Merge remote-tracking branch 'upstream/main' into asynclinq

Author: stephentoub

eng/pipelines/runtime-ioslike.yml

@@ -4,6 +4,24 @@
 
 trigger: none
 
+# To reduce the load on the pipeline, enable it only for PRs that affect Mono LLVM related code.
+pr:
+  branches:
+    include:
+    - main
+    - release/*.*
+
+  paths:
+    include:
+      - src/mono/mono/mini/aot-*.*
+      - src/mono/mono/mini/llvm-*.*
+      - src/mono/mono/mini/mini-llvm-*.*
+      - src/mono/mono/mini/intrinsics.c
+      - src/mono/mono/mini/simd-*.*
+      - src/mono/mono/mini/decompose.c
+      - src/mono/mono/mini/method-to-ir.c
+      - src/mono/mono/mini/mini.c
+
 variables:
   - template: /eng/pipelines/common/variables.yml
 

eng/pipelines/runtime-llvm.yml

@@ -28,6 +28,24 @@ schedules:
       - main
     always: false # run only if there were changes since the last successful scheduled run.
 
+# To reduce the load on the pipeline, enable it only for PRs that affect Mono LLVM related code.
+pr:
+  branches:
+    include:
+    - main
+    - release/*.*
+
+  paths:
+    include:
+      - src/mono/mono/mini/aot-*.*
+      - src/mono/mono/mini/llvm-*.*
+      - src/mono/mono/mini/mini-llvm-*.*
+      - src/mono/mono/mini/intrinsics.c
+      - src/mono/mono/mini/simd-*.*
+      - src/mono/mono/mini/decompose.c
+      - src/mono/mono/mini/method-to-ir.c
+      - src/mono/mono/mini/mini.c
+
 variables:
   - template: /eng/pipelines/common/variables.yml
 

src/coreclr/CMakeLists.txt

@@ -170,10 +170,7 @@ endif(CLR_CMAKE_HOST_WIN32)
 #----------------------------------
 include(clrdefinitions.cmake)
 
-if(FEATURE_STANDALONE_GC)
-  add_definitions(-DFEATURE_STANDALONE_GC)
-  add_subdirectory(gc)
-endif(FEATURE_STANDALONE_GC)
+add_subdirectory(gc)
 
 if (CLR_CMAKE_HOST_UNIX)
   include_directories("pal/inc")

src/coreclr/gc/CMakeLists.txt

@@ -1,12 +1,10 @@
 set(CMAKE_INCLUDE_CURRENT_DIR ON)
 
-# Local GC meta-issue: https://github.com/dotnet/runtime/issues/8061
-
-# https://github.com/dotnet/runtime/issues/8059
-remove_definitions(-DSTRESS_HEAP)
-
-# https://github.com/dotnet/runtime/issues/8062
-remove_definitions(-DWRITE_BARRIER_CHECK)
+if(FEATURE_STANDALONE_GC)
+  add_definitions(-DFEATURE_STANDALONE_GC)
+  remove_definitions(-DSTRESS_HEAP)
+  remove_definitions(-DWRITE_BARRIER_CHECK)
+endif(FEATURE_STANDALONE_GC)
 
 set(GC_SOURCES
   gceventstatus.cpp
@@ -100,37 +98,39 @@ list(APPEND GC_SOURCES ${GC_HEADERS})
 
 convert_to_absolute_path(GC_SOURCES ${GC_SOURCES})
 
