https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120403
--- Comment #4 from Shivam Gupta <shivam98.tkg at gmail dot com> ---
Hello,
I have been experimenting with a prototype tree pass for this issue, I planned
to do this in two phases.
Phase 1 – Linearize equality chains (new tree pass)
The idea is to transform the conditional equality chain:
<bb 2> [local count: 1073741822]:
_1 = s1.x;
_2 = s2.x;
if (_1 == _2)
goto <bb 3>; [34.00%]
else
goto <bb 4>; [66.00%]
<bb 3> [local count: 365072224]:
_3 = s1.y;
_4 = s2.y;
_7 = _3 == _4;
<bb 4> [local count: 1073741824]:
# iftmp.0_5 = PHI <_7(3), 0(2)>
return iftmp.0_5;
into a straight-line XOR/OR sequence:
<bb 2> [local count: 1073741824]:
_1 = s1.x;
_2 = s2.x;
_3 = s1.y;
_4 = s2.y;
_7 = _1 ^ _2;
_8 = _3 ^ _4;
_9 = _7 | _8;
_10 = _9 == 0;
return _10;
This eliminates the branch and PHI node entirely. The resulting IR is becomes
simpler.
Phase 2 – Widening
Once the chain is linearized, look for opportunities to widen the XOR/OR
sequence across multiple fields into a single wider comparison — similar to
what LLVM does to produce cmp rdi, rsi for the struct s case and cmp edi, esi
for struct t.
I'm wondering whether a dedicated tree pass (running before if-combine) would
be cleaner, given that the widening step in Phase 2 may need to reason about
field layout and alignment separately.
