Hi, I think this is a pass-ordering issue: the generic fold already
exists, but the expression is exposed after the last generic
simplification opportunity. I traced the pass in which the RTL form
actually first appears: combine - does not contain the pattern
late_combine1 - does not contain the pattern ce2 -
contains it for the first time The first occurrence is: (set (reg:DI
128) (if_then_else:DI (eq:DI (reg/v:DI 82 [c])
(const_int 0)) (reg/v:DI 82 [c]) (const_int 0)))
{*seleqdi_using_di} pass_cse2 makes the condition operand and the
true arm the same register through common-subexpression substitution,
exposing the expression whose value is always zero. However, CSE does
not re-run simplify_ternary_operation after changing the SET_SRC.
late_combine2 later sees the already-formed IF_THEN_ELSE, but
late-combine does not simplify an existing SET_SRC. The current split is
a local cleanup for the target-specific expansion exposed at this point.
Thanks.
在 2026/11/13 06:27, Jeff Law 写道:
On 11/12/26 2:34 PM, Andrea Pinski wrote:
On Thu, Nov 12, 2026 at 1:33 AM Guo Jie <[email protected]> wrote:
RTL can contain:
(set D (if_then_else:GPR (eq:X OP 0) OP 0))
Both outcomes are zero. Fold this expression to a move of zero.
gcc/ChangeLog:
* config/loongarch/loongarch.md: New split.
This should really be done in simplify-rt.cc instead if not already.
Can you check on why it is not?
Yea, 100% agreed.
jeff
