On 9/12/23 00:18, Lehua Ding wrote:
Hi Jeff,
On 2023/9/12 11:47, Jeff Law wrote:
But that condition is _not_ generally sufficient to prevent these
insns from existing during sched1. ie, a pass between split1 and
sched1 could create these patterns and successfully match them. That
in turn would trigger the assertion.
can_create_pseudo_p is true up through the register allocator. As a
result a condition like TARGET_VECTOR && can_create_pseudo_p() is
_not_ sufficient to ensure the pattern does not exist during sched1.
While no pass likely creates these kinds of insns right now in that
window between split1 and sched1, there's no guarantee that will
always be true.
But if a pass between split1 and sched1 creates these patterns, then an
unrecognized error will throw after reload. Is that right? That is to
say, this insn patterns is designed to exist only before split1, but now
the conditions are a little looser, a little tighter is better if we
can. If this is the case, I feel it makes no difference whether the
error is thrwoed by sched pass or a pass after reload.
If someone was to create one of these patterns without an associated
insn type, then the assert would trigger during sched1, and that is
good. The earlier we can catch an inconsistency, the better.
The simple rule is easy to follow. Every insn should have a type.
That also gives us a degree of future-proof, even if someone adds
additional passes/capabilities between split1 and sched1.
However, adding content that you don't need feels even more difficult to
understand, and this is just my feeling. It would be clearer if we could
set the type according to the purpose of the insn pattern.
I understand your position, but respectfully disagree with the
conclusion in this case.
jeff
