GitLab

create_compilerlib_entity() does it already, so it is not necessary to do it again in replace_with_call().

If this Conv is inserted, then the existing Conv becomes a nop.
So let the existing Conv handle the conversion.

Strange casts may cause that a Load ends up with a method entity to load from.

This prevented optimizing x < 0 ? -1 : 0, when the result is unsigned.

The same test is done by an enclosing if.

This simplifies to understand the pattern when reading the code.

It is a sign extension when the operand is signed, not just when both the operand and the result are signed.

This prevents accidently hitting a Switch, though this code can only handle Cond.
This code probably is older than the Cond-Switch schism and later tests removed switch-like Conds.

This typo was always there (even before Cmp relations), so this optimization never worked before.

It still assumed that Cmps have Projs.
Nobody noticed, because no -O level in cparser activates this optimization.

Truncate operands only get bigger, so the condition is always true.

The verifier already ensures that Not cannot have float mode.

mode_b cannot/must be present in these places, otherwise the verifier would have complained.

* It can never trigger, because its implementation is wrong.
* If it were correct, it would be subsumed by the same check using bitinfo.

They can be handled during instruction selection.

This fixes opt/fehler295.c.

Especially move rules that expects a constant on the right-hand side
after the normalization step.

1 << x has only one bit but can also occur on the left-hand side.

This fixes opt/fehler296.c.

Adapt only optimization that has created such nodes.