https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96245
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Looking at godbolt and LLVM release notes, this optimization appeared in LLVM 9
and looks to be a switch optimization:
LLVM can now sink similar instructions to a common successor block also when
the instructions have no uses, such as calls to void functions. This allows
code such as
void g(int);
enum e { A, B, C, D };
void f(e x, int y, int z) {
switch(x) {
case A: g(6); break;
case B: g(3); break;
case C: g(9); break;
case D: g(2); break;
}
}
to be optimized to a single call to g, with the argument loaded from a lookup
table.
I bet it is https://reviews.llvm.org/D59936
We don't optimize even:
void
foo (int *p, int x)
{
switch (x)
{
case 7: *p = 14; break;
case 8: *p = 15; break;
case 9: *p = 16; break;
case 10: *p = 17; break;
case 11: *p = 18; break;
case 12: *p = 19; break;
default: return;
}
}
but do optimize if one performs the switch parametrized tail merging manually:
void
bar (int *p, int x)
{
int y;
switch (x)
{
case 7: y = 14; break;
case 8: y = 15; break;
case 9: y = 16; break;
case 10: y = 17; break;
case 11: y = 18; break;
case 12: y = 19; break;
default: return;
}
*p = y;
}
So, do you prefer to do this in tree-ssa-tail-merge.c (for blocks starting with
switches) or in tree-switch-conversion.c, and should we handle both the case
where the constant(s) are linear vs. the switch expression, or also any other
(let the switchconv pass then create the arrays)?