[llvm] [DXIL] Add GroupMemoryBarrierWithGroupSync intrinsic (PR #114349)
Adam Yang via llvm-commits
llvm-commits at lists.llvm.org
Wed Oct 30 20:49:22 PDT 2024
https://github.com/adam-yang created https://github.com/llvm/llvm-project/pull/114349
fixes #112974
partially fixes #70103
This change was reverted so some issues could be fixed.
### Changes
- Added new tablegen based way of lowering dx intrinsics to DXIL ops.
- Added int_dx_group_memory_barrier_with_group_sync intrinsic in IntrinsicsDirectX.td
- Added expansion for int_dx_group_memory_barrier_with_group_sync in DXILIntrinsicExpansion.cpp`
- Added DXIL backend test case
### Related PRs
* [[clang][HLSL] Add GroupMemoryBarrierWithGroupSync intrinsic #111883](https://github.com/llvm/llvm-project/pull/111883)
* [[SPIRV] Add GroupMemoryBarrierWithGroupSync intrinsic #111888](https://github.com/llvm/llvm-project/pull/111888)
>From c283c9f95ad18a8c0fce8b676752d30ec777390f Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 8 Oct 2024 00:53:11 -0700
Subject: [PATCH 01/16] Added GroupMemoryBarrierWithGroupSync intrinsic for
DXIL
---
llvm/include/llvm/IR/IntrinsicsDirectX.td | 2 ++
llvm/lib/Target/DirectX/DXIL.td | 9 ++++++
llvm/lib/Target/DirectX/DXILConstants.h | 7 +++++
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 31 +++++++++++++++++++
.../GroupMemoryBarrierWithGroupSync.ll | 8 +++++
5 files changed, 57 insertions(+)
create mode 100644 llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index e30d37f69f781e..015cc0ddab4026 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -92,4 +92,6 @@ def int_dx_step : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, L
def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>],
[LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [IntrNoMem]>;
def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+
+def int_dx_groupMemoryBarrierWithGroupSync : DefaultAttrsIntrinsic<[], [], []>;
}
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index 1e8dc63ffa257e..d7b08ff102cf69 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -278,6 +278,7 @@ def IsFeedback : DXILAttribute;
def IsWave : DXILAttribute;
def NeedsUniformInputs : DXILAttribute;
def IsBarrier : DXILAttribute;
+def NoDuplicate : DXILAttribute;
class Overloads<Version ver, list<DXILOpParamType> ols> {
Version dxil_version = ver;
@@ -829,3 +830,11 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
let stages = [Stages<DXIL1_0, [all_stages]>];
let attributes = [Attributes<DXIL1_0, [ReadNone]>];
}
+
+def Barrier : DXILOp<80, barrier> {
+ let Doc = "inserts a memory barrier in the shader";
+ let arguments = [Int32Ty];
+ let result = VoidTy;
+ let stages = [Stages<DXIL1_0, [compute, library]>];
+ let attributes = [Attributes<DXIL1_0, [NoDuplicate]>];
+}
diff --git a/llvm/lib/Target/DirectX/DXILConstants.h b/llvm/lib/Target/DirectX/DXILConstants.h
index 022cd57795a063..38984727761bb3 100644
--- a/llvm/lib/Target/DirectX/DXILConstants.h
+++ b/llvm/lib/Target/DirectX/DXILConstants.h
@@ -30,6 +30,13 @@ enum class OpParamType : unsigned {
#include "DXILOperation.inc"
};
+enum class BarrierMode : unsigned {
+ SyncThreadGroup = 0x00000001,
+ UAVFenceGlobal = 0x00000002,
+ UAVFenceThreadGroup = 0x00000004,
+ TGSMFence = 0x00000008,
+};
+
} // namespace dxil
} // namespace llvm
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 8acc9c1efa08c0..ecfa120f3d7557 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -234,6 +234,34 @@ class OpLowerer {
});
}
+ [[nodiscard]] bool lowerBarrier(Function& F, Intrinsic::ID IntrId) {
+ IRBuilder<> &IRB = OpBuilder.getIRB();
+ return replaceFunction(F, [&](CallInst *CI) -> Error {
+ unsigned BarrierMode = 0;
+ switch (IntrId) {
+ default:
+ report_fatal_error("Unhandled barrier operation type.");
+ break;
+ case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
+ BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence | (unsigned)dxil::BarrierMode::SyncThreadGroup;
+ break;
+ }
+
+ std::array<Value *, 1> Args{IRB.getInt32(BarrierMode)};
+
+ IRB.SetInsertPoint(CI);
+ Expected<CallInst *> OpCall =
+ OpBuilder.tryCreateOp(OpCode::Barrier, Args, CI->getName());
+ if (Error E = OpCall.takeError())
+ return E;
+
+ CI->replaceAllUsesWith(OpCall.get());
+ CI->eraseFromParent();
+
+ return Error::success();
+ });
+ }
+
[[nodiscard]] bool lowerToBindAndAnnotateHandle(Function &F) {
IRBuilder<> &IRB = OpBuilder.getIRB();
@@ -588,6 +616,9 @@ class OpLowerer {
HasErrors |= replaceFunctionWithOp(F, OpCode); \
break;
#include "DXILOperation.inc"
+ case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
+ HasErrors |= lowerBarrier(F, ID);
+ break;
case Intrinsic::dx_handle_fromBinding:
HasErrors |= lowerHandleFromBinding(F);
break;
diff --git a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll b/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
new file mode 100644
index 00000000000000..a99c6757814f3b
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
@@ -0,0 +1,8 @@
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+
+define void @test_group_memory_barrier_with_group_sync() {
+entry:
+ ; CHECK: call void @dx.op.barrier(i32 80, i32 9)
+ call void @llvm.dx.groupMemoryBarrierWithGroupSync()
+ ret void
+}
\ No newline at end of file
>From 127ac317636042dac1abb44caf4a83b0206888a5 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 12:25:26 -0700
Subject: [PATCH 02/16] Changed naming convention and fixed formatting
---
llvm/include/llvm/IR/IntrinsicsDirectX.td | 2 +-
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 23 ++++++++-----------
...> group_memory_barrier_with_group_sync.ll} | 2 +-
3 files changed, 12 insertions(+), 15 deletions(-)
rename llvm/test/CodeGen/DirectX/{GroupMemoryBarrierWithGroupSync.ll => group_memory_barrier_with_group_sync.ll} (80%)
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 015cc0ddab4026..dada426368995d 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -93,5 +93,5 @@ def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>
[LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [IntrNoMem]>;
def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
-def int_dx_groupMemoryBarrierWithGroupSync : DefaultAttrsIntrinsic<[], [], []>;
+def int_dx_group_memory_barrier_with_group_sync : DefaultAttrsIntrinsic<[], [], []>;
}
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index ecfa120f3d7557..9f9da213c48419 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -234,19 +234,14 @@ class OpLowerer {
