[Attributor] New attribute to identify what byte ranges are alive for an allocation

[vidsinghal]

Summary

Changes the size of allocations automatically.
Only implements the case when a single range from start of the allocation is alive.

[vidsinghal]

Fix typo in assert.
Add a positive test case for malloc with a TODO.

[vidsinghal]

This comment was removed by vidsinghal.

[jdoerfert]

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

12686

12694

Don't cast it to an "Impl". There is no need anymore anyway.

12695

State could be anything. At least call it PIState.

12699–12705

This is basically BinSize != 1, return, no?

12718

Why does this not trigger for the ptr/i64 case? Is one of them in bytes the other in bits?

12721

This is fine for OffsetEnd = 2,4,8, but then it becomes a little iffy. Check for non default sizes and go with a char array instead.

12739

Same as above.

llvm/test/Transforms/Attributor/value-simplify-local-remote.ll

82 ↗

(On Diff #551856)

This doesn't actually make it smaller, we should avoid such changes.

135 ↗

(On Diff #551856)

Same here.

[vidsinghal]

Revision based on comments

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

12699–12705

umm, I think BinSize == 0

12718

if its a ptr

public: std::optional<TypeSize> getAllocationSize(const DataLayout &DL)

returns a size of 0 which is why that if condition is not triggered.
At the LLVM IR level, I think the size of pointer is just 0.

[vidsinghal]

Fix bugs

[jdoerfert]

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

12727

Why do you give up on struct type like {i32, i32}?
I don't understand why we need this conditional at all.

[vidsinghal]

Remove un-necessary if condition for struct with only one element.

[vidsinghal]

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

12727

I was trying to indicate a fix point for structs like
type S = {ptr}
where the struct has only one element in it.
But I think that OffsetEnd == AllocationSize check should work for this now (after fixing the size of pointer issue) I'll remove this check.
Previously this was not working which is why I had to had this check and i think
getNumContainedTypes should have been getStructNumElements instead.

[vidsinghal]

check for zero alloca and indicate fixpoint.

[tianshilei1992]

llvm/lib/Transforms/IPO/Attributor.cpp

3653

Better to use [[maybe_unused]] now.

[jdoerfert]

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

12689–12690

As I said before, no casting to implementations of the interface. put the method you need into AAPointerInfo.

12738–12740

llvm/test/Transforms/Attributor/value-simplify-local-remote.ll

384 ↗

(On Diff #552461)

unrelated, just keep the old version of this file.

[vidsinghal]

1.) Insert a new instruction instead of modiying the old one (Char array)
2.) Restore value-simplify-local-remote.ll
3.) Add a new positive test to reduce size of array alloca

[vidsinghal]

1.) Put method to get offsetBins in AAPointerInfo
2.) work on other comments.

[vidsinghal]

add maybe_unused pragma

[vidsinghal]

edits

[jdoerfert]

llvm/include/llvm/Transforms/IPO/Attributor.h

6127

This is not a good idea. What happend to the old API you had?
We do not want to leak out the underlying map if we don't have to.
All you want is a size and begin/end, no?

[vidsinghal]

llvm/include/llvm/Transforms/IPO/Attributor.h

6127

Yeah I see.
The underlying map is protected in the state I think.
Yeah for now, size begin and end should be good.
I think I will try making those virtual pure in AAPointerInfo and then override them in the state?

[vidsinghal]

Add API for begin, end size in AAPointerInfo

[vidsinghal]

Add a positive test with TODO

[vidsinghal]

Make switch to indentify alloca vs malloc instruction

[tianshilei1992]

llvm/lib/Transforms/IPO/Attributor.cpp

3653

Then there is no need (void)Success;.

[vidsinghal]

remove (void)success

[vidsinghal]

Support for modifying malloc instructions
TODO: The original malloc call is not being removed by changeAfterManifest even though it is dead.

[vidsinghal]

edits

[vidsinghal]

fix name variable

[vidsinghal]

edit

[vidsinghal]

Delete malloc call after attributor update

[vidsinghal]

fix example IR

[vidsinghal]

edits

[vidsinghal]

Don't call AAPointerInfoGetOrCreate other than Alloca and Malloc Instruction. Results in infinite loop for instance when calling a ptr

[vidsinghal]

Clean up code

[vidsinghal]

style

[vidsinghal]

Add new test where argument to malloc isn't a constant.
Use AAPotentialConstantValues to get the constant value.

[vidsinghal]

Add null check for IntOperand

[vidsinghal]

New example and question about what the optimization should do for an example like @baz in allocator.ll

[vidsinghal]

Remove run commands in depgraph.ll that were not being parsed