Update llpc from commit 495b235c

[Continuations] Refactor 'TypeLowering::lowerFunctionArguments'
Fix performance gap caused by LDS bank conflict
[CompilerUtils] Assert return types are not changed
[CompilerUtils] Fixup alloca alignment in type lowering
[llvmraytracing] Remove plugin pass info function for dynamically loaded plugins
[CompilerUtils] Add LLVM plugin
Remove deprecated assert
Cleanup useless VgtGsVertItemsize calculation assignment
[Continuations] Support '_AmdIsLlpc' intrinsic
Remove deprecated interface forceEnablePrimStats
Fix build failure due to llvm memory issue
[Continuations] Introduce 'RCR' argument for 'lgc.cps.jump', refactor wait mask handling
Add support for no return image sample intrinsics in builder
[Continuations] Improve payload size reporting
lgc : replace threadMaskedSelect function to use inverseballot
[Continuations] Move 'lgc.ilcps.return' op building to 'LowerRaytracingPipeline'
Compiler::buildShaderCacheHash: Fix fragment hash
[Continuations] Improve stats reporting
Force wave64 mode for legacy GS on GFX10
Change in/out semantics representation
[Continuations] Remove payload size reporting in bytes
Fix component addressing issue of 64-bit FS input
lgc : fix logic issues of tryOptimizeWorkgroupId
Implement the mapping from CLIP_DIST_4/5/6/7 to CLIP_DIST_0/1/2/3 accordingly
[Continuations] Add 'complete' intrinsic, generate 'lgc.cps.jump'
Disable AllowReassoc for reflect operations
Correct processing of mesh shader output semantics
[RT] Don't compile kernel entry for pipeline library
lgc : fix global pointer access for buffer strided pointer
[RT] Support AmdTraceRayInitStaticId
Record debugprintf usermapping to the elf
[Continuations] Add payload serialization alloca for Traversal in lgc.cps mode
[Continuations] Use 'lgc.cps.complete' for program termination
Rename some classes and files in lower pass
Add strix1 support

.clang-format

@@ -4,10 +4,12 @@ AllowShortFunctionsOnASingleLine: InlineOnly
 IncludeBlocks: Merge
 IncludeCategories:
   - Regex:           '^"lgc/'
-    Priority:        2
-  - Regex:           '^"(llvm|llvm-c|llvm-dialects|clang|clang-c)/'
     Priority:        3
-  - Regex:           '^(<|"(gtest|gmock|isl|json)/)'
+  - Regex:           '^"(llvm|llvm-c|llvm-dialects|clang|clang-c)/'
     Priority:        4
+  - Regex:           '^(<|"(gtest|gmock|isl|json)/)'
+    Priority:        5
+  - Regex:           '.*/'
+    Priority:        2
   - Regex:           '.*'
     Priority:        1

.typos.toml

@@ -13,6 +13,7 @@ UE = "UE"
 OpBuildNDRange = "OpBuildNDRange"
 serDataNode = "serDataNode"
 rcall = "rcall"
+Fo = "Fo"
 
 [default.extend-words]
 ba = "ba"
@@ -28,4 +29,3 @@ dne = "dne"
 offen = "offen"
 varing = "varing"
 Derivate = "Derivate"
-Fo = "Fo"

compilerutils/CMakeLists.txt

@@ -14,6 +14,7 @@ endfunction()
 add_llvm_library(LLVMCompilerUtils
   lib/ArgPromotion.cpp
   lib/CompilerUtils.cpp
+  lib/DxilToLlvm.cpp
   lib/TypeLowering.cpp
   lib/TypesMetadata.cpp
 
@@ -39,5 +40,6 @@ set_compiler_options(LLVMCompilerUtils)
 target_compile_features(LLVMCompilerUtils PUBLIC cxx_std_17)
 set_target_properties(LLVMCompilerUtils PROPERTIES CXX_EXTENSIONS OFF)
 
+add_subdirectory(plugin)
 add_subdirectory(tool/cross-module-inline)
 add_subdirectory(test)

compilerutils/include/compilerutils/CompilerUtils.h

@@ -38,8 +38,17 @@
 #include "llvm/IR/Attributes.h"
 #include "llvm/IR/IRBuilder.h"
 
+namespace llvm {
+
+class PassBuilder;
+
+} // namespace llvm
+
 namespace CompilerUtils {
 