});
}
- [[nodiscard]] bool lowerBarrier(Function& F, Intrinsic::ID IntrId) {
+ [[nodiscard]] bool lowerBarrier(Function &F, Intrinsic::ID IntrId,
+ ArrayRef<dxil::BarrierMode> BarrierModes) {
+ unsigned BarrierMode = 0;
+ for (const dxil::BarrierMode B : BarrierModes) {
+ BarrierMode |= (unsigned)B;
+ }
IRBuilder<> &IRB = OpBuilder.getIRB();
return replaceFunction(F, [&](CallInst *CI) -> Error {
- unsigned BarrierMode = 0;
- switch (IntrId) {
- default:
- report_fatal_error("Unhandled barrier operation type.");
- break;
- case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
- BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence | (unsigned)dxil::BarrierMode::SyncThreadGroup;
- break;
- }
-
std::array<Value *, 1> Args{IRB.getInt32(BarrierMode)};
IRB.SetInsertPoint(CI);
@@ -616,8 +611,10 @@ class OpLowerer {
HasErrors |= replaceFunctionWithOp(F, OpCode); \
break;
#include "DXILOperation.inc"
- case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
- HasErrors |= lowerBarrier(F, ID);
+ case Intrinsic::dx_group_memory_barrier_with_group_sync:
+ HasErrors |= lowerBarrier(
+ F, ID,
+ {dxil::BarrierMode::TGSMFence, dxil::BarrierMode::SyncThreadGroup});
break;
case Intrinsic::dx_handle_fromBinding:
HasErrors |= lowerHandleFromBinding(F);
diff --git a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
similarity index 80%
rename from llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
rename to llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index a99c6757814f3b..48907647c660f8 100644
--- a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -3,6 +3,6 @@
define void @test_group_memory_barrier_with_group_sync() {
entry:
; CHECK: call void @dx.op.barrier(i32 80, i32 9)
- call void @llvm.dx.groupMemoryBarrierWithGroupSync()
+ call void @llvm.dx.group.memory.barrier.with.group.sync()
ret void
}
\ No newline at end of file
>From 6bc974608eb18abbf72aa296c4fa0228c0a8757e Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 12:59:52 -0700
Subject: [PATCH 03/16] Got rid of the noduplicate attr
---
llvm/lib/Target/DirectX/DXIL.td | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index d7b08ff102cf69..c195887c754690 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -278,7 +278,6 @@ def IsFeedback : DXILAttribute;
def IsWave : DXILAttribute;
def NeedsUniformInputs : DXILAttribute;
def IsBarrier : DXILAttribute;
-def NoDuplicate : DXILAttribute;
class Overloads<Version ver, list<DXILOpParamType> ols> {
Version dxil_version = ver;
@@ -836,5 +835,5 @@ def Barrier : DXILOp<80, barrier> {
let arguments = [Int32Ty];
let result = VoidTy;
let stages = [Stages<DXIL1_0, [compute, library]>];
- let attributes = [Attributes<DXIL1_0, [NoDuplicate]>];
+ let attributes = [Attributes<DXIL1_0, []>];
}
>From 0483bd89b73637476f0ae9621e0ed40e17d5307e Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:03:52 -0700
Subject: [PATCH 04/16] Shader model changed to 6.0
---
.../CodeGen/DirectX/group_memory_barrier_with_group_sync.ll | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index 48907647c660f8..7cacbe778ac952 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.0-library < %s | FileCheck %s --check-prefix=CHECK
define void @test_group_memory_barrier_with_group_sync() {
entry:
>From 5b0ccac939936fa002f2e5f5d5e0c2091f90610b Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:07:03 -0700
Subject: [PATCH 05/16] Fixed the incorrect arguments list
---
llvm/lib/Target/DirectX/DXIL.td | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index c195887c754690..baa64d90ae64d9 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -832,7 +832,7 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
def Barrier : DXILOp<80, barrier> {
let Doc = "inserts a memory barrier in the shader";
- let arguments = [Int32Ty];
+ let arguments = [];
let result = VoidTy;
let stages = [Stages<DXIL1_0, [compute, library]>];
let attributes = [Attributes<DXIL1_0, []>];
>From c3cc475f8c5d9e6553913292aa8577884f103f57 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:13:21 -0700
Subject: [PATCH 06/16] Revert "Shader model changed to 6.0"
This reverts commit c7d83cf8ae32d98b0677ff1a88f74fe4827dd61f.
---
.../CodeGen/DirectX/group_memory_barrier_with_group_sync.ll | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index 7cacbe778ac952..48907647c660f8 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.0-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
define void @test_group_memory_barrier_with_group_sync() {
entry:
>From b48c0fe46af37c6ac20d69e371f45d96e6411e02 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 15 Oct 2024 15:56:50 -0700
Subject: [PATCH 07/16] Added another intermediate memory barrier intrinsic
that maps directly to the barrier dxil op
---
llvm/include/llvm/IR/IntrinsicsDirectX.td | 1 +
llvm/lib/Target/DirectX/DXIL.td | 3 +-
.../Target/DirectX/DXILIntrinsicExpansion.cpp | 26 +++++++++++++++++
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 28 -------------------
.../group_memory_barrier_with_group_sync.ll | 2 +-
5 files changed, 30 insertions(+), 30 deletions(-)
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index dada426368995d..d16a71f717b8db 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -93,5 +93,6 @@ def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>
[LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [IntrNoMem]>;
def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+def int_dx_memory_barrier : DefaultAttrsIntrinsic<[], [llvm_i32_ty], []>;
def int_dx_group_memory_barrier_with_group_sync : DefaultAttrsIntrinsic<[], [], []>;
}
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index baa64d90ae64d9..d00253b0b7c03e 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -832,7 +832,8 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
def Barrier : DXILOp<80, barrier> {
let Doc = "inserts a memory barrier in the shader";
- let arguments = [];
+ let LLVMIntrinsic = int_dx_memory_barrier;
+ let arguments = [Int32Ty];
let result = VoidTy;
let stages = [Stages<DXIL1_0, [compute, library]>];
let attributes = [Attributes<DXIL1_0, []>];
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index fb5383b3514a5a..56dca69b42f1fe 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -10,6 +10,7 @@
// opcodes in DirectX Intermediate Language (DXIL).
