https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112935
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords|needs-bisection |missed-optimization
Target|loongarch64-*-* |loongson
Component|tree-optimization |target
--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Either r14-1655-g52c92fb3f40050 or r14-1654-g7ceed7e3e29c33 causes the
difference in the gimple level.
BUT
`a * boolean` is Canonical form and expr.cc expands it as `a & -boolean`:
```
/* Expand X*Y as X&-Y when Y must be zero or one. */
...
bool speed = optimize_insn_for_speed_p ();
int cost = add_cost (speed, mode) + neg_cost (speed, mode);
struct algorithm algorithm;
enum mult_variant variant;
if (CONST_INT_P (op1)
? !choose_mult_variant (mode, INTVAL (op1),
&algorithm, &variant, cost)
: cost < mul_cost (speed, mode))
{
target = bit0_p ? expand_and (mode, negate_rtx (mode, op0),
op1, target)
: expand_and (mode, op0,
negate_rtx (mode, op1),
target);
return REDUCE_BIT_FIELD (target);
}
```
Which means the cost model for loongson is wrong here.
Though it is comparing the cost of doing + and - to a * here though it should
be & and - but most target's + and & have a similar cost.
