Optimize access of array member in a structure.

[shwqf]

This change adds a simple pattern match for a short AIR sequence of 'struct_field_ptr -> load -> array_elem_val', and generates same LLVM IR as the sequence of 'struct_field_ptr -> ptr_elem_ptr -> load'.

Although this change will still generate CallInst of Memcpy, LLVM will successfully identify the CallInst as dead code and eliminate it, preventing unuseful memcpy of the array field.

Resolves #13938

[Vexu]

It is good to fix things like this but doing it in llvm.zig means that it either has to be implemented for every backend or that it needs to be reimplemented in an earlier stage of the compiler.

[shwqf]

It is good to fix things like this but doing it in llvm.zig means that it either has to be implemented for every backend or that it needs to be reimplemented in an earlier stage of the compiler.

I am familiar with LLVM, so I changed llvm.zig first. I will try to optimize it in Sema.zig tomorrow.

[Vexu]

I am familiar with LLVM, so I changed llvm.zig first. I will try to optimize it in Sema.zig tomorrow.

Note that it is not a blocker for this PR since the other backends (besides maybe the C backend) are not ready for use.

[shwqf]

I am familiar with LLVM, so I changed llvm.zig first. I will try to optimize it in Sema.zig tomorrow.

Note that it is not a blocker for this PR since the other backends (besides maybe the C backend) are not ready for use.

lol, it would be great if we have a LLVM-like pass manager for AIR

[IntegratedQuantum]

Although this change will still generate CallInst of Memcpy, LLVM will successfully identify the CallInst as dead code and eliminate it

Does that also work in debug mode?
Because if it doesn't and the debug performance is still 10000 times worse, then this PR wouldn't help me(being the author of #13938) since I want to be able to test my program in debug mode.

[shwqf]

Although this change will still generate CallInst of Memcpy, LLVM will successfully identify the CallInst as dead code and eliminate it

Does that also work in debug mode? Because if it doesn't and the debug performance is still 10000 times worse, then this PR wouldn't help me(being the author of #13938) since I want to be able to test my program in debug mode.

Now it can

[matu3ba]

Can you rebase on top of master and run zig fmt? It fails:

-- Install configuration: "Debug"
+ echo Looking for non-conforming code formatting...
Looking for non-conforming code formatting...
+ stage3-debug/bin/zig fmt --check .. --exclude ../test/cases/ --exclude ../build-debug
../src/codegen/llvm.zig
Error: Process completed with exit code 1.

[shwqf]

Can you rebase on top of master and run zig fmt? It fails:

-- Install configuration: "Debug"
+ echo Looking for non-conforming code formatting...
Looking for non-conforming code formatting...
+ stage3-debug/bin/zig fmt --check .. --exclude ../test/cases/ --exclude ../build-debug
../src/codegen/llvm.zig
Error: Process completed with exit code 1.

fixed

[andrewrk]

Thank you for sending this enhancement.

In the future, I want to implement #13265. This will make the LLVM backend no longer use the LLVM IR builder API and instead output bitcode directly. Under such circumstances, we will not have access to these C++ APIs, or this SSA structure that LLVM creates. This patch takes us further away from that goal, and so I request that the optimization be done without doing either of these things:

• Introducing new C++ code
• Depending on more of LLVM's API surface area

[shwqf]

Thank you for sending this enhancement.

In the future, I want to implement #13265. This will make the LLVM backend no longer use the LLVM IR builder API and instead output bitcode directly. Under such circumstances, we will not have access to these C++ APIs, or this SSA structure that LLVM creates. This patch takes us further away from that goal, and so I request that the optimization be done without doing either of these things:

• Introducing new C++ code
• Depending on more of LLVM's API surface area

The new C++ code is for removing useless memcpy in debug mode, LLVM can optimize it automatically in release mode.

I think we can't avoid it because we need to track the number of users of the AllocaInst, to make sure that users don't intend to copy the field, that is why I call 'hasOneUser'.

I guess it is ok, if we want to generate .bc directly, we should implement use-def and def-use in air, this function is easy to reimplement.

[Vexu]

• Introducing new C++ code

• Depending on more of LLVM's API surface area

One possible but slightly harder solution would be to implement this in AstGen.

[shwqf]

• Introducing new C++ code
• Depending on more of LLVM's API surface area

One possible but slightly harder solution would be to implement this in AstGen.

I tried but I found AST is untyped.

[Vexu]

Types are not needed to fix this, field access can result in a pointer via field_ptr and array access can take a pointer with elem_ptr and then turn it into a value using load. The issue is that currently AstGen uses field_val and elem_val and then relies on LLVM to fix it.

[shwqf]

Types are not needed to fix this, field access can result in a pointer via field_ptr and array access can take a pointer with elem_ptr and then turn it into a value using load. The issue is that currently AstGen uses field_val and elem_val and then relies on LLVM to fix it.

How can we tell if a field is a slice or an array?

My first implementation is in AstGen as you suggest, statg3 cannot be compiled.

[Vexu]

elem_ptr works for all indexable types including arrays and slices.

[shwqf]

I got COVID in China, too tired to continue to finish AstGen version.

Some comptime_field_ptr problems appear, looks not naive.

I remove the redundant C++ code, now this PR only work in non-Debug mode.

@Vexu @matu3ba @andrewrk

[matu3ba]

Thanks, I think you need to apply zig fmt:

+ echo Looking for non-conforming code formatting...
+ stage3-debug/bin/zig fmt --check .. --exclude ../test/cases/ --exclude ../build-debug
Looking for non-conforming code formatting...
../lib/std/c/darwin.zig
Error: Process completed with exit code 1.