+// Register compiler utils passes.
+void RegisterPasses(llvm::PassBuilder &PB);
+
 // Create an LLVM function call to the named function. The callee is built
 // automatically based on return type and its parameters.
 //

compilerutils/include/compilerutils/DxilToLlvm.h

@@ -0,0 +1,78 @@
+/*
+ ***********************************************************************************************************************
+ *
+ *  Copyright (c) 2024 Advanced Micro Devices, Inc. All Rights Reserved.
+ *
+ *  Permission is hereby granted, free of charge, to any person obtaining a copy
+ *  of this software and associated documentation files (the "Software"), to
+ *  deal in the Software without restriction, including without limitation the
+ *  rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ *  sell copies of the Software, and to permit persons to whom the Software is
+ *  furnished to do so, subject to the following conditions:
+ *
+ *  The above copyright notice and this permission notice shall be included in all
+ *  copies or substantial portions of the Software.
+ *
+ *  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ *  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ *  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ *  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ *  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ *  FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ *  IN THE SOFTWARE.
+ *
+ **********************************************************************************************************************/
+
+//===- DxilToLlvm.h -  --------------------------------------------------------------------------------------------===//
+//
+// This pass converts a DXIL module into an LLVM module by fixing constructs that have different semantics in the two.
+// The output module will still contain DXIL intrinsics and metadata, because we only fix incompatibilities, and don't
+// lower away DXIL.
+//
+// The following modifications are made:
+//
+// * i1 vectors are replaced by i32 vectors
+//   This works around a more general difference between DXIL and LLVM:
+//   In LLVM, vectors are always bit-packed and ignore the elements' alignment.
+//   In DXIL, vectors respect the elements' alignment, and i1s have 32-bit alignment.
+//   Thus, in DXIL, the elements of <2 x i1> are 32 bits apart, while they are bit-packed in LLVM,
+//   and DXC relies on this by bit-casting allocas between <2 x i1> and <2 x i32>.
+//   This only seems to affect HLSL i1 *matrices*, which are lowered to arrays of i1 vectors in DXIL,
+//   and not HLSL i1 vectors, which are lowered to i32 arrays in DXIL.
+//   To fix this, we replace all i1 vectors by i32 vectors.
+//   We don't apply the same to other vectors that were overaligned in the original DXIL data layout (e.g. i16)
+//   because this may harm performance, and we haven't observed cases yet where DXC relies on this layout.
+//   See https://github.com/microsoft/DirectXShaderCompiler/issues/6082 for some background.
+//
+// Further known, not yet handled differences:
+//
+// * vectors of non-i1 elements that are overaligned in DXIL (see above)
+// * potentially: overaligned types in general
+//   After importing DXIL modules, we change the data layout to match what the backend does. Doing so potentially
+//   breaks the module if it relies on the existing DL. For instance, after changing the alignment of i16 from 32 to 16,
+//   storing as [4 x i16] and reading back the second dword behaves differently. Strictly speaking, when changing the
+//   DL, we would need to update such occurrences. We don't do that because we haven't yet observed such cases, and
+//   because it is difficult in general. For instance, we could transparently replace i16s by i32s to preserve the
+//   32-bit size, but replacing half by float is more problematic. Although there is no spec, it appears DXC tries to
+//   emit DXIL that supports such DL changes, by only using structured GEPs and avoiding transformations based on byte
+//   offsets. Also, such fixups are only possible locally (e.g. for allocas), and not through opaque memory.
+// * UDiv/URem/FPTrunc differences
+// * fast math flags
+//
+//===--------------------------------------------------------------------------------------------------------------===//
+
+#pragma once
+
+#include "llvm/IR/PassManager.h"
+#include "llvm/Pass.h"
+
+namespace CompilerUtils {
+
+class DxilToLlvmPass : public llvm::PassInfoMixin<DxilToLlvmPass> {
+public:
+  llvm::PreservedAnalyses run(llvm::Module &Module, llvm::ModuleAnalysisManager &AnalysisManager);
+
+  static llvm::StringRef name() { return "Convert DXIL to LLVM IR"; }
+};
+
+} // namespace CompilerUtils

compilerutils/lib/CompilerUtils.cpp

@@ -24,6 +24,7 @@
  **********************************************************************************************************************/
 
 #include "compilerutils/CompilerUtils.h"
+#include "compilerutils/DxilToLlvm.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/iterator_range.h"
 #include "llvm/IR/Attributes.h"
@@ -32,6 +33,7 @@
 #include "llvm/IR/Module.h"
 #include "llvm/IR/Type.h"
 #include "llvm/IR/Value.h"
+#include "llvm/Passes/PassBuilder.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/Cloning.h"
@@ -544,3 +546,20 @@ void CompilerUtils::replaceAllPointerUses(IRBuilder<> *builder, Value *oldPointe
   assert(PhiElems.empty() && "All phi inputs need to be handled, otherwise we end in an inconsistent state");
 #endif
 }
+
+void CompilerUtils::RegisterPasses(llvm::PassBuilder &PB) {
+#define HANDLE_PASS(NAME, CREATE_PASS)                                                                                 \
+  if (innerPipeline.empty() && name == NAME) {                                                                         \
+    passMgr.addPass(CREATE_PASS);                                                                                      \
+    return true;                                                                                                       \
+  }
+
+  PB.registerPipelineParsingCallback(
+      [](StringRef name, ModulePassManager &passMgr, ArrayRef<PassBuilder::PipelineElement> innerPipeline) {
+        StringRef Params;
+        (void)Params;
+#define COMPILERUTILS_PASS HANDLE_PASS
+#include "PassRegistry.inc"
+        return false;
+      });
+}