http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46306
Summary: inefficient code generated for array accesses
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: davi...@gcc.gnu.org
//Example:
int foo (int i, int *p, int t)
{
int p2 = p[i];
int temp = 0;
int temp2 = 1;
int temp3 = 4;
if (p[i+1] > t)
{
temp = p2;
temp2 = p2 + 2;
temp3 = p2 + 3;
}
return p[temp] + p [temp2] + p[temp3];
}
Two problems seen the code generated by trunk gcc at -O2
1) all the shift operation are redundant and should be folded as the stride in
the memory operand
2) unnecessary code duplication
(may be handled by a pass that converts memory access with linear address into
target memref in straight line code)
foo:
.LFB0:
.cfi_startproc
movslq %edi, %rdi
movl (%rsi,%rdi,4), %eax
cmpl %edx, 4(%rsi,%rdi,4)
jle .L3
movslq %eax, %rdi
leal 2(%rax), %ecx
salq $2, %rdi
leal 3(%rax), %edx
movslq %ecx, %rcx
movl (%rsi,%rdi), %eax
salq $2, %rcx
movslq %edx, %rdx
addl (%rsi,%rcx), %eax
salq $2, %rdx
addl (%rsi,%rdx), %eax
ret
.p2align 4,,10
.p2align 3
.L3:
movl $16, %edx
movl $4, %ecx
xorl %edi, %edi
movl (%rsi,%rdi), %eax
addl (%rsi,%rcx), %eax
addl (%rsi,%rdx), %eax
ret
// The following code is generated by another compiler -- not ideal, but
better:
foo:
.Leh_func_begin0:
pushq %rbp
.Ltmp0:
movq %rsp, %rbp
.Ltmp1:
movslq %edi, %rax
leal 1(%rax), %ecx
movslq %ecx, %rcx
cmpl %edx, (%rsi,%rcx,4)
jg .LBB0_2
movl $1, %eax
xorl %ecx, %ecx
movl $4, %edx
jmp .LBB0_3
.LBB0_2:
movslq (%rsi,%rax,4), %rcx
leal 3(%rcx), %eax
movslq %eax, %rdx
leal 2(%rcx), %eax
movslq %eax, %rax
.LBB0_3:
movl (%rsi,%rax,4), %eax
addl (%rsi,%rcx,4), %eax
addl (%rsi,%rdx,4), %eax
popq %rbp
ret