//===----------------------------------------------------------------------===//
+#include "DXILConstants.h"
#include "DXILIntrinsicExpansion.h"
#include "DirectX.h"
#include "llvm/ADT/STLExtras.h"
@@ -66,6 +67,7 @@ static bool isIntrinsicExpansion(Function &F) {
case Intrinsic::dx_sign:
case Intrinsic::dx_step:
case Intrinsic::dx_radians:
+ case Intrinsic::dx_group_memory_barrier_with_group_sync:
return true;
}
return false;
@@ -452,6 +454,27 @@ static Value *expandRadiansIntrinsic(CallInst *Orig) {
return Builder.CreateFMul(X, PiOver180);
}
+static Value *expandMemoryBarrier(CallInst *Orig, Intrinsic::ID IntrinsicId) {
+ assert(IntrinsicId == Intrinsic::dx_group_memory_barrier_with_group_sync);
+ unsigned BarrierMode = 0;
+ switch (IntrinsicId) {
+ case Intrinsic::dx_group_memory_barrier_with_group_sync:
+ BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence |
+ (unsigned)dxil::BarrierMode::SyncThreadGroup;
+ break;
+ default:
+ report_fatal_error(Twine("Unexpected memory barrier intrinsic."),
+ /* gen_crash_diag=*/false);
+ break;
+ }
+
+ IRBuilder<> Builder(Orig);
+ return Builder.CreateIntrinsic(
+ Builder.getVoidTy(), Intrinsic::dx_memory_barrier,
+ ArrayRef<Value *>{Builder.getInt32(BarrierMode)}, nullptr,
+ Orig->getName());
+}
+
static Intrinsic::ID getMaxForClamp(Type *ElemTy,
Intrinsic::ID ClampIntrinsic) {
if (ClampIntrinsic == Intrinsic::dx_uclamp)
@@ -586,6 +609,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
case Intrinsic::dx_radians:
Result = expandRadiansIntrinsic(Orig);
break;
+ case Intrinsic::dx_group_memory_barrier_with_group_sync:
+ Result = expandMemoryBarrier(Orig, IntrinsicId);
+ break;
}
if (Result) {
Orig->replaceAllUsesWith(Result);
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 9f9da213c48419..8acc9c1efa08c0 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -234,29 +234,6 @@ class OpLowerer {
});
}
- [[nodiscard]] bool lowerBarrier(Function &F, Intrinsic::ID IntrId,
- ArrayRef<dxil::BarrierMode> BarrierModes) {
- unsigned BarrierMode = 0;
- for (const dxil::BarrierMode B : BarrierModes) {
- BarrierMode |= (unsigned)B;
- }
- IRBuilder<> &IRB = OpBuilder.getIRB();
- return replaceFunction(F, [&](CallInst *CI) -> Error {
- std::array<Value *, 1> Args{IRB.getInt32(BarrierMode)};
-
- IRB.SetInsertPoint(CI);
- Expected<CallInst *> OpCall =
- OpBuilder.tryCreateOp(OpCode::Barrier, Args, CI->getName());
- if (Error E = OpCall.takeError())
- return E;
-
- CI->replaceAllUsesWith(OpCall.get());
- CI->eraseFromParent();
-
- return Error::success();
- });
- }
-
[[nodiscard]] bool lowerToBindAndAnnotateHandle(Function &F) {
IRBuilder<> &IRB = OpBuilder.getIRB();
@@ -611,11 +588,6 @@ class OpLowerer {
HasErrors |= replaceFunctionWithOp(F, OpCode); \
break;
#include "DXILOperation.inc"
- case Intrinsic::dx_group_memory_barrier_with_group_sync:
- HasErrors |= lowerBarrier(
- F, ID,
- {dxil::BarrierMode::TGSMFence, dxil::BarrierMode::SyncThreadGroup});
- break;
case Intrinsic::dx_handle_fromBinding:
HasErrors |= lowerHandleFromBinding(F);
break;
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index 48907647c660f8..c43625755d6efc 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
define void @test_group_memory_barrier_with_group_sync() {
entry:
>From c056247010394feeebef8b97a33a4a9fe47050aa Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 15 Oct 2024 16:27:13 -0700
Subject: [PATCH 08/16] Format
---
llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 56dca69b42f1fe..649398fc5bd2b4 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -10,8 +10,8 @@
// opcodes in DirectX Intermediate Language (DXIL).
//===----------------------------------------------------------------------===//
-#include "DXILConstants.h"
#include "DXILIntrinsicExpansion.h"
+#include "DXILConstants.h"
#include "DirectX.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
>From 3d869fe60ad297ca0157e7c520f1455032c10a16 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Wed, 16 Oct 2024 17:48:25 -0700
Subject: [PATCH 09/16] Tablegen simple
---
llvm/include/llvm/IR/IntrinsicsDirectX.td | 1 -
llvm/lib/Target/DirectX/DXIL.td | 14 +++-
.../Target/DirectX/DXILIntrinsicExpansion.cpp | 26 --------
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 25 +++----
.../group_memory_barrier_with_group_sync.ll | 2 +-
llvm/utils/TableGen/DXILEmitter.cpp | 66 +++++++++++++++----
6 files changed, 80 insertions(+), 54 deletions(-)
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index d16a71f717b8db..dada426368995d 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -93,6 +93,5 @@ def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>
[LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [IntrNoMem]>;
def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
-def int_dx_memory_barrier : DefaultAttrsIntrinsic<[], [llvm_i32_ty], []>;
def int_dx_group_memory_barrier_with_group_sync : DefaultAttrsIntrinsic<[], [], []>;
}
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index d00253b0b7c03e..c2392b8f90eb7c 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -294,6 +294,11 @@ class Attributes<Version ver = DXIL1_0, list<DXILAttribute> attrs> {
list<DXILAttribute> op_attrs = attrs;
}
+class IntrinsicSelect<Intrinsic intr, list<string> extra_args> {
+ Intrinsic Intr = intr;
+ list<string> ExtraArgs = extra_args;
+}
+
// Abstraction DXIL Operation
class DXILOp<int opcode, DXILOpClass opclass> {
// A short description of the operation
@@ -308,6 +313,8 @@ class DXILOp<int opcode, DXILOpClass opclass> {
// LLVM Intrinsic DXIL Operation maps to
Intrinsic LLVMIntrinsic = ?;
+ list<IntrinsicSelect> intrinsic_selects = [];
