Instcombine should break up insane integer phi nodes

[lattner]

Bugzilla Link	3451
Resolution	FIXED
Resolved on	Nov 08, 2009 21:16
Version	1.0
OS	All
Attachments	example

Extended Description

188.ammp has a few aggregates promoted to large integers, but they are live across blocks. If they were local to a block instcombine would be able to eliminate the bitcast+zext+shift+shift+trunc+bitcast dance that is happening, but it can't. We should clean this up for a major perf boost (assuming the loop is hot). Search the attached .bc file for i576.

[lattner]

Fixed in
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20090126/073023.html

Thanks!

[lattner]

Oops, wrong bug.

[lattner]

Planned approach: visitPHI should notice that the phi is an unusual size (not i1/i8/i16/i32/i64). If so, it checks to see if all uses of the phi are "extractions" or uses of itself (x = phi x, ...). If so, split it up into the extracted chunks.

extractions are considered to be a truncate or a lshr+truncate.

[lattner]

Amusingly, with my latest changes to SROA, we now make a <9 x double> vector in 188.ammp!

[lattner]

r63638 teaches SROA to prefer insane integers to insane vectors.

[llvmbot]

Codegen of funky sized vectors isn't as robust as I'd like.
First off, it doesn't even try to handle things like <4 x i24>,
i.e. vectors of funky integer types. That's because it has not
yet been decided what the memory representation of these should
be (the target independent optimizers are inconsistent about it,
see #2156 ). Secondly, the vector widening stuff is quite new
and my impression (maybe unjustified) is that it could do with
some more solidifying.

[lattner]

xref: rdar://6628605

[lattner]

Another example from zsh, producing i96's

[lattner]

*** Bug llvm/llvm-bugzilla-archive#5159 has been marked as a duplicate of this bug. ***

[lattner]

This has been largely fixed in the lead up to r86503. The testcase in llvm/llvm-bugzilla-archive#5159 is still not fixed so I'm closing this and reopening it.

[lattner]

mentioned in issue llvm/llvm-bugzilla-archive#5159