Skip to content

[DirectX] Eliminate resource global variables from module #114105

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 8 commits into from
Nov 12, 2024
Merged

[DirectX] Eliminate resource global variables from module #114105

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
3775e77
[DirectX] Mark buffer load/store as mem read/write
Oct 29, 2024
d013656
[DirectX] Eliminate resource global variables from module
Oct 30, 2024
127d582
Make GV removal more robust add opt dxil-op-lower test
Oct 31, 2024
5cc11c7
clang-format
Oct 31, 2024
2a18dca
Respond to sundry feedback
Nov 4, 2024
b9dc4fa
Merge remote-tracking branch 'refs/remotes/origin/main' into elim_res…
Nov 7, 2024
1edb01d
restore test for correct memory attributes applied to intrinsics
Nov 7, 2024
4ba7021
Merge remote-tracking branch 'refs/remotes/origin/main' into elim_res…
Nov 8, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions llvm/include/llvm/IR/IntrinsicsDirectX.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,12 +28,12 @@ def int_dx_handle_fromBinding
[IntrNoMem]>;

def int_dx_typedBufferLoad
: DefaultAttrsIntrinsic<[llvm_any_ty], [llvm_any_ty, llvm_i32_ty]>;
: DefaultAttrsIntrinsic<[llvm_any_ty], [llvm_any_ty, llvm_i32_ty], [IntrReadMem]>;
def int_dx_typedBufferLoad_checkbit
: DefaultAttrsIntrinsic<[llvm_any_ty, llvm_i1_ty],
[llvm_any_ty, llvm_i32_ty]>;
[llvm_any_ty, llvm_i32_ty], [IntrReadMem]>;
def int_dx_typedBufferStore
: DefaultAttrsIntrinsic<[], [llvm_any_ty, llvm_i32_ty, llvm_anyvector_ty]>;
: DefaultAttrsIntrinsic<[], [llvm_any_ty, llvm_i32_ty, llvm_anyvector_ty], [IntrWriteMem]>;
Copy link
Contributor

@bogner bogner Nov 1, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This file isn't exactly clang-format clean, but it would be nice to keep this formatting for this part to what clang-format says, ie:

def int_dx_typedBufferLoad
    : DefaultAttrsIntrinsic<[llvm_any_ty], [llvm_any_ty, llvm_i32_ty],
                            [IntrReadMem]>;
def int_dx_typedBufferLoad_checkbit
    : DefaultAttrsIntrinsic<[llvm_any_ty, llvm_i1_ty],
                            [llvm_any_ty, llvm_i32_ty], [IntrReadMem]>;
def int_dx_typedBufferStore
    : DefaultAttrsIntrinsic<[], [llvm_any_ty, llvm_i32_ty, llvm_anyvector_ty],
                            [IntrWriteMem]>;

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done. I don't know how to apply clang-format-diff to a specific file that's not included by the existing rules


// Cast between target extension handle types and dxil-style opaque handles
def int_dx_cast_handle : Intrinsic<[llvm_any_ty], [llvm_any_ty]>;
Expand Down
20 changes: 20 additions & 0 deletions llvm/lib/Target/DirectX/DXILOpLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,6 +204,22 @@ class OpLowerer {
CleanupCasts.clear();
}

// Remove the resource global associated with the handleFromBinding call instruction
// and their uses as they aren't needed anymore.
void removeResourceGlobals(CallInst *CI) {
for (User *User : make_early_inc_range(CI->users())) {
if(StoreInst *Store = dyn_cast<StoreInst>(User)) {
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Store->getOperand(1))) {
Copy link
Contributor

@coopp coopp Oct 31, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is special about operand 1? in Store->getOperand(1)?

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's the location of the global that gets stored to. In DXC, we would declare constants that made clear what these refer to, but I see no similar practice in LLVM.

Store->eraseFromParent();
assert(GV->use_empty() && "Buffer global still has users");
GV->removeDeadConstantUsers();
GV->eraseFromParent();
}
}
}
}


[[nodiscard]] bool lowerToCreateHandle(Function &F) {
IRBuilder<> &IRB = OpBuilder.getIRB();
Type *Int8Ty = IRB.getInt8Ty();
Expand All @@ -228,6 +244,8 @@ class OpLowerer {

Value *Cast = createTmpHandleCast(*OpCall, CI->getType());

removeResourceGlobals(CI);

CI->replaceAllUsesWith(Cast);
CI->eraseFromParent();
return Error::success();
Expand Down Expand Up @@ -272,6 +290,8 @@ class OpLowerer {

Value *Cast = createTmpHandleCast(*OpAnnotate, CI->getType());

removeResourceGlobals(CI);

CI->replaceAllUsesWith(Cast);
CI->eraseFromParent();

Expand Down
93 changes: 93 additions & 0 deletions llvm/test/CodeGen/DirectX/ResourceGlobalElimination.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,93 @@
; RUN: opt -S -passes='early-cse<memssa>' %s -o %t
; RUN: FileCheck --check-prefixes=CSE,CHECK %s < %t
; finish compiling to verify that dxil-op-lower removes the globals entirely
; RUN: llc -mtriple=dxil-pc-shadermodel6.0-compute --filetype=asm -o - %t | FileCheck --check-prefixes=LLC,CHECK %s
Copy link
Contributor

@coopp coopp Oct 31, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why are both shader model 6.0 and 6.6 used here? Is there some special difference?