+
// Result type of the op
DXILOpParamType result;
@@ -832,7 +839,12 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
def Barrier : DXILOp<80, barrier> {
let Doc = "inserts a memory barrier in the shader";
- let LLVMIntrinsic = int_dx_memory_barrier;
+ let intrinsic_selects = [
+ IntrinsicSelect<
+ int_dx_group_memory_barrier_with_group_sync,
+ [ "OpBuilder.getIRB().getInt32((unsigned)BarrierMode::SyncThreadGroup | (unsigned)BarrierMode::TGSMFence)" ]>,
+ ];
+
let arguments = [Int32Ty];
let result = VoidTy;
let stages = [Stages<DXIL1_0, [compute, library]>];
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 649398fc5bd2b4..fb5383b3514a5a 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -11,7 +11,6 @@
//===----------------------------------------------------------------------===//
#include "DXILIntrinsicExpansion.h"
-#include "DXILConstants.h"
#include "DirectX.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
@@ -67,7 +66,6 @@ static bool isIntrinsicExpansion(Function &F) {
case Intrinsic::dx_sign:
case Intrinsic::dx_step:
case Intrinsic::dx_radians:
- case Intrinsic::dx_group_memory_barrier_with_group_sync:
return true;
}
return false;
@@ -454,27 +452,6 @@ static Value *expandRadiansIntrinsic(CallInst *Orig) {
return Builder.CreateFMul(X, PiOver180);
}
-static Value *expandMemoryBarrier(CallInst *Orig, Intrinsic::ID IntrinsicId) {
- assert(IntrinsicId == Intrinsic::dx_group_memory_barrier_with_group_sync);
- unsigned BarrierMode = 0;
- switch (IntrinsicId) {
- case Intrinsic::dx_group_memory_barrier_with_group_sync:
- BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence |
- (unsigned)dxil::BarrierMode::SyncThreadGroup;
- break;
- default:
- report_fatal_error(Twine("Unexpected memory barrier intrinsic."),
- /* gen_crash_diag=*/false);
- break;
- }
-
- IRBuilder<> Builder(Orig);
- return Builder.CreateIntrinsic(
- Builder.getVoidTy(), Intrinsic::dx_memory_barrier,
- ArrayRef<Value *>{Builder.getInt32(BarrierMode)}, nullptr,
- Orig->getName());
-}
-
static Intrinsic::ID getMaxForClamp(Type *ElemTy,
Intrinsic::ID ClampIntrinsic) {
if (ClampIntrinsic == Intrinsic::dx_uclamp)
@@ -609,9 +586,6 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
case Intrinsic::dx_radians:
Result = expandRadiansIntrinsic(Orig);
break;
- case Intrinsic::dx_group_memory_barrier_with_group_sync:
- Result = expandMemoryBarrier(Orig, IntrinsicId);
- break;
}
if (Result) {
Orig->replaceAllUsesWith(Result);
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 8acc9c1efa08c0..73b43801eabfa0 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -106,20 +106,21 @@ class OpLowerer {
return false;
}
- [[nodiscard]]
- bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp) {
+ [[nodiscard]] bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp,
+ ArrayRef<Value *> ExtraArgs) {
bool IsVectorArgExpansion = isVectorArgExpansion(F);
return replaceFunction(F, [&](CallInst *CI) -> Error {
- SmallVector<Value *> Args;
- OpBuilder.getIRB().SetInsertPoint(CI);
+ SmallVector<Value *> NewArgs;
if (IsVectorArgExpansion) {
- SmallVector<Value *> NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
- Args.append(NewArgs.begin(), NewArgs.end());
+ NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
} else
- Args.append(CI->arg_begin(), CI->arg_end());
+ NewArgs.append(CI->arg_begin(), CI->arg_end());
- Expected<CallInst *> OpCall =
- OpBuilder.tryCreateOp(DXILOp, Args, CI->getName(), F.getReturnType());
+ NewArgs.append(ExtraArgs.begin(), ExtraArgs.end());
+
+ OpBuilder.getIRB().SetInsertPoint(CI);
+ Expected<CallInst *> OpCall = OpBuilder.tryCreateOp(
+ DXILOp, NewArgs, CI->getName(), F.getReturnType());
if (Error E = OpCall.takeError())
return E;
@@ -583,9 +584,9 @@ class OpLowerer {
switch (ID) {
default:
continue;
-#define DXIL_OP_INTRINSIC(OpCode, Intrin) \
- case Intrin: \
- HasErrors |= replaceFunctionWithOp(F, OpCode); \
+#define DXIL_OP_INTRINSIC(OpCode, Intrin, ExtraArgs) \
+ case Intrin: \
+ HasErrors |= replaceFunctionWithOp(F, OpCode, ExtraArgs); \
break;
#include "DXILOperation.inc"
case Intrinsic::dx_handle_fromBinding:
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index c43625755d6efc..cd588d464e302b 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
define void @test_group_memory_barrier_with_group_sync() {
entry:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index e74fc00015b404..5ab6b0edd02ab2 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -32,6 +32,11 @@ using namespace llvm::dxil;
namespace {
+struct DXILIntrinsicSelect {
+ StringRef Intrinsic;
+ SmallVector<StringRef, 4> ExtraArgs;
+};
+
struct DXILOperationDesc {
std::string OpName; // name of DXIL operation
int OpCode; // ID of DXIL operation
@@ -42,8 +47,7 @@ struct DXILOperationDesc {
SmallVector<const Record *> OverloadRecs;
SmallVector<const Record *> StageRecs;
SmallVector<const Record *> AttrRecs;
- StringRef Intrinsic; // The llvm intrinsic map to OpName. Default is "" which
- // means no map exists
+ SmallVector<DXILIntrinsicSelect> IntrinsicSelects;
SmallVector<StringRef, 4>
ShaderStages; // shader stages to which this applies, empty for all.
int OverloadParamIndex; // Index of parameter with overload type.
@@ -157,14 +161,43 @@ DXILOperationDesc::DXILOperationDesc(const Record *R) {
OpName);
}
- const RecordVal *RV = R->getValue("LLVMIntrinsic");
- if (RV && RV->getValue()) {
- if (const DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
- auto *IntrinsicDef = DI->getDef();
- auto DefName = IntrinsicDef->getName();
- assert(DefName.starts_with("int_") && "invalid intrinsic name");
- // Remove the int_ from intrinsic name.