-# clrgcexp is build with standalone+regions
-if (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
-  add_library_clr(clrgcexp SHARED ${GC_SOURCES})
-  add_dependencies(clrgcexp eventing_headers)
-  target_link_libraries(clrgcexp PRIVATE ${GC_LINK_LIBRARIES})
-  target_compile_definitions(clrgcexp PRIVATE -DUSE_REGIONS)
-  install_clr(TARGETS clrgcexp DESTINATIONS . COMPONENT runtime)
-endif (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
-
-# clrgc is build with standalone+segments
-add_library_clr(clrgc SHARED ${GC_SOURCES})
-add_dependencies(clrgc eventing_headers)
-target_link_libraries(clrgc PRIVATE ${GC_LINK_LIBRARIES})
-install_clr(TARGETS clrgc DESTINATIONS . COMPONENT runtime)
-
-add_definitions(-DBUILD_AS_STANDALONE)
-add_definitions(-DFEATURE_CONSERVATIVE_GC)
-
-add_definitions(-DFX_VER_INTERNALNAME_STR=clrgc.dll)
-add_definitions(-DVERIFY_HEAP)
-if(CLR_CMAKE_HOST_APPLE)
-  # The implementation of GCToOSInterface on Apple platforms makes use of non-POSIX
-  # pthreads APIs, which by default are not included in the pthreads header
-  # unless we define this macro.
-  add_definitions(-D_DARWIN_C_SOURCE)
-endif(CLR_CMAKE_HOST_APPLE)
-
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-include_directories(${CMAKE_CURRENT_SOURCE_DIR}/env)
-
-install_clr(TARGETS clrgc DESTINATIONS . sharedFramework COMPONENT runtime)
-if (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
-  install_clr(TARGETS clrgcexp DESTINATIONS . sharedFramework COMPONENT runtime)
-endif (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
+if(FEATURE_STANDALONE_GC)
+  # clrgcexp is build with standalone+regions
+  if (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
+    add_library_clr(clrgcexp SHARED ${GC_SOURCES})
+    add_dependencies(clrgcexp eventing_headers)
+    target_link_libraries(clrgcexp PRIVATE ${GC_LINK_LIBRARIES})
+    target_compile_definitions(clrgcexp PRIVATE -DUSE_REGIONS)
+    install_clr(TARGETS clrgcexp DESTINATIONS . COMPONENT runtime)
+  endif (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
+
+  # clrgc is build with standalone+segments
+  add_library_clr(clrgc SHARED ${GC_SOURCES})
+  add_dependencies(clrgc eventing_headers)
+  target_link_libraries(clrgc PRIVATE ${GC_LINK_LIBRARIES})
+  install_clr(TARGETS clrgc DESTINATIONS . COMPONENT runtime)
+
+  add_definitions(-DBUILD_AS_STANDALONE)
+  add_definitions(-DFEATURE_CONSERVATIVE_GC)
+
+  add_definitions(-DFX_VER_INTERNALNAME_STR=clrgc.dll)
+  add_definitions(-DVERIFY_HEAP)
+  if(CLR_CMAKE_HOST_APPLE)
+    # The implementation of GCToOSInterface on Apple platforms makes use of non-POSIX
+    # pthreads APIs, which by default are not included in the pthreads header
+    # unless we define this macro.
+    add_definitions(-D_DARWIN_C_SOURCE)
+  endif(CLR_CMAKE_HOST_APPLE)
+
+  include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+  include_directories(${CMAKE_CURRENT_SOURCE_DIR}/env)
+
+  install_clr(TARGETS clrgc DESTINATIONS . sharedFramework COMPONENT runtime)
+  if (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
+    install_clr(TARGETS clrgcexp DESTINATIONS . sharedFramework COMPONENT runtime)
+  endif (CLR_CMAKE_TARGET_ARCH_ARM64 OR CLR_CMAKE_TARGET_ARCH_AMD64)
+endif(FEATURE_STANDALONE_GC)

src/coreclr/jit/block.h

@@ -1197,6 +1197,18 @@ struct BasicBlock : private LIR::Range
     // getBBWeight -- get the normalized weight of this block
     weight_t getBBWeight(Compiler* comp) const;
 
