https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90571
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Turning indirect calls into direct ones might be important enough to also
handle
int x, y;
int f() { return x; }
int g() { return y; }
int t0(bool b) { int (*i)() = b ? &f : &g; x = 1; return i(); }
int main(int ac, char**) { return t0(ac & 1); }
like where there are statements before the indirect call that prevent it
from being simply duplicated into the predecessor blocks.
The transformation "primitive" would then be to duplicate the
joiner up to the call and the "interesting" part of it is
creating all required PHI nodes (unless you want to make SSA rewrite
deal with this somehow). Sinking stmts below the call and limiting
the amount of copying is important. Note there are related PRs for
that we miss to sink/hoist stmts through PHI nodes when that reduces the
number of PHI nodes.
The transform would likely split the block, insert "block-closed" PHI
nodes in the tail part for all SSA names defined in the first half and
live over the new edge and then duplicate the first half re-wiring
edges as needed.
This is as opposed to the original testcase where a simpler pattern-matching
scheme could be invented. I wonder how the "original" testcase looked
like - the one in this bug is probably simplified from real-world code?