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, the difference is whether the handle is created or created and annotated. We have separate paths for them, so even though they call the same utility function, I thought it better to test both paths.

; RUN: llc -mtriple=dxil-pc-shadermodel6.6-compute --filetype=asm -o - %t | FileCheck --check-prefixes=LLC,CHECK %s

; Ensure that EarlyCSE is able to eliminate unneeded loads of resource globals across typedBufferLoad.
; Also that DXILOpLowering eliminates the globals entirely.

target datalayout = "e-m:e-p:32:32-i1:32-i8:8-i16:16-i32:32-i64:64-f16:16-f32:32-f64:64-n8:16:32:64"
Copy link
Contributor

@bogner bogner Nov 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is the default datalayout for the triple. Best to remove it since it doesn't really affect anything

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Removed

target triple = "dxilv1.6-unknown-shadermodel6.6-compute"
Copy link
Contributor

@coopp coopp Oct 31, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does this triple get overridden by the -mtriple set above in the tests?

Copy link
Contributor

@bogner bogner Nov 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It does. We should only set the triple one way or the other here, so I would probably drop the truple from the llc command line

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd rather be able to run both options above, I've removed this too.


%"class.hlsl::RWBuffer" = type { target("dx.TypedBuffer", <4 x float>, 1, 0, 0) }

; LLC-NOT: @In = global
; LLC-NOT: @Out = global
@In = global %"class.hlsl::RWBuffer" zeroinitializer, align 4
@Out = global %"class.hlsl::RWBuffer" zeroinitializer, align 4

; Function Attrs: convergent noinline norecurse
Copy link
Contributor

@bogner bogner Nov 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I generally drop these function attr comments in tests - they're mostly just noise and can easily get out of sync with the actual attributes if we update those for one reason or another.

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Okay, removed.

; CHECK-LABEL define void @main()
define void @main() local_unnamed_addr #0 {
entry:
; LLC: %In_h.i1 = call %dx.types.Handle @dx.op.createHandle
; LLC: %Out_h.i2 = call %dx.types.Handle @dx.op.createHandle
%In_h.i = call target("dx.TypedBuffer", <4 x float>, 1, 0, 0) @llvm.dx.handle.fromBinding.tdx.TypedBuffer_v4f32_1_0_0t(i32 0, i32 0, i32 1, i32 0, i1 false)
store target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %In_h.i, ptr @In, align 4
%Out_h.i = call target("dx.TypedBuffer", <4 x float>, 1, 0, 0) @llvm.dx.handle.fromBinding.tdx.TypedBuffer_v4f32_1_0_0t(i32 4, i32 1, i32 1, i32 0, i1 false)
store target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %Out_h.i, ptr @Out, align 4
; CSE: call i32 @llvm.dx.flattened.thread.id.in.group()
%0 = call i32 @llvm.dx.flattened.thread.id.in.group()
; CHECK-NOT: load {{.*}} ptr @In
%1 = load target("dx.TypedBuffer", <4 x float>, 1, 0, 0), ptr @In, align 4
; CSE: call noundef <4 x float> @llvm.dx.typedBufferLoad.v4f32.tdx.TypedBuffer_v4f32_1_0_0t
%2 = call noundef <4 x float> @llvm.dx.typedBufferLoad.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %1, i32 %0)
; CHECK-NOT: load {{.*}} ptr @In
%3 = load target("dx.TypedBuffer", <4 x float>, 1, 0, 0), ptr @In, align 4
%4 = call noundef <4 x float> @llvm.dx.typedBufferLoad.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %3, i32 %0)
%add.i = fadd <4 x float> %2, %4
call void @llvm.dx.typedBufferStore.tdx.TypedBuffer_v4f32_1_0_0t.v4f32(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %Out_h.i, i32 %0, <4 x float> %add.i)
; CHECK: ret void
ret void
}

; Function Attrs: mustprogress nofree nosync nounwind willreturn memory(none)
declare i32 @llvm.dx.flattened.thread.id.in.group() #1

; Function Attrs: mustprogress nocallback nofree nosync nounwind willreturn
; CSE: declare <4 x float> @llvm.dx.typedBufferLoad.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0), i32) [[ROAttr:#[0-9]+]]
declare <4 x float> @llvm.dx.typedBufferLoad.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0), i32) #2
Copy link
Contributor

@bogner bogner Nov 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We're declaring the intrinsic with the wrong attributes here and then testing that it gets overwritten with the correct ones? That's kind of a weird thing to do...

I'd probably leave the declarations of the intrinsics out entirely since they'll be declared implicitly anyway.

