[PATCH] D29265: [legalize-types] Remove stale entries from SoftenedFloats.

[Chih-Hung Hsieh via Phabricator via llvm-commits]

chh added a comment.

Could we have more info in this change or https://reviews.llvm.org/D29180?
Which recent change broke which tests? Could it be reverted?
What were the before and after regression IL trees?
Do we need new unit test?

Last time I fixed a regression related to fp128 in https://reviews.llvm.org/D26942.
The regression was introduced by a cleanup with nice intention,
but we just do not have enough unit tests. We only found
the regression too late to revert, when we tried to compile
a large Android code base. Yes, fp128 could work for all other
targets except Android's x86-64. The SoftenFloat* functions got into
infinite loop. So I would be very careful now about removing nodes
from SoftenedFloats.


https://reviews.llvm.org/D29265