Restructuring lowerGCAllocBytesLate pass

JuliaLang · Dec 2, 2024 · 52c76b1 · 52c76b1
1 parent 1cc0814
commit 52c76b1
Show file tree

Hide file tree

Showing 5 changed files with 117 additions and 117 deletions.
diff --git a/src/Makefile b/src/Makefile
@@ -65,7 +65,14 @@ RT_LLVMLINK :=
 CG_LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
+# Currently these files are used by both GCs. But we should make the list specific to stock, and MMTk should have its own implementation.
 GC_CODEGEN_SRCS := llvm-final-gc-lowering llvm-late-gc-lowering llvm-gc-invariant-verifier
+ifeq ($(WITH_MMTK), 1)
+FLAGS += -I$(MMTK_API_INC)
+GC_CODEGEN_SRCS += llvm-late-gc-lowering-mmtk
+else
+GC_CODEGEN_SRCS += llvm-late-gc-lowering-stock
+endif
 CODEGEN_SRCS := codegen jitlayers aotcompile debuginfo disasm llvm-simdloop \
 	llvm-pass-helpers llvm-ptls \
 	llvm-lower-handlers llvm-propagate-addrspaces \

diff --git a/src/llvm-gc-interface-passes.h b/src/llvm-gc-interface-passes.h
@@ -368,9 +368,7 @@ struct LateLowerGCFrame:  private JuliaPassContext {
     void RefineLiveSet(LargeSparseBitVector &LS, State &S, ArrayRef<int> CalleeRoots);
     Value *EmitTagPtr(IRBuilder<> &builder, Type *T, Type *T_size, Value *V);
     Value *EmitLoadTag(IRBuilder<> &builder, Type *T_size, Value *V);
-#ifdef MMTK_GC
     Value* lowerGCAllocBytesLate(CallInst *target, Function &F);
-#endif
 };
 