- Intrinsic = DefName.substr(4);
+ auto GetIntrinsicName = [](const RecordVal *RV) -> StringRef {
+ if (RV && RV->getValue()) {
+ if (const DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
+ auto *IntrinsicDef = DI->getDef();
+ auto DefName = IntrinsicDef->getName();
+ assert(DefName.starts_with("int_") && "invalid intrinsic name");
+ // Remove the int_ from intrinsic name.
+ return DefName.substr(4);
+ }
+ }
+ return "";
+ };
+
+ {
+ DXILIntrinsicSelect IntrSelect;
+ IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("LLVMIntrinsic"));
+ if (IntrSelect.Intrinsic.size())
+ IntrinsicSelects.emplace_back(std::move(IntrSelect));
+ }
+
+ Recs = R->getValueAsListOfDefs("intrinsic_selects");
+ if (Recs.size()) {
+ if (IntrinsicSelects.size()) {
+ PrintFatalError(R,
+ Twine("LLVMIntrinsic and intrinsic_match cannot be both "
+ "defined for DXIL operation - ") +
+ OpName);
+ } else {
+ for (const Record *R : Recs) {
+ DXILIntrinsicSelect IntrSelect;
+ IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("Intr"));
+ auto ExtraArgs = R->getValueAsListOfStrings("ExtraArgs");
+ for (StringRef Arg : ExtraArgs) {
+ IntrSelect.ExtraArgs.push_back(Arg);
+ }
+ IntrinsicSelects.emplace_back(std::move(IntrSelect));
+ }
}
}
}
@@ -377,10 +410,17 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
OS << "#ifdef DXIL_OP_INTRINSIC\n";
OS << "\n";
for (const auto &Op : Ops) {
- if (Op.Intrinsic.empty())
+ if (Op.IntrinsicSelects.empty()) {
continue;
- OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
- << ", Intrinsic::" << Op.Intrinsic << ")\n";
+ }
+ for (const DXILIntrinsicSelect &MappedIntr : Op.IntrinsicSelects) {
+ OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
+ << ", Intrinsic::" << MappedIntr.Intrinsic << ", (ArrayRef<Value *> {";
+ for (const StringRef &Arg : MappedIntr.ExtraArgs) {
+ OS << Arg << ", ";
+ }
+ OS << "}))\n";
+ }
}
OS << "\n";
OS << "#undef DXIL_OP_INTRINSIC\n";
>From c6dcd202d1157ed9300d66e6af7306acb0999bc8 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 21 Oct 2024 10:53:23 -0700
Subject: [PATCH 10/16] All working now
---
llvm/lib/Target/DirectX/DXIL.td | 40 +++++++-
llvm/lib/Target/DirectX/DXILConstants.h | 7 --
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 44 +++++++--
llvm/utils/TableGen/DXILEmitter.cpp | 104 +++++++++++++++------
4 files changed, 146 insertions(+), 49 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index c2392b8f90eb7c..57e5d3a6f1d5e7 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -294,9 +294,41 @@ class Attributes<Version ver = DXIL1_0, list<DXILAttribute> attrs> {
list<DXILAttribute> op_attrs = attrs;
}
-class IntrinsicSelect<Intrinsic intr, list<string> extra_args> {
- Intrinsic Intr = intr;
- list<string> ExtraArgs = extra_args;
+class DXILConstant<int value_> {
+ int value = value_;
+}
+
+defset list<DXILConstant> BarrierModes = {
+ def BarrierMode_DeviceMemoryBarrier : DXILConstant<2>;
+ def BarrierMode_DeviceMemoryBarrierWithGroupSync : DXILConstant<3>;
+ def BarrierMode_GroupMemoryBarrier : DXILConstant<8>;
+ def BarrierMode_GroupMemoryBarrierWithGroupSync : DXILConstant<9>;
+ def BarrierMode_AllMemoryBarrier : DXILConstant<10>;
+ def BarrierMode_AllMemoryBarrierWithGroupSync : DXILConstant<11>;
+}
+
+// Intrinsic arg selection
+class Arg {
+ int index = -1;
+ DXILConstant value;
+ bit is_i8 = 0;
+ bit is_i32 = 0;
+}
+class ArgSelect<int index_> : Arg {
+ let index = index_;
+}
+class ArgI32<DXILConstant value_> : Arg {
+ let value = value_;
+ let is_i32 = 1;
+}
+class ArgI8<DXILConstant value_> : Arg {
+ let value = value_;
+ let is_i8 = 1;
+}
+
+class IntrinsicSelect<Intrinsic intrinsic_, list<Arg> args_=?> {
+ Intrinsic intrinsic = intrinsic_;
+ list<Arg> args = args_;
}
// Abstraction DXIL Operation
@@ -842,7 +874,7 @@ def Barrier : DXILOp<80, barrier> {
let intrinsic_selects = [
IntrinsicSelect<
int_dx_group_memory_barrier_with_group_sync,
- [ "OpBuilder.getIRB().getInt32((unsigned)BarrierMode::SyncThreadGroup | (unsigned)BarrierMode::TGSMFence)" ]>,
+ [ ArgI32<BarrierMode_GroupMemoryBarrierWithGroupSync> ]>,
];
let arguments = [Int32Ty];
diff --git a/llvm/lib/Target/DirectX/DXILConstants.h b/llvm/lib/Target/DirectX/DXILConstants.h
index 38984727761bb3..022cd57795a063 100644
--- a/llvm/lib/Target/DirectX/DXILConstants.h
+++ b/llvm/lib/Target/DirectX/DXILConstants.h
@@ -30,13 +30,6 @@ enum class OpParamType : unsigned {
#include "DXILOperation.inc"
};
-enum class BarrierMode : unsigned {
- SyncThreadGroup = 0x00000001,
- UAVFenceGlobal = 0x00000002,
- UAVFenceThreadGroup = 0x00000004,
- TGSMFence = 0x00000008,
-};
-
} // namespace dxil
} // namespace llvm
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 73b43801eabfa0..2c9791ff7fd739 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -106,19 +106,44 @@ class OpLowerer {
return false;
}
+ struct Arg {
+ enum class Type {
+ Index,
+ I8,
+ I32,
+ };
+ Type Type = Type::Index;
+ int Value = -1;
+ };
+
[[nodiscard]] bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp,
- ArrayRef<Value *> ExtraArgs) {
+ ArrayRef<Arg> Args) {
bool IsVectorArgExpansion = isVectorArgExpansion(F);
return replaceFunction(F, [&](CallInst *CI) -> Error {
+ OpBuilder.getIRB().SetInsertPoint(CI);
SmallVector<Value *> NewArgs;
- if (IsVectorArgExpansion) {
+ if (Args.size()) {
+ for (const Arg &A : Args) {
+ switch (A.Type) {
+ case Arg::Type::Index:
+ NewArgs.push_back(CI->getArgOperand(A.Value));
+ break;
+ case Arg::Type::I8:
+ NewArgs.push_back(OpBuilder.getIRB().getInt8((uint8_t)A.Value));
+ break;
+ case Arg::Type::I32:
+ NewArgs.push_back(OpBuilder.getIRB().getInt32(A.Value));
+ break;
+ default:
+ llvm_unreachable("Invalid type of intrinsic arg.");
+ }
+ }
+ } else if (IsVectorArgExpansion) {
NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