+    // computeIncomingWeight -- sum the weights of the flow edges into this block
+    weight_t computeIncomingWeight() const
+    {
+        weight_t incomingWeight = BB_ZERO_WEIGHT;
+        for (FlowEdge* const predEdge : PredEdges())
+        {
+            incomingWeight += predEdge->getLikelyWeight();
+        }
+
+        return incomingWeight;
+    }
+
     // hasProfileWeight -- Returns true if this block's weight came from profile data
     bool hasProfileWeight() const
     {

src/coreclr/jit/compiler.cpp

@@ -5127,6 +5127,9 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
     // Expand thread local access
     DoPhase(this, PHASE_EXPAND_TLS, &Compiler::fgExpandThreadLocalAccess);
 
+    // Expand stack allocated arrays
+    DoPhase(this, PHASE_EXPAND_STACK_ARR, &Compiler::fgExpandStackArrayAllocations);
+
     // Insert GC Polls
     DoPhase(this, PHASE_INSERT_GC_POLLS, &Compiler::fgInsertGCPolls);
 
@@ -5153,6 +5156,11 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
         DoPhase(this, PHASE_SWITCH_RECOGNITION, &Compiler::optSwitchRecognition);
     }
 
+    // Drop back to just checking profile likelihoods.
+    //
+    activePhaseChecks &= ~PhaseChecks::CHECK_PROFILE;
+    activePhaseChecks |= PhaseChecks::CHECK_LIKELIHOODS;
+
 #ifdef DEBUG
     // Stash the current estimate of the function's size if necessary.
     if (verbose && opts.OptimizationEnabled())

src/coreclr/jit/compiler.h

@@ -691,8 +691,8 @@ class LclVarDsc
     unsigned char lvSingleDefDisqualifyReason = 'H';
 #endif
 
-    unsigned char lvAllDefsAreNoGc    : 1; // For pinned locals: true if all defs of this local are no-gc
-    unsigned char lvStackAllocatedBox : 1; // Local is a stack allocated box
+    unsigned char lvAllDefsAreNoGc       : 1; // For pinned locals: true if all defs of this local are no-gc
+    unsigned char lvStackAllocatedObject : 1; // Local is a stack allocated object (class, box, array, ...)
 
 #if FEATURE_MULTIREG_ARGS
     regNumber lvRegNumForSlot(unsigned slotNum)
@@ -807,9 +807,9 @@ class LclVarDsc
         return lvIsMultiRegArg || lvIsMultiRegRet;
     }
 
-    bool IsStackAllocatedBox() const
+    bool IsStackAllocatedObject() const
     {
-        return lvStackAllocatedBox;
+        return lvStackAllocatedObject;
     }
 
 #if defined(DEBUG)
@@ -6106,6 +6106,9 @@ class Compiler
     PhaseStatus fgExpandStaticInit();
     bool fgExpandStaticInitForCall(BasicBlock** pBlock, Statement* stmt, GenTreeCall* call);
 
+    PhaseStatus fgExpandStackArrayAllocations();
+    bool fgExpandStackArrayAllocation(BasicBlock* pBlock, Statement* stmt, GenTreeCall* call);
+
     PhaseStatus fgVNBasedIntrinsicExpansion();
     bool fgVNBasedIntrinsicExpansionForCall(BasicBlock** pBlock, Statement* stmt, GenTreeCall* call);
     bool fgVNBasedIntrinsicExpansionForCall_ReadUtf8(BasicBlock** pBlock, Statement* stmt, GenTreeCall* call);
@@ -7537,6 +7540,7 @@ class Compiler
 #define OMF_HAS_SPECIAL_INTRINSICS             0x00020000 // Method contains special intrinsics expanded in late phases
 #define OMF_HAS_RECURSIVE_TAILCALL             0x00040000 // Method contains recursive tail call
 #define OMF_HAS_EXPANDABLE_CAST                0x00080000 // Method contains casts eligible for late expansion
+#define OMF_HAS_STACK_ARRAY                    0x00100000 // Method contains stack allocated arrays
 
     // clang-format on
 
@@ -7627,6 +7631,16 @@ class Compiler
         optMethodFlags |= OMF_HAS_RECURSIVE_TAILCALL;
     }
 