 // The final GC lowering pass. This pass lowers platform-agnostic GC

diff --git a/src/llvm-late-gc-lowering-mmtk.cpp b/src/llvm-late-gc-lowering-mmtk.cpp
@@ -0,0 +1,99 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "llvm-gc-interface-passes.h"
+
+Value* LateLowerGCFrame::lowerGCAllocBytesLate(CallInst *target, Function &F)
+{
+    assert(target->arg_size() == 3);
+
+    IRBuilder<> builder(target);
+    auto ptls = target->getArgOperand(0);
+    auto type = target->getArgOperand(2);
+    if (auto CI = dyn_cast<ConstantInt>(target->getArgOperand(1))) {
+        size_t sz = (size_t)CI->getZExtValue();
+        // This is strongly architecture and OS dependent
+        int osize;
+        int offset = jl_gc_classify_pools(sz, &osize);
+        if (offset >= 0) {
+            // In this case instead of lowering julia.gc_alloc_bytes to jl_gc_small_alloc
+            // We do a slowpath/fastpath check and lower it only on the slowpath, returning
+            // the cursor and updating it in the fastpath.
+            auto pool_osize_i32 = ConstantInt::get(Type::getInt32Ty(F.getContext()), osize);
+            auto pool_osize = ConstantInt::get(Type::getInt64Ty(F.getContext()), osize);
+
+            // Should we generate fastpath allocation sequence here? We should always generate fastpath here for MMTk.
+            // Setting this to false will increase allocation overhead a lot, and should only be used for debugging.
+            const bool INLINE_FASTPATH_ALLOCATION = true;
+
+            if (INLINE_FASTPATH_ALLOCATION) {
+                // Assuming we use the first immix allocator.
+                // FIXME: We should get the allocator index and type from MMTk.
+                auto allocator_offset = offsetof(jl_tls_states_t, gc_tls) + offsetof(jl_gc_tls_states_t, mmtk_mutator) + offsetof(MMTkMutatorContext, allocators) + offsetof(Allocators, immix);
+
+                auto cursor_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, cursor));
+                auto limit_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()),  allocator_offset + offsetof(ImmixAllocator, limit));
+
+                auto cursor_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, cursor_pos);
+                auto cursor_ptr = builder.CreateBitCast(cursor_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "cursor_ptr");
+                auto cursor = builder.CreateLoad(Type::getInt64Ty(target->getContext()), cursor_ptr, "cursor");
+
+                // offset = 8
+                auto delta_offset = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), ConstantInt::get(Type::getInt64Ty(target->getContext()), 8));
+                auto delta_cursor = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), cursor);
+                auto delta_op = builder.CreateNSWAdd(delta_offset, delta_cursor);
+                // alignment 16 (15 = 16 - 1)
+                auto delta = builder.CreateAnd(delta_op, ConstantInt::get(Type::getInt64Ty(target->getContext()), 15), "delta");
+                auto result = builder.CreateNSWAdd(cursor, delta, "result");
+
+                auto new_cursor = builder.CreateNSWAdd(result, pool_osize);
+
+                auto limit_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, limit_pos);
+                auto limit_ptr = builder.CreateBitCast(limit_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "limit_ptr");
+                auto limit = builder.CreateLoad(Type::getInt64Ty(target->getContext()), limit_ptr, "limit");
+
+                auto gt_limit = builder.CreateICmpSGT(new_cursor, limit);
+
+                auto slowpath = BasicBlock::Create(target->getContext(), "slowpath", target->getFunction());
+                auto fastpath = BasicBlock::Create(target->getContext(), "fastpath", target->getFunction());
+
+                auto next_instr = target->getNextNode();
+                SmallVector<uint32_t, 2> Weights{1, 9};
+
+                MDBuilder MDB(F.getContext());
+                SplitBlockAndInsertIfThenElse(gt_limit, next_instr, &slowpath, &fastpath, false, false, MDB.createBranchWeights(Weights));
+
+                builder.SetInsertPoint(next_instr);
+                auto phiNode = builder.CreatePHI(target->getCalledFunction()->getReturnType(), 2, "phi_fast_slow");
+
+                // slowpath
+                builder.SetInsertPoint(slowpath);
+                auto pool_offs = ConstantInt::get(Type::getInt32Ty(F.getContext()), 1);
+                auto new_call = builder.CreateCall(smallAllocFunc, { ptls, pool_offs, pool_osize_i32, type });
+                new_call->setAttributes(new_call->getCalledFunction()->getAttributes());
+                builder.CreateBr(next_instr->getParent());
+
+                // fastpath
+                builder.SetInsertPoint(fastpath);
+                builder.CreateStore(new_cursor, cursor_ptr);
+
+                // ptls->gc_tls.gc_num.allocd += osize;
+                auto pool_alloc_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), offsetof(jl_tls_states_t, gc_tls_common) + offsetof(jl_gc_tls_states_common_t, gc_num));
+                auto pool_alloc_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, pool_alloc_pos);
+                auto pool_alloc_tls = builder.CreateBitCast(pool_alloc_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "pool_alloc");
+                auto pool_allocd = builder.CreateLoad(Type::getInt64Ty(target->getContext()), pool_alloc_tls);
+                auto pool_allocd_total = builder.CreateAdd(pool_allocd, pool_osize);
+                builder.CreateStore(pool_allocd_total, pool_alloc_tls);
+
+                auto v_raw = builder.CreateNSWAdd(result, ConstantInt::get(Type::getInt64Ty(target->getContext()), sizeof(jl_taggedvalue_t)));
+                auto v_as_ptr = builder.CreateIntToPtr(v_raw, smallAllocFunc->getReturnType());
+                builder.CreateBr(next_instr->getParent());
+
+                phiNode->addIncoming(new_call, slowpath);
+                phiNode->addIncoming(v_as_ptr, fastpath);
+                phiNode->takeName(target);
+                return phiNode;
+            }
+        }
+    }
+    return target;
+}
diff --git a/src/llvm-late-gc-lowering-stock.cpp b/src/llvm-late-gc-lowering-stock.cpp
@@ -0,0 +1,9 @@
+// This file is a part of Julia. License is MIT: https://julialang.org/license
+
+#include "llvm-gc-interface-passes.h"
+
+Value* LateLowerGCFrame::lowerGCAllocBytesLate(CallInst *target, Function &F)
+{
+    // Do nothing for the stock GC
+    return target;
+}
diff --git a/src/llvm-late-gc-lowering.cpp b/src/llvm-late-gc-lowering.cpp
@@ -2465,119 +2465,6 @@ void LateLowerGCFrame::PlaceRootsAndUpdateCalls(ArrayRef<int> Colors, int PreAss
     }
 }
 
