AllocOpt: Fix stack lowering where object contains both julia and non-julia objects

wsmoses · wsmoses · commit 81e683a8d0bd · 2024-07-29T19:02:30.000-04:00
diff --git a/src/llvm-alloc-helpers.cpp b/src/llvm-alloc-helpers.cpp
@@ -88,6 +88,9 @@ bool AllocUseInfo::addMemOp(Instruction *inst, unsigned opno, uint32_t offset,
     memop.isaggr = isa<StructType>(elty) || isa<ArrayType>(elty) || isa<VectorType>(elty);
     memop.isobjref = hasObjref(elty);
     auto &field = getField(offset, size, elty);
+            
+    field.second.hasbits |= !hasObjref(elty) || (hasObjref(elty) && !isa<PointerType>(elty));
+
     if (field.second.hasobjref != memop.isobjref)
         field.second.multiloc = true; // can't split this field, since it contains a mix of references and bits
     if (!isstore)
@@ -198,6 +201,7 @@ void jl_alloc::runEscapeAnalysis(llvm::CallInst *I, EscapeAnalysisRequiredArgs r
                 auto elty = inst->getType();
                 required.use_info.has_unknown_objref |= hasObjref(elty);
                 required.use_info.has_unknown_objrefaggr |= hasObjref(elty) && !isa<PointerType>(elty);
+                required.use_info.has_unknown_bits |= !hasObjref(elty) || (hasObjref(elty) && !isa<PointerType>(elty));
                 required.use_info.hasunknownmem = true;
             } else if (!required.use_info.addMemOp(inst, 0, cur.offset,
                                                                inst->getType(),
@@ -289,6 +293,7 @@ void jl_alloc::runEscapeAnalysis(llvm::CallInst *I, EscapeAnalysisRequiredArgs r
                 auto elty = storev->getType();
                 required.use_info.has_unknown_objref |= hasObjref(elty);
                 required.use_info.has_unknown_objrefaggr |= hasObjref(elty) && !isa<PointerType>(elty);
+                required.use_info.has_unknown_bits |= !hasObjref(elty) || (hasObjref(elty) && !isa<PointerType>(elty));
                 required.use_info.hasunknownmem = true;
             } else if (!required.use_info.addMemOp(inst, use->getOperandNo(),
                                                                cur.offset, storev->getType(),
@@ -310,10 +315,14 @@ void jl_alloc::runEscapeAnalysis(llvm::CallInst *I, EscapeAnalysisRequiredArgs r
             }
             required.use_info.hasload = true;
             auto storev = isa<AtomicCmpXchgInst>(inst) ? cast<AtomicCmpXchgInst>(inst)->getNewValOperand() : cast<AtomicRMWInst>(inst)->getValOperand();
+            Type *elty = storev->getType();
             if (cur.offset == UINT32_MAX || !required.use_info.addMemOp(inst, use->getOperandNo(),
-                                                               cur.offset, storev->getType(),
+                                                               cur.offset, elty,
                                                                true, required.DL)) {
                 LLVM_DEBUG(dbgs() << "Atomic inst has unknown offset\n");
+                required.use_info.has_unknown_objref |= hasObjref(elty);
+                required.use_info.has_unknown_objrefaggr |= hasObjref(elty) && !isa<PointerType>(elty);
+                required.use_info.has_unknown_bits |= !hasObjref(elty) || (hasObjref(elty) && !isa<PointerType>(elty));
                 required.use_info.hasunknownmem = true;
             }
             required.use_info.refload = true;
diff --git a/src/llvm-alloc-helpers.h b/src/llvm-alloc-helpers.h
@@ -46,6 +46,8 @@ namespace jl_alloc {
         bool hasaggr:1;
         bool multiloc:1;
         bool hasload:1;
+        // The alloc has a non-julia object at this offset.
