memset/memcpy analysis/optimizations need improving

[llvmbot]

Bugzilla Link	452
Resolution	FIXED
Resolved on	Apr 22, 2009 04:37
Version	trunk
OS	All
Depends On	llvm/llvm-bugzilla-archive#2058 llvm/llvm-bugzilla-archive#1226 llvm/llvm-bugzilla-archive#1952 llvm/llvm-bugzilla-archive#2059
Reporter	LLVM Bugzilla Contributor

Extended Description

Just so this doesn't get lost, I'm filing a "wishlist" item for this.

Date: Wed, 6 Oct 2004 14:25:30 -0500 (CDT)
From: Chris Lattner sabre@nondot.org
Subject: memset optimizations

memset currently inhibits multiple optimizations, e.g. memset/load doesn't
get copy-propagated, memset doesn't cause dead stores to die, and memset
inhibits SRA. This should be fixed!

Also: memset(&X, 4) -> store int 0, X iff X is aligned

Also, memset should not kill unaliased loads/stores.

[llvmbot]

Edits a hard-coded string in an unrolled loop
I was told in the IRC channel that the non optimized LLVM code I was getting of
the attached C++ code was related to this issue.
It becomes optimized when the string 's' becomes small, like "this".

[lattner]

Bug 1309 is also related to this.

[llvmbot]

example with poor memcpy optimization
In the last line
return r + x.A;
the value of x.A is zero, but LLVM doesn't understand this.
Presumably this is because it thinks memcpy has modified x,
even though it copies from x and not to it.

[lattner]

Bug 1952 is one specific optimization that we can do, relating it.

[llvmbot]

This has gotten a lot better recently, with work Chris and I have done which is now in the MemCpyOpt pass. There's still a lot to do, though.

[llvmbot]

I'm closing this open-ended bug, as we'll never catch EVERY case. If you still have specific cases that you'd like to see addressed, please file individual PRs for them.