AMDGPU: Do not not force alwaysinline with calls and module LDS enabled

[arsenm]

Summary

Function calls are now well supported. We used to also need this for
dealing with LDS uses in functions, but now that should also work.
This also removes the stress calls option, which would require moving
the flag to the TargetMachine and I no longer think it's useful.

We should probably drop the amdgpu-function-calls too but I think
hipcc is still using it.

This trims the pass list nicely (I'm surprised this saved a dominator
and alias analysis run).

[scchan]

We should probably drop the amdgpu-function-calls too but I think
hipcc is still using it.

The amdgpu-function-calls switch is needed to disable function calls as a prereq to enable alwaysinline, which I don't understand why.

[arsenm]

In D152414#4406127, @scchan wrote:

We should probably drop the amdgpu-function-calls too but I think
hipcc is still using it.

The amdgpu-function-calls switch is needed to disable function calls as a prereq to enable alwaysinline, which I don't understand why.

I'm failing to parse this sentence

[foad]

Typo in description "not not"

[scchan]

In D152414#4406186, @arsenm wrote:

In D152414#4406127, @scchan wrote:

We should probably drop the amdgpu-function-calls too but I think
hipcc is still using it.

The amdgpu-function-calls switch is needed to disable function calls as a prereq to enable alwaysinline, which I don't understand why.

I'm failing to parse this sentence

hipcc uses this switch because the enablement of alwaysinline requires that the function call support be turned off why do we have to turn off function call?

[arsenm]

In D152414#4406423, @scchan wrote:

hipcc uses this switch because the enablement of alwaysinline requires that the function call support be turned off why do we have to turn off function call?

I'm still confused about what you're asking. Function calls are supposed to always be on, they can't actually be completely avoided. amdgpu-function-calls was for bringup of call support and should be retired. This maintains the flag for the force off usage hipcc was relying on, for now.

[arsenm]

ping

[jmmartinez]

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

977–983

Currently the inliner gives a boost to the inline-threshold for functions that take alloca instructions as arguments (ArgAllocaCost in AMDGPUTargetTransformInfo.cpp).

Should that inline bonus be removed too?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

977–983

Those would be unrelated. That's just regular inline heuristics, not related to hacking out a call to avoid codegen limitations

[arsenm]

Rebase

[arsenm]

ping

[JonChesterfield]

LG here. Can we get rid of EnableFunctionCalls?

[arsenm]

In D152414#4492691, @JonChesterfield wrote:

LG here. Can we get rid of EnableFunctionCalls?

If we can get rid of hipcc which is still using it

[scchan]

In D152414#4493174, @arsenm wrote:

In D152414#4492691, @JonChesterfield wrote:

LG here. Can we get rid of EnableFunctionCalls?

If we can get rid of hipcc which is still using it

hipcc only needs a way to force EarlyInlineAll but that pass is predicated by !EnableFunctionCalls. If we don't think it's warranted then we can deprecate EnableFunctionCalls by first removing it in hipcc.