+        bool hasbits:1;
         llvm::Type *elty;
         llvm::SmallVector<MemOp,4> accesses;
         Field(uint32_t size, llvm::Type *elty)
@@ -54,6 +56,7 @@ namespace jl_alloc {
               hasaggr(false),
               multiloc(false),
               hasload(false),
+              hasbits(false),
               elty(elty)
         {
         }
@@ -95,6 +98,9 @@ namespace jl_alloc {
         // The alloc has an aggregate Julia object reference not in an explicit field.
         bool has_unknown_objrefaggr:1;
 
+        // The alloc has a non-julia object at an unknown offset.
+        bool has_unknown_bits:1;
+
         void reset()
         {
             escaped = false;
@@ -110,6 +116,7 @@ namespace jl_alloc {
             allockind = llvm::AllocFnKind::Unknown;
             has_unknown_objref = false;
             has_unknown_objrefaggr = false;
+            has_unknown_bits = false;
             uses.clear();
             preserves.clear();
             memops.clear();
diff --git a/src/llvm-alloc-opt.cpp b/src/llvm-alloc-opt.cpp
@@ -252,10 +252,12 @@ void Optimizer::optimizeAll()
             removeAlloc(orig);
             continue;
         }
+        bool has_bits = use_info.has_unknown_bits;
         bool has_ref = use_info.has_unknown_objref;
         bool has_refaggr = use_info.has_unknown_objrefaggr;
         for (auto memop: use_info.memops) {
             auto &field = memop.second;
+            has_bits |= field.hasbits;
             if (field.hasobjref) {
                 has_ref = true;
                 // This can be relaxed a little based on hasload
@@ -284,6 +286,15 @@ void Optimizer::optimizeAll()
             splitOnStack(orig);
             continue;
         }
+        if (has_bits && has_ref) {
+            REMARK([&]() {
+                return OptimizationRemarkMissed(DEBUG_TYPE, "Escaped", orig)
+                    << "GC allocation could not be split and contained both julia object and non-julia-object data, unable to move to stack " << ore::NV("GC Allocation", orig);
+            });
+            if (use_info.hastypeof)
+                optimizeTag(orig);
+            continue;
+        }
         REMARK([&](){
             return OptimizationRemark(DEBUG_TYPE, "Stack Move Allocation", orig)
                 << "GC allocation moved to stack " << ore::NV("GC Allocation", orig);
diff --git a/test/llvmpasses/alloc-opt-bits.ll b/test/llvmpasses/alloc-opt-bits.ll
@@ -0,0 +1,36 @@
+; This file is a part of Julia. License is MIT: https://julialang.org/license
+
+; RUN: opt --load-pass-plugin=libjulia-codegen%shlibext -passes='function(AllocOpt)' -S %s | FileCheck %s --check-prefixes=CHECK,OPAQUE
+
+
+@tag = external addrspace(10) global {}
+
+@glob = external addrspace(10) global {}
+
+; Test that the gc_preserve intrinsics are deleted directly.
+
+; CHECK-LABEL: @ptr_and_bits
+; CHECK: call noalias ptr addrspace(10) @julia.gc_alloc_obj
+
+define void @ptr_and_bits(ptr %fptr, i1 %b, i1 %b2, i32 %idx) {
+  %pgcstack = call ptr @julia.get_pgcstack()
+  %ptls = call ptr @julia.ptls_states()
+  %ptls_i8 = bitcast ptr %ptls to ptr
+  %v = call noalias ptr addrspace(10) @julia.gc_alloc_obj(ptr %ptls_i8, i64 16, ptr addrspace(10) @tag)
+  
+  %g0 = getelementptr { i64, ptr addrspace(10) }, ptr addrspace(10) %v, i32 %idx, i32 1
+  store ptr addrspace(10) @glob, ptr addrspace(10) %g0
+  
+  %g1 = getelementptr { i64, ptr addrspace(10) }, ptr addrspace(10) %v, i32 %idx, i32 0
+  store i64 7, ptr addrspace(10) %g1
+
+  %res = load ptr addrspace(10), ptr addrspace(10) %g0
+  %res2 = load i64, ptr addrspace(10) %g1
+  ret void
+}
+
+declare noalias ptr addrspace(10) @julia.gc_alloc_obj(ptr, i64, ptr addrspace(10))
+
+declare ptr @julia.ptls_states()
+
+declare ptr @julia.get_pgcstack()