-#ifdef MMTK_GC
-Value* LateLowerGCFrame::lowerGCAllocBytesLate(CallInst *target, Function &F)
-{
-    assert(target->arg_size() == 3);
-
-    IRBuilder<> builder(target);
-    auto ptls = target->getArgOperand(0);
-    auto type = target->getArgOperand(2);
-    if (auto CI = dyn_cast<ConstantInt>(target->getArgOperand(1))) {
-        size_t sz = (size_t)CI->getZExtValue();
-        // This is strongly architecture and OS dependent
-        int osize;
-        int offset = jl_gc_classify_pools(sz, &osize);
-        if (offset >= 0) {
-            // In this case instead of lowering julia.gc_alloc_bytes to jl_gc_small_alloc
-            // We do a slowpath/fastpath check and lower it only on the slowpath, returning
-            // the cursor and updating it in the fastpath.
-            auto pool_osize_i32 = ConstantInt::get(Type::getInt32Ty(F.getContext()), osize);
-            auto pool_osize = ConstantInt::get(Type::getInt64Ty(F.getContext()), osize);
-
-            // Should we generate fastpath allocation sequence here? We should always generate fastpath here for MMTk.
-            // Setting this to false will increase allocation overhead a lot, and should only be used for debugging.
-            const bool INLINE_FASTPATH_ALLOCATION = true;
-
-            if (INLINE_FASTPATH_ALLOCATION) {
-                // Assuming we use the first immix allocator.
-                // FIXME: We should get the allocator index and type from MMTk.
-                auto allocator_offset = offsetof(jl_tls_states_t, gc_tls) + offsetof(jl_gc_tls_states_t, mmtk_mutator) + offsetof(MMTkMutatorContext, allocators) + offsetof(Allocators, immix);
-
-                auto cursor_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), allocator_offset + offsetof(ImmixAllocator, cursor));
-                auto limit_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()),  allocator_offset + offsetof(ImmixAllocator, limit));
-
-                auto cursor_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, cursor_pos);
-                auto cursor_ptr = builder.CreateBitCast(cursor_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "cursor_ptr");
-                auto cursor = builder.CreateLoad(Type::getInt64Ty(target->getContext()), cursor_ptr, "cursor");
-
-                // offset = 8
-                auto delta_offset = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), ConstantInt::get(Type::getInt64Ty(target->getContext()), 8));
-                auto delta_cursor = builder.CreateNSWSub(ConstantInt::get(Type::getInt64Ty(target->getContext()), 0), cursor);
-                auto delta_op = builder.CreateNSWAdd(delta_offset, delta_cursor);
-                // alignment 16 (15 = 16 - 1)
-                auto delta = builder.CreateAnd(delta_op, ConstantInt::get(Type::getInt64Ty(target->getContext()), 15), "delta");
-                auto result = builder.CreateNSWAdd(cursor, delta, "result");
-
-                auto new_cursor = builder.CreateNSWAdd(result, pool_osize);
-
-                auto limit_tls_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, limit_pos);
-                auto limit_ptr = builder.CreateBitCast(limit_tls_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "limit_ptr");
-                auto limit = builder.CreateLoad(Type::getInt64Ty(target->getContext()), limit_ptr, "limit");
-
-                auto gt_limit = builder.CreateICmpSGT(new_cursor, limit);
-
-                auto slowpath = BasicBlock::Create(target->getContext(), "slowpath", target->getFunction());
-                auto fastpath = BasicBlock::Create(target->getContext(), "fastpath", target->getFunction());