- } else
+ } else {
NewArgs.append(CI->arg_begin(), CI->arg_end());
+ }
- NewArgs.append(ExtraArgs.begin(), ExtraArgs.end());
-
- OpBuilder.getIRB().SetInsertPoint(CI);
Expected<CallInst *> OpCall = OpBuilder.tryCreateOp(
DXILOp, NewArgs, CI->getName(), F.getReturnType());
if (Error E = OpCall.takeError())
@@ -584,9 +609,10 @@ class OpLowerer {
switch (ID) {
default:
continue;
-#define DXIL_OP_INTRINSIC(OpCode, Intrin, ExtraArgs) \
- case Intrin: \
- HasErrors |= replaceFunctionWithOp(F, OpCode, ExtraArgs); \
+#define DXIL_OP_INTRINSIC(OpCode, Intrin, ...) \
+ case Intrin: \
+ HasErrors |= replaceFunctionWithOp(F, OpCode, \
+ ArrayRef<Arg>{ __VA_ARGS__ }); \
break;
#include "DXILOperation.inc"
case Intrinsic::dx_handle_fromBinding:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 5ab6b0edd02ab2..06f5f89615c6f3 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -32,9 +32,18 @@ using namespace llvm::dxil;
namespace {
+struct DXILArgSelect {
+ enum class Type {
+ Index,
+ I32,
+ I8,
+ };
+ Type Type = Type::Index;
+ int Value = 0;
+};
struct DXILIntrinsicSelect {
StringRef Intrinsic;
- SmallVector<StringRef, 4> ExtraArgs;
+ SmallVector<DXILArgSelect, 4> Args;
};
struct DXILOperationDesc {
@@ -75,6 +84,21 @@ static void ascendingSortByVersion(std::vector<const Record *> &Recs) {
});
}
+/// Take a `int_{intrinsic_name}` and return just the intrinsic_name part if available.
+/// Otherwise return the empty string.
+static StringRef GetIntrinsicName(const RecordVal *RV){
+ if (RV && RV->getValue()) {
+ if (const DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
+ auto *IntrinsicDef = DI->getDef();
+ auto DefName = IntrinsicDef->getName();
+ assert(DefName.starts_with("int_") && "invalid intrinsic name");
+ // Remove the int_ from intrinsic name.
+ return DefName.substr(4);
+ }
+ }
+ return "";
+}
+
/// Construct an object using the DXIL Operation records specified
/// in DXIL.td. This serves as the single source of reference of
/// the information extracted from the specified Record R, for
@@ -161,19 +185,6 @@ DXILOperationDesc::DXILOperationDesc(const Record *R) {
OpName);
}
- auto GetIntrinsicName = [](const RecordVal *RV) -> StringRef {
- if (RV && RV->getValue()) {
- if (const DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
- auto *IntrinsicDef = DI->getDef();
- auto DefName = IntrinsicDef->getName();
- assert(DefName.starts_with("int_") && "invalid intrinsic name");
- // Remove the int_ from intrinsic name.
- return DefName.substr(4);
- }
- }
- return "";
- };
-
{
DXILIntrinsicSelect IntrSelect;
IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("LLVMIntrinsic"));
@@ -181,20 +192,43 @@ DXILOperationDesc::DXILOperationDesc(const Record *R) {
IntrinsicSelects.emplace_back(std::move(IntrSelect));
}
- Recs = R->getValueAsListOfDefs("intrinsic_selects");
- if (Recs.size()) {
+ auto IntrinsicSelectRecords = R->getValueAsListOfDefs("intrinsic_selects");
+ if (IntrinsicSelectRecords.size()) {
if (IntrinsicSelects.size()) {
- PrintFatalError(R,
- Twine("LLVMIntrinsic and intrinsic_match cannot be both "
- "defined for DXIL operation - ") +
- OpName);
+ PrintFatalError(
+ R, Twine("LLVMIntrinsic and intrinsic_selects cannot be both "
+ "defined for DXIL operation - ") +
+ OpName);
} else {
- for (const Record *R : Recs) {
+ for (const Record *R : IntrinsicSelectRecords) {
DXILIntrinsicSelect IntrSelect;
- IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("Intr"));
- auto ExtraArgs = R->getValueAsListOfStrings("ExtraArgs");
- for (StringRef Arg : ExtraArgs) {
- IntrSelect.ExtraArgs.push_back(Arg);
+ IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("intrinsic"));
+ auto Args = R->getValueAsListOfDefs("args");
+ for (const Record *Arg : Args) {
+ bool IsI8 = Arg->getValueAsBit("is_i8");
+ bool IsI32 = Arg->getValueAsBit("is_i32");
+ int Index = Arg->getValueAsInt("index");
+ const Record *ValueRec = Arg->getValueAsDef("value");
+
+ DXILArgSelect ArgSelect;
+ if (IsI8) {
+ ArgSelect.Type = DXILArgSelect::Type::I8;
+ ArgSelect.Value = ValueRec->getValueAsInt("value");
+ } else if (IsI32) {
+ ArgSelect.Type = DXILArgSelect::Type::I32;
+ ArgSelect.Value = ValueRec->getValueAsInt("value");
+ } else {
+ if (Index < 0) {
+ PrintFatalError(
+ R, Twine("Index in ArgSelect<index> must be equal to or "
+ "greater than 0 for DXIL operation - ") +
+ OpName);
+ }
+ ArgSelect.Type = DXILArgSelect::Type::Index;
+ ArgSelect.Value = Index;
+ }
+
+ IntrSelect.Args.emplace_back(std::move(ArgSelect));
}
IntrinsicSelects.emplace_back(std::move(IntrSelect));
}
@@ -415,11 +449,23 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
}
for (const DXILIntrinsicSelect &MappedIntr : Op.IntrinsicSelects) {
OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
- << ", Intrinsic::" << MappedIntr.Intrinsic << ", (ArrayRef<Value *> {";
- for (const StringRef &Arg : MappedIntr.ExtraArgs) {
- OS << Arg << ", ";
+ << ", Intrinsic::" << MappedIntr.Intrinsic;
+ for (const DXILArgSelect &ArgSelect : MappedIntr.Args) {
+ OS << ", (Arg { ";
+ switch (ArgSelect.Type) {
+ case DXILArgSelect::Type::Index:
+ OS << "Arg::Type::Index, ";
+ break;
+ case DXILArgSelect::Type::I8:
+ OS << "Arg::Type::I8, ";
+ break;
+ case DXILArgSelect::Type::I32:
+ OS << "Arg::Type::I32, ";
+ break;
+ }
+ OS << ArgSelect.Value << "})";
}
- OS << "}))\n";
+ OS << ")\n";
}
}
OS << "\n";
>From 661190139b041239388ab7f4f196ea6bd24b93cf Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 21 Oct 2024 11:17:01 -0700
Subject: [PATCH 11/16] Rename and clang-format
---
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 22 +++++++++++-----------
llvm/utils/TableGen/DXILEmitter.cpp | 8 ++++----
2 files changed, 15 insertions(+), 15 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 2c9791ff7fd739..78ae366001bb0c 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -106,7 +106,7 @@ class OpLowerer {