+    bool doesMethodHaveStackAllocatedArray()
+    {
+        return (optMethodFlags & OMF_HAS_STACK_ARRAY) != 0;
+    }
+
+    void setMethodHasStackAllocatedArray()
+    {
+        optMethodFlags |= OMF_HAS_STACK_ARRAY;
+    }
+
     void pickGDV(GenTreeCall*           call,
                  IL_OFFSET              ilOffset,
                  bool                   isInterface,

src/coreclr/jit/compphases.h

@@ -109,6 +109,7 @@ CompPhaseNameMacro(PHASE_EXPAND_RTLOOKUPS,           "Expand runtime lookups",
 CompPhaseNameMacro(PHASE_EXPAND_STATIC_INIT,         "Expand static init",             false, -1, true)
 CompPhaseNameMacro(PHASE_EXPAND_CASTS,               "Expand casts",                   false, -1, true)
 CompPhaseNameMacro(PHASE_EXPAND_TLS,                 "Expand TLS access",              false, -1, true)
+CompPhaseNameMacro(PHASE_EXPAND_STACK_ARR,           "Expand stack array allocation",  false, -1, true)
 CompPhaseNameMacro(PHASE_INSERT_GC_POLLS,            "Insert GC Polls",                false, -1, true)
 CompPhaseNameMacro(PHASE_CREATE_THROW_HELPERS,       "Create throw helper blocks",     false, -1, true)
 CompPhaseNameMacro(PHASE_DETERMINE_FIRST_COLD_BLOCK, "Determine first cold block",     false, -1, true)

src/coreclr/jit/fgdiagnostic.cpp

@@ -955,6 +955,10 @@ bool Compiler::fgDumpFlowGraph(Phases phase, PhasePosition pos)
             {
                 fprintf(fgxFile, "\n            hot=\"true\"");
             }
+            if (block->HasFlag(BBF_HAS_NEWARR))
+            {
+                fprintf(fgxFile, "\n            callsNewArr=\"true\"");
+            }
             if (block->HasFlag(BBF_HAS_NEWOBJ))
             {
                 fprintf(fgxFile, "\n            callsNew=\"true\"");

src/coreclr/jit/gentree.cpp

@@ -13192,6 +13192,8 @@ const char* Compiler::gtGetWellKnownArgNameForArgMsg(WellKnownArg arg)
             return "swift self";
         case WellKnownArg::X86TailCallSpecialArg:
             return "tail call";
+        case WellKnownArg::StackArrayLocal:
+            return "&lcl arr";
         default:
             return nullptr;
     }
@@ -19814,7 +19816,10 @@ void GenTreeArrAddr::ParseArrayAddress(Compiler* comp, GenTree** pArr, ValueNum*
 /* static */ void GenTreeArrAddr::ParseArrayAddressWork(
     GenTree* tree, Compiler* comp, target_ssize_t inputMul, GenTree** pArr, ValueNum* pInxVN, target_ssize_t* pOffset)
 {
-    if (tree->TypeIs(TYP_REF))
+    ValueNum  vn = comp->GetValueNumStore()->VNLiberalNormalValue(tree->gtVNPair);
+    VNFuncApp vnf;
+
+    if (tree->TypeIs(TYP_REF) || comp->GetValueNumStore()->IsVNNewArr(vn, &vnf))
     {
         // This must be the array pointer.
         assert(*pArr == nullptr);
@@ -19917,7 +19922,7 @@ void GenTreeArrAddr::ParseArrayAddress(Compiler* comp, GenTree** pArr, ValueNum*
         // If we didn't return above, must be a contribution to the non-constant part of the index VN.
         // We don't get here for GT_CNS_INT, GT_ADD, or GT_SUB, or for GT_MUL by constant, or GT_LSH of
         // constant shift. Thus, the generated index VN does not include the parsed constant offset.
-        ValueNum vn = comp->GetValueNumStore()->VNLiberalNormalValue(tree->gtVNPair);
+        //
         if (inputMul != 1)
         {
             ValueNum mulVN = comp->GetValueNumStore()->VNForLongCon(inputMul);