Fix reverse translation of ptr.annotation for multiple UserSemantic decorations

[vmaksimo]

It was found that the decorated variables (results of ptr.annotation intrinsic) are not used in further instructions unlike the input source. The issue is present for the reverse translation of multiple UserSemantic decorations.

This change fixes the reverse translation of multiple ptr.annotation intrinsics on the same class field. With this, the result of intrinsic is not lost and actually reused in further intrinsics and instructions.

[vmaksimo]

@MrSidims @asudarsa please take a look. Thanks!

[MrSidims]

Overall LGTM
@tiwaria1 may I ask you to check if this patch causes regressions on your side?

[MrSidims]

Restarting CI with a new base

[vmaksimo]

The test failure is not related to this patch. I found that the translation fails on this line call ptr @llvm.ptr.annotation.p0.p0(ptr %16, ptr %17, ptr @.str.1, i32 12, ptr null), because we expect the string constant instead of bitcast as a second argument. The IR has changed after type scavenger making input incorrect.

// input
%11 = getelementptr inbounds %class.A, ptr %2, i32 0, i32 1 
%12 = call ptr @llvm.ptr.annotation.p0.p0(ptr %11, ptr @.str.2, ptr @.str.1, i32 12, ptr null) 
%13 = call ptr @llvm.ptr.annotation.p0.p0(ptr %12, ptr @.str.3, ptr @.str.1, i32 12, ptr null)

// IR after type scavenger magic
%15 = getelementptr inbounds %class.A, ptr %2, i32 0, i32 1 
%16 = bitcast ptr %15 to ptr 
%17 = bitcast ptr @.str.2 to ptr 
%18 = call ptr @llvm.ptr.annotation.p0.p0(ptr %16, ptr %17, ptr @.str.1, i32 12, ptr null) 
%19 = bitcast ptr @.str.3 to ptr 
%20 = call ptr @llvm.ptr.annotation.p0.p0(ptr %18, ptr %19, ptr @.str.1, i32 12, ptr null)

Found that #1775 fixes this issue. I believe we should merge it before this one.

[MrSidims]

LGTM
But would prefer if someone else from Intel took a look.

[vmaksimo]

@dwoodwor-intel @sarnex @asudarsa please take a look. Thanks!

[sarnex]

I don't fully understand this but if nobody else responds ping me and we can go over it

[MrSidims]

I don't fully understand this but if nobody else responds ping me and we can go over it

That's actually the point - to share some expertise and have someone with a fresh look at annotations translations, since, it does look like the initial 'design' was not the best and two authors of this 'design' are author and reviewer of this patch :D

[vmaksimo]

@MrSidims can we still merge this one? I believe no one is concerned about the change :)