return false;
}
- struct Arg {
+ struct ArgSelect {
enum class Type {
Index,
I8,
@@ -117,25 +117,25 @@ class OpLowerer {
};
[[nodiscard]] bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp,
- ArrayRef<Arg> Args) {
+ ArrayRef<ArgSelect> Args) {
bool IsVectorArgExpansion = isVectorArgExpansion(F);
return replaceFunction(F, [&](CallInst *CI) -> Error {
OpBuilder.getIRB().SetInsertPoint(CI);
SmallVector<Value *> NewArgs;
if (Args.size()) {
- for (const Arg &A : Args) {
+ for (const ArgSelect &A : Args) {
switch (A.Type) {
- case Arg::Type::Index:
+ case ArgSelect::Type::Index:
NewArgs.push_back(CI->getArgOperand(A.Value));
break;
- case Arg::Type::I8:
+ case ArgSelect::Type::I8:
NewArgs.push_back(OpBuilder.getIRB().getInt8((uint8_t)A.Value));
break;
- case Arg::Type::I32:
+ case ArgSelect::Type::I32:
NewArgs.push_back(OpBuilder.getIRB().getInt32(A.Value));
break;
default:
- llvm_unreachable("Invalid type of intrinsic arg.");
+ llvm_unreachable("Invalid type of intrinsic arg select.");
}
}
} else if (IsVectorArgExpansion) {
@@ -609,10 +609,10 @@ class OpLowerer {
switch (ID) {
default:
continue;
-#define DXIL_OP_INTRINSIC(OpCode, Intrin, ...) \
- case Intrin: \
- HasErrors |= replaceFunctionWithOp(F, OpCode, \
- ArrayRef<Arg>{ __VA_ARGS__ }); \
+#define DXIL_OP_INTRINSIC(OpCode, Intrin, ...) \
+ case Intrin: \
+ HasErrors |= replaceFunctionWithOp(F, OpCode, \
+ ArrayRef<ArgSelect>{ __VA_ARGS__ }); \
break;
#include "DXILOperation.inc"
case Intrinsic::dx_handle_fromBinding:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 06f5f89615c6f3..540af4869787ac 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -451,16 +451,16 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
<< ", Intrinsic::" << MappedIntr.Intrinsic;
for (const DXILArgSelect &ArgSelect : MappedIntr.Args) {
- OS << ", (Arg { ";
+ OS << ", (ArgSelect { ";
switch (ArgSelect.Type) {
case DXILArgSelect::Type::Index:
- OS << "Arg::Type::Index, ";
+ OS << "ArgSelect::Type::Index, ";
break;
case DXILArgSelect::Type::I8:
- OS << "Arg::Type::I8, ";
+ OS << "ArgSelect::Type::I8, ";
break;
case DXILArgSelect::Type::I32:
- OS << "Arg::Type::I32, ";
+ OS << "ArgSelect::Type::I32, ";
break;
}
OS << ArgSelect.Value << "})";
>From d6a9e27f62d57e9e328064d5eba2757ff420b033 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 21 Oct 2024 11:21:25 -0700
Subject: [PATCH 12/16] Comment
---
llvm/lib/Target/DirectX/DXIL.td | 1 +
1 file changed, 1 insertion(+)
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index 57e5d3a6f1d5e7..89aae5e49be0c4 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -345,6 +345,7 @@ class DXILOp<int opcode, DXILOpClass opclass> {
// LLVM Intrinsic DXIL Operation maps to
Intrinsic LLVMIntrinsic = ?;
+ // Non-trivial LLVM Intrinsics DXIL Operation maps to
list<IntrinsicSelect> intrinsic_selects = [];
// Result type of the op
>From 39ce663c2817ca707a313c5e05a58b6d3251343b Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 21 Oct 2024 12:21:10 -0700
Subject: [PATCH 13/16] Handling the arg select version for other ops
---
llvm/lib/Target/DirectX/DXIL.td | 2 +-
llvm/utils/TableGen/DXILEmitter.cpp | 12 +++++++++++-
2 files changed, 12 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index 89aae5e49be0c4..263ca50011aa7b 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -326,7 +326,7 @@ class ArgI8<DXILConstant value_> : Arg {
let is_i8 = 1;
}
-class IntrinsicSelect<Intrinsic intrinsic_, list<Arg> args_=?> {
+class IntrinsicSelect<Intrinsic intrinsic_, list<Arg> args_> {
Intrinsic intrinsic = intrinsic_;
list<Arg> args = args_;
}
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 540af4869787ac..d68d19320c803a 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -208,13 +208,23 @@ DXILOperationDesc::DXILOperationDesc(const Record *R) {
bool IsI8 = Arg->getValueAsBit("is_i8");
bool IsI32 = Arg->getValueAsBit("is_i32");
int Index = Arg->getValueAsInt("index");
- const Record *ValueRec = Arg->getValueAsDef("value");
+ const Record *ValueRec = Arg->getValueAsOptionalDef("value");
DXILArgSelect ArgSelect;
if (IsI8) {
+ if (!ValueRec) {
+ PrintFatalError(R, Twine("'value' must be defined for i8 "
+ "ArgSelect for DXIL operation - ") +
+ OpName);
+ }
ArgSelect.Type = DXILArgSelect::Type::I8;
ArgSelect.Value = ValueRec->getValueAsInt("value");
} else if (IsI32) {
+ if (!ValueRec) {
+ PrintFatalError(R, Twine("'value' must be defined for i32 "
+ "ArgSelect for DXIL operation - ") +
+ OpName);
+ }
ArgSelect.Type = DXILArgSelect::Type::I32;
ArgSelect.Value = ValueRec->getValueAsInt("value");
} else {
>From 51f6632da871531832a1ebc13f1c2500bcdd2261 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 21 Oct 2024 12:43:53 -0700
Subject: [PATCH 14/16] Formatting
---
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 4 ++--
llvm/utils/TableGen/DXILEmitter.cpp | 6 +++---
2 files changed, 5 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 78ae366001bb0c..1ee03b622a7145 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -611,8 +611,8 @@ class OpLowerer {
continue;
#define DXIL_OP_INTRINSIC(OpCode, Intrin, ...) \
case Intrin: \
- HasErrors |= replaceFunctionWithOp(F, OpCode, \
- ArrayRef<ArgSelect>{ __VA_ARGS__ }); \
+ HasErrors |= \
+ replaceFunctionWithOp(F, OpCode, ArrayRef<ArgSelect>{__VA_ARGS__}); \
break;
#include "DXILOperation.inc"
case Intrinsic::dx_handle_fromBinding:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index d68d19320c803a..011a5cad186492 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -84,9 +84,9 @@ static void ascendingSortByVersion(std::vector<const Record *> &Recs) {
});
}
-/// Take a `int_{intrinsic_name}` and return just the intrinsic_name part if available.
