[memcpyopt] Handle partially overlapping memcpys

[llvmbot]

Bugzilla Link	2058
Version	unspecified
OS	Linux
Blocks	#824
Attachments	testcase .ll, another slightly simpler example of the same thing
Reporter	LLVM Bugzilla Contributor
CC	@asl

Extended Description

In the attached testcase there are two memcpy's. In theory one
could be eliminated, for example by using %agg.result instead of
%temp everywhere. I admit that this is tricky - but not impossible!

[lattner]

This is interesting, but is going to be difficult to catch and doesn't arise commonly in C code. This is basically stuff like this:

... stuff...
memcpy(alloca+4, arg+4, 512)
memcpy(dst, alloca, 516)

This could be optimized into:

... stuff...
memcpy(alloca+4, arg+4, 512)
memcpy(dst, alloca, 4)
memcpy(dst, arg+4, 512)

which would hopefully allow the rest of the code to dissolve away. This gets into complex cost model stuff though.

[llvmbot]

The second testcase could be described as:

load 1 byte from src
store 1 byte to alloca
skip 3 bytes of alignment padding
memcpy(alloca+4, src+4, 512)

this could reasonably be combined into:

memcpy(alloca, src, 516)

If this was done, then I guess the second memcpy

memcpy(agg.result, alloca, 516)

would be zapped by the existing logic.

PS: In the testcase
src = getelementptr %struct.integers__big_numbers__big_signed* %x, i32 0, i32 1

[llvmbot]

mentioned in issue #824

[arsenm]

Modernized test integers__big_numbers__conversion__unsigned2signed.ll.txt

I only see 1 memcpy call in opt -O3