https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120837
--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
For e.g. restrict-2.c, the assembly difference is
- movslq %ecx, %rcx
- movss 200(%rsi,%rcx,4), %xmm0
- addss -200(%rsi,%rcx,4), %xmm0
+ addl $50, %ecx
+ movslq %ecx, %rax
+ movss -400(%rsi,%rax,4), %xmm0
+ addss (%rsi,%rax,4), %xmm0
so worse code. gimple dump difference
- _1 = (sizetype) j;
- _2 = _1 + 50;
+ _1 = j + 50;
+ _2 = (long unsigned int) _1;
_3 = _2 * 4;
_4 = b + _3;
_5 = *_4;
- _6 = (long unsigned int) j;
- _7 = _6 * 4;
- _8 = _7 + 18446744073709551416;
+ _6 = j + -50;
+ _7 = (long unsigned int) _6;
+ _8 = _7 * 4;
and 12 Moving statement messages in lim2 instead of 11.
For copy-headers-5.c gimple difference
- _5 = (sizetype) i;
- _6 = _5 + 1;
+ _5 = i + 1;
+ _6 = (long unsigned int) _5;
and the loop header copying doesn't happen, assembly is shorter though:
- cmpl $1, %esi
- jle .L4
- leal -2(%rsi), %ecx
- movl (%rdi), %edx
- leaq 4(%rdi), %rax
- leaq 8(%rdi,%rcx,4), %rsi
- jmp .L3
- .p2align 5
- .p2align 4,,10
- .p2align 3
-.L8:
- addq $4, %rax
- cmpq %rsi, %rax
- je .L4
-.L3:
- movl %edx, %ecx
- movl (%rax), %edx
- cmpl %ecx, %edx
- jge .L8
+ subl $1, %esi
xorl %eax, %eax
- ret
+ jmp .L2
+ .p2align 4
.p2align 4,,10
.p2align 3
.L4:
+ movl (%rdi,%rax,4), %edx
+ addq $1, %rax
+ cmpl (%rdi,%rax,4), %edx
+ jg .L5
+.L2:
+ cmpl %eax, %esi
+ jg .L4
movl $1, %eax
ret
+ .p2align 4,,10
+ .p2align 3
+.L5:
+ xorl %eax, %eax
+ ret