-/// Otherwise return the empty string.
-static StringRef GetIntrinsicName(const RecordVal *RV){
+/// Take a `int_{intrinsic_name}` and return just the intrinsic_name part if
+/// available. Otherwise return the empty string.
+static StringRef GetIntrinsicName(const RecordVal *RV) {
if (RV && RV->getValue()) {
if (const DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
auto *IntrinsicDef = DI->getDef();
>From 29180abc3585c46e06a7ba3f501ecbdbb58ae605 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Mon, 28 Oct 2024 14:39:41 -0700
Subject: [PATCH 15/16] Addressed feedback
---
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 22 +++++++++----------
.../group_memory_barrier_with_group_sync.ll | 2 +-
llvm/utils/TableGen/DXILEmitter.cpp | 2 +-
3 files changed, 13 insertions(+), 13 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 1ee03b622a7145..b5cf1654181c6c 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -117,35 +117,35 @@ class OpLowerer {
};
[[nodiscard]] bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp,
- ArrayRef<ArgSelect> Args) {
+ ArrayRef<ArgSelect> ArgSelects) {
bool IsVectorArgExpansion = isVectorArgExpansion(F);
return replaceFunction(F, [&](CallInst *CI) -> Error {
OpBuilder.getIRB().SetInsertPoint(CI);
- SmallVector<Value *> NewArgs;
- if (Args.size()) {
- for (const ArgSelect &A : Args) {
+ SmallVector<Value *> Args;
+ if (ArgSelects.size()) {
+ for (const ArgSelect &A : ArgSelects) {
switch (A.Type) {
case ArgSelect::Type::Index:
- NewArgs.push_back(CI->getArgOperand(A.Value));
+ Args.push_back(CI->getArgOperand(A.Value));
break;
case ArgSelect::Type::I8:
- NewArgs.push_back(OpBuilder.getIRB().getInt8((uint8_t)A.Value));
+ Args.push_back(OpBuilder.getIRB().getInt8((uint8_t)A.Value));
break;
case ArgSelect::Type::I32:
- NewArgs.push_back(OpBuilder.getIRB().getInt32(A.Value));
+ Args.push_back(OpBuilder.getIRB().getInt32(A.Value));
break;
default:
llvm_unreachable("Invalid type of intrinsic arg select.");
}
}
} else if (IsVectorArgExpansion) {
- NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
+ Args = argVectorFlatten(CI, OpBuilder.getIRB());
} else {
- NewArgs.append(CI->arg_begin(), CI->arg_end());
+ Args.append(CI->arg_begin(), CI->arg_end());
}
- Expected<CallInst *> OpCall = OpBuilder.tryCreateOp(
- DXILOp, NewArgs, CI->getName(), F.getReturnType());
+ Expected<CallInst *> OpCall =
+ OpBuilder.tryCreateOp(DXILOp, Args, CI->getName(), F.getReturnType());
if (Error E = OpCall.takeError())
return E;
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index cd588d464e302b..baf93d4e177f0f 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s
define void @test_group_memory_barrier_with_group_sync() {
entry:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 011a5cad186492..8bebe608eece47 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -39,7 +39,7 @@ struct DXILArgSelect {
I8,
};
Type Type = Type::Index;
- int Value = 0;
+ int Value = -1;
};
struct DXILIntrinsicSelect {
StringRef Intrinsic;
>From 6b7e6b4ad43244f8f64bd37b4a5c7477624cb1f7 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Wed, 30 Oct 2024 20:47:05 -0700
Subject: [PATCH 16/16] Fixed the warnings on clang
---
llvm/lib/Target/DirectX/DXILOpLowering.cpp | 2 --
llvm/utils/TableGen/DXILEmitter.cpp | 6 +++---
2 files changed, 3 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index b5cf1654181c6c..5f278b609690e2 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -134,8 +134,6 @@ class OpLowerer {
case ArgSelect::Type::I32:
Args.push_back(OpBuilder.getIRB().getInt32(A.Value));
break;
- default:
- llvm_unreachable("Invalid type of intrinsic arg select.");
}
}
} else if (IsVectorArgExpansion) {
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 8bebe608eece47..4656438d64c80a 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -459,9 +459,9 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
}
for (const DXILIntrinsicSelect &MappedIntr : Op.IntrinsicSelects) {
OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
- << ", Intrinsic::" << MappedIntr.Intrinsic;
+ << ", Intrinsic::" << MappedIntr.Intrinsic << ", ";
for (const DXILArgSelect &ArgSelect : MappedIntr.Args) {
- OS << ", (ArgSelect { ";
+ OS << "(ArgSelect { ";
switch (ArgSelect.Type) {
case DXILArgSelect::Type::Index:
OS << "ArgSelect::Type::Index, ";
@@ -473,7 +473,7 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
OS << "ArgSelect::Type::I32, ";
break;
}
- OS << ArgSelect.Value << "})";
+ OS << ArgSelect.Value << "}), ";
}
OS << ")\n";
}
More information about the llvm-commits
mailing list