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's the only way I have at the moment to directly test the attributes applied in IntrinsicsDirectX.td. That does go better with a CodeGen test when these can be generated. I'm happy to leave that for that change and limit this testing to the indirect effects of the attribute here.

Copy link
Contributor

@bogner bogner Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So the LLVM IR parser has a feature where if you use an intrinsic without declaring it, it will find the declaration implicitly. So what I'm saying is that you could omit the declare <4 x float> @llvm.dx.typedBufferLoad... line entirely and still check the generated attributes, rather than having a declaration whose attributes are intentionally wrong. The only potential pitfall is that the order of the declarations might not be guaranteed - it probably depends on the traversal order of some pass or other over the IR.

For example, consider https://hlsl.godbolt.org/z/6bohv41s5 - the source doesn't declare any intrinsics, but they're all present in the output of opt running no passes. Bonus - you don't even have to spell out the overload parts of the intrinsic names if you don't want to.

Copy link
Author

@pow2clk pow2clk Nov 7, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've restored the tests for intrinsic attributes as you suggested.


; Function Attrs: mustprogress nocallback nofree nosync nounwind willreturn
; CSE: declare void @llvm.dx.typedBufferStore.tdx.TypedBuffer_v4f32_1_0_0t.v4f32(target("dx.TypedBuffer", <4 x float>, 1, 0, 0), i32, <4 x float>) [[WOAttr:#[0-9]+]]
declare void @llvm.dx.typedBufferStore.tdx.TypedBuffer_v4f32_1_0_0t.v4f32(target("dx.TypedBuffer", <4 x float>, 1, 0, 0), i32, <4 x float>) #2

; Function Attrs: mustprogress nocallback nofree nosync nounwind willreturn memory(none)
declare target("dx.TypedBuffer", <4 x float>, 1, 0, 0) @llvm.dx.handle.fromBinding.tdx.TypedBuffer_v4f32_1_0_0t(i32, i32, i32, i32, i1) #3

; CSE: attributes [[ROAttr]] = { {{.*}} memory(read) }
; CSE: attributes [[WOAttr]] = { {{.*}} memory(write) }

attributes #0 = { convergent noinline norecurse "frame-pointer"="all" "hlsl.numthreads"="8,1,1" "hlsl.shader"="compute" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
attributes #1 = { mustprogress nofree nosync nounwind willreturn memory(none) }
attributes #2 = { mustprogress nocallback nofree nosync nounwind willreturn }
attributes #3 = { mustprogress nocallback nofree nosync nounwind willreturn memory(none) }

!llvm.module.flags = !{!0, !1}
!dx.valver = !{!2}
!llvm.ident = !{!3}

!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 7, !"frame-pointer", i32 2}
!2 = !{i32 1, i32 8}
!3 = !{!"clang version 20.0.0git ([email protected]:llvm/llvm-project.git 54dc966bd3d375d7c1604fac5fdac20989c1072a)"}
!4 = !{!5}
!5 = distinct !{!5, !6, !"_ZN4hlsl8RWBufferIDv4_fEixEi: %agg.result"}
!6 = distinct !{!6, !"_ZN4hlsl8RWBufferIDv4_fEixEi"}
!7 = !{!8, !9, i64 0}
!8 = !{!"_ZTSN4hlsl8RWBufferIDv4_fEE", !9, i64 0}
!9 = !{!"omnipotent char", !10, i64 0}
!10 = !{!"Simple C++ TBAA"}
!11 = !{!12}
!12 = distinct !{!12, !13, !"_ZN4hlsl8RWBufferIDv4_fEixEi: %agg.result"}
!13 = distinct !{!13, !"_ZN4hlsl8RWBufferIDv4_fEixEi"}
!14 = !{!15}
!15 = distinct !{!15, !16, !"_ZN4hlsl8RWBufferIDv4_fEixEi: %agg.result"}
!16 = distinct !{!16, !"_ZN4hlsl8RWBufferIDv4_fEixEi"}
!17 = !{!18, !9, i64 0}
!18 = !{!"_ZTSN4hlsl8__detail18TypedResourceProxyIU9_Res_u_CTDv4_fu17__hlsl_resource_tS2_EE", !9, i64 0, !19, i64 4}
!19 = !{!"int", !9, i64 0}
!20 = !{!21}
!21 = distinct !{!21, !22, !"_ZN4hlsl8__detail18TypedResourceProxyIU9_Res_u_CTDv4_fu17__hlsl_resource_tS2_EaSES2_: %agg.result"}
!22 = distinct !{!22, !"_ZN4hlsl8__detail18TypedResourceProxyIU9_Res_u_CTDv4_fu17__hlsl_resource_tS2_EaSES2_"}
Copy link
Contributor

@bogner bogner Nov 1, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You should be able to remove all of this metadata. Most of it isn't even referenced currently.

Copy link
Author

@pow2clk pow2clk Nov 4, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Removed. Funny that this shows as commenting on a large number of lines, but besides the text at the top, there's no visual indication of it.

Loading