Cleanup LclVar alignment requirements.

[sandreenko]

The current alignment mechanism for jit local vars is very complex, I think it is worth a cleanup.

Requirements:

1. There are different requirements for incoming arguments and local vars in the method:
   1.1 the incoming argument requirements are simple, esp. without vector calling convention.
2. There are correctness and performance requirements for local vars:
   2.1 there are more correctness requirements for armArch, less for XARCH;

I propose to create 2 functions:

• getAlignmentForArg();
• getAlignmentForLocal(bool correctness/performance);

What do we have now:

1. There is genTypeAlignments table that is defined in "typelist.h" and is used only in InferOpSizeAlign for ArmArch, for arm64 it is used only in a stress mode, for arm32 it is used in fgInitArgInfo;

2. How do we get alignment for arm64? It happens in lvaAllocLocalAndSetVirtualOffset, but this function has correct code only for 64-bit target, for arm32 we set the correct alignment before calling it:
   

   runtime/src/coreclr/jit/lclvars.cpp

   Line 6496 in d0121ea

   /* Need to align the offset? */

   
   and then call the function that handles 64-bit"
   

   runtime/src/coreclr/jit/lclvars.cpp

   Line 6536 in d0121ea

   stkOffs = lvaAllocLocalAndSetVirtualOffset(lclNum, lvaLclSize(lclNum), stkOffs);

the correctness requirements are hardcoded into the code, they are not coming from "typelist.h".

3. for arguments we usually just set alignment to TARGET_POINTER_SIZE and it works as long as we don't have vector calling convention, if we get it we will have the wrong alignment for arm/arm64.

With the new requirements for arm64 apple I had to check and change all of these places for the new ABI.

I suggest to delete alignment information from "typelist.h", delete genTypeAlignments, replace code in lvaAllocLocalAndSetVirtualOffset and lvaAssignVirtualFrameOffsetsToLocals with a call to getAlignmentForLocal;

Stress code in lvaStressLclFldCB also should be changed to use getAlignmentForLocal, the old code that was working only for armArch should be deleted.

category:design
theme:alignment
skill-level:intermediate
cost:medium
impact:small

[sandreenko]

@AndyAyersMS I would like to get your opinion about it if you think it is worth doing now.

[AndyAyersMS]

Generally I'm in favor of things that remove complexity (or at least try and limit the scope of complexity). This proposal seems like it is likely to do just that. So it seems like a good idea.