed|0000-00-00 00:00:00 |2009-02-09 08:28:40
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22031