-
-                auto next_instr = target->getNextNode();
-                SmallVector<uint32_t, 2> Weights{1, 9};
-
-                MDBuilder MDB(F.getContext());
-                SplitBlockAndInsertIfThenElse(gt_limit, next_instr, &slowpath, &fastpath, false, false, MDB.createBranchWeights(Weights));
-
-                builder.SetInsertPoint(next_instr);
-                auto phiNode = builder.CreatePHI(target->getCalledFunction()->getReturnType(), 2, "phi_fast_slow");
-
-                // slowpath
-                builder.SetInsertPoint(slowpath);
-                auto pool_offs = ConstantInt::get(Type::getInt32Ty(F.getContext()), 1);
-                auto new_call = builder.CreateCall(smallAllocFunc, { ptls, pool_offs, pool_osize_i32, type });
-                new_call->setAttributes(new_call->getCalledFunction()->getAttributes());
-                builder.CreateBr(next_instr->getParent());
-
-                // fastpath
-                builder.SetInsertPoint(fastpath);
-                builder.CreateStore(new_cursor, cursor_ptr);
-
-                // ptls->gc_tls.gc_num.allocd += osize;
-                auto pool_alloc_pos = ConstantInt::get(Type::getInt64Ty(target->getContext()), offsetof(jl_tls_states_t, gc_tls_common) + offsetof(jl_gc_tls_states_common_t, gc_num));
-                auto pool_alloc_i8 = builder.CreateGEP(Type::getInt8Ty(target->getContext()), ptls, pool_alloc_pos);
-                auto pool_alloc_tls = builder.CreateBitCast(pool_alloc_i8, PointerType::get(Type::getInt64Ty(target->getContext()), 0), "pool_alloc");
-                auto pool_allocd = builder.CreateLoad(Type::getInt64Ty(target->getContext()), pool_alloc_tls);
-                auto pool_allocd_total = builder.CreateAdd(pool_allocd, pool_osize);
-                builder.CreateStore(pool_allocd_total, pool_alloc_tls);
-
-                auto v_raw = builder.CreateNSWAdd(result, ConstantInt::get(Type::getInt64Ty(target->getContext()), sizeof(jl_taggedvalue_t)));
-                auto v_as_ptr = builder.CreateIntToPtr(v_raw, smallAllocFunc->getReturnType());
-                builder.CreateBr(next_instr->getParent());
-
-                phiNode->addIncoming(new_call, slowpath);
-                phiNode->addIncoming(v_as_ptr, fastpath);
-                phiNode->takeName(target);
-                return phiNode;
-            }
-        }
-    }
-    return target;
-}
-
-template<typename TIterator>
-static void replaceInstruction(
-    Instruction *oldInstruction,
-    Value *newInstruction,
-    TIterator &it)
-{
-    if (newInstruction != oldInstruction) {
-        oldInstruction->replaceAllUsesWith(newInstruction);
-        it = oldInstruction->eraseFromParent();
-    }
-    else {
-        ++it;
-    }
-}
-#endif
-
 bool LateLowerGCFrame::runOnFunction(Function &F, bool *CFGModified) {
     initAll(*F.getParent());
     smallAllocFunc = getOrDeclare(jl_well_known::GCSmallAlloc);
@@ -2596,8 +2483,9 @@ bool LateLowerGCFrame::runOnFunction(Function &F, bool *CFGModified) {
     PlaceRootsAndUpdateCalls(Colors.first, Colors.second, S, CallFrames);
     CleanupIR(F, &S, CFGModified);
 
-#ifdef MMTK_GC
+
     // We lower the julia.gc_alloc_bytes intrinsic in this pass to insert slowpath/fastpath blocks for MMTk
+    // For now, we do nothing for the Stock GC
     auto GCAllocBytes = getOrNull(jl_intrinsics::GCAllocBytes);
 
     if (GCAllocBytes) {
@@ -2618,7 +2506,6 @@ bool LateLowerGCFrame::runOnFunction(Function &F, bool *CFGModified) {
             ++it;
         }
     }
-#endif
 
     return true;
 }