https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87245
Bug ID: 87245
Summary: [missed optimization] switching on indices of struct
members
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: herring at lanl dot gov
Target Milestone: ---
Consider a simple struct whose members can be accessed by name or by index:
struct vec {double x,y,z;};
double get(const struct vec *v,unsigned long i) {
switch(i) {
case 0: return v->x;
case 1: return v->y;
case 2: return v->z;
}
}
Godbolt's trunk x86_64 build produces at -O3
get:
cmpq $1, %rsi
ja .L2
je .L11
movsd (%rdi), %xmm0
ret
.L2:
cmpq $2, %rsi
jne .L12
movsd 16(%rdi), %xmm0
ret
.L11:
movsd 8(%rdi), %xmm0
ret
.L12:
ret
when of course
get:
movsd (%rdi,%rsi,8), %xmm0
ret
would suffice. (Apologies if I picked the wrong -optimization component.)
Bonus points for collapsing
double get(const struct vec *arr,unsigned long i) {
arr+=i/3;
switch(i%3) {
case 0: return arr->x;
case 1: return arr->y;
case 2: return arr->z;
}
}
to the same two instructions.
