[SER] MaybeReorderThread + Make(Nop|Miss) HLSL -> DXIL lowering and tests #7262
Conversation
- HLSL -> DXIL lowering - ast, hlsl->dxil, dxilgen, and ScalarReplAggregatesHLSL tests SER implementation tracker (microsoft#7214)
tools/clang/test/DXC/Passes/ScalarReplHLSL/hitobject_make_scalarrepl.ll Outdated Show resolved Hide resolved
| %tmp2 = alloca %dx.types.HitObject, align 4 | ||
| %0 = bitcast %dx.types.HitObject* %hit to i8*, !dbg !23 ; line:42 col:3 | ||
| call void @llvm.lifetime.start(i64 4, i8* %0) #0, !dbg !23 ; line:42 col:3 | ||
| ; CHECK: %{{[^ ]+}} = call %dx.types.HitObject* @"dx.hl.op..%dx.types.HitObject* (i32, %dx.types.HitObject*)"(i32 358, %dx.types.HitObject* %{{[^ ]+}}) |
There was a problem hiding this comment.
This match would succeed even if the IR is unchanged from with is in the test case now, so we probably need some more specific matches, maybe a variable match in order to verify that this IR is actually being transformed.
There was a problem hiding this comment.
That is because this call to HitObject_MakeNop is not expected to change. SroaHLSL must only flatten the RayDesc param in the DxHitObject_MakeMiss call below and the test is checking for exactly that.
@llvm-beanz If that is not what we want, could you provide more specific guidance of what you want to see changed?
There was a problem hiding this comment.
If the goal is to make sure this doesn't get scalarized, we should have some variable matches to make sure that the values move as expected through the operations.
There was a problem hiding this comment.
Added explicit CHECKs that RayDesc gets scalarized and all other params remain in place. Also checking that the ptr from the default ctr remains unused.
@llvm-beanz does this resolve this request?
Co-authored-by: Chris B <beanz@abolishcrlf.org>
simoll changed the title 
tools/clang/test/CodeGenDXIL/hlsl/objects/HitObject/hitobject_make.hlsl Outdated Show resolved Hide resolved
This reverts commit 57b5fdd.
…return is a Clang codegen artifact)
4 participants
SER implementation tracker (#7214)