Missed optimization due to instcombine + sroa creating inttoptr on 32bit platforms

[dotdash]

Bugzilla Link	22786
Version	trunk
OS	Linux
Attachments	Testcase

Extended Description

InstCombine turns memcpys of 8 bytes into a load/store which SROA might later split up into smaller values. In the attached test case, this ends up extracting a pointer to doing a trunc + inttoptr operation, which stops the second conditional branch from being eliminated, although it's clear that the branch to "panic" will never be taken.

[dotdash]

Maybe instcombine should only convert memcpy to load/store if the size is <= pointer size on the target?