Initialize hoisted object allocations (#49584)

JuliaLang · May 4, 2023 · 5032a1a · 5032a1a
1 parent 93ce36c
commit 5032a1a
Show file tree

Hide file tree

Showing 2 changed files with 23 additions and 3 deletions.
diff --git a/src/llvm-julia-licm.cpp b/src/llvm-julia-licm.cpp
@@ -134,7 +134,6 @@ struct JuliaLICMPassLegacy : public LoopPass {
             getLoopAnalysisUsage(AU);
         }
 };
-
 struct JuliaLICM : public JuliaPassContext {
     function_ref<DominatorTree &()> GetDT;
     function_ref<LoopInfo &()> GetLI;
@@ -339,6 +338,19 @@ struct JuliaLICM : public JuliaPassContext {
                     });
                     ++HoistedAllocation;
                     moveInstructionBefore(*call, *preheader->getTerminator(), MSSAU, SE);
+                    IRBuilder<> builder(preheader->getTerminator());
+                    builder.SetCurrentDebugLocation(call->getDebugLoc());
+                    auto obj_i8 = builder.CreateBitCast(call, Type::getInt8PtrTy(call->getContext(), call->getType()->getPointerAddressSpace()));
+                    // Note that this alignment is assuming the GC allocates at least pointer-aligned memory
+                    auto align = Align(DL.getPointerSize(0));
+                    auto clear_obj = builder.CreateMemSet(obj_i8, ConstantInt::get(Type::getInt8Ty(call->getContext()), 0), call->getArgOperand(1), align);
+                    if (MSSAU.getMemorySSA()) {
+                        auto alloc_mdef = MSSAU.getMemorySSA()->getMemoryAccess(call);
+                        assert(isa<MemoryDef>(alloc_mdef) && "Expected alloc to be associated with a memory def!");
+                        auto clear_mdef = MSSAU.createMemoryAccessAfter(clear_obj, nullptr, alloc_mdef);
+                        assert(isa<MemoryDef>(clear_mdef) && "Expected memset to be associated with a memory def!");
+                        (void) clear_mdef;
+                    }
                     changed = true;
                 }
             }
@@ -395,6 +407,10 @@ PreservedAnalyses JuliaLICMPass::run(Loop &L, LoopAnalysisManager &AM,
     };
     auto juliaLICM = JuliaLICM(GetDT, GetLI, GetMSSA, GetSE);
     if (juliaLICM.runOnLoop(&L, ORE)) {
+#ifdef JL_DEBUG_BUILD
+        if (AR.MSSA)
+            AR.MSSA->verifyMemorySSA();
+#endif
         auto preserved = getLoopPassPreservedAnalyses();
         preserved.preserveSet<CFGAnalyses>();
         preserved.preserve<MemorySSAAnalysis>();

diff --git a/test/llvmpasses/julia-licm.ll b/test/llvmpasses/julia-licm.ll
@@ -29,13 +29,15 @@ L4:                                               ; preds = %top
   %current_task112 = getelementptr inbounds {}**, {}*** %1, i64 -12
   %current_task1 = bitcast {}*** %current_task112 to {}**
   ; CHECK: %3 = call noalias nonnull {} addrspace(10)* @julia.gc_alloc_obj({}** nonnull %current_task1, i64 8, {} addrspace(10)* @tag)
+  ; CHECK-NEXT: %4 = bitcast {} addrspace(10)* %3 to i8 addrspace(10)*
+  ; CHECK-NEXT: call void @llvm.memset.p10i8.i64(i8 addrspace(10)* align {{[0-9]+}} %4, i8 0, i64 8, i1 false)
   ; CHECK-NEXT: br label %L22
   br label %L22
 
 L22:                                              ; preds = %L4, %L22
   %value_phi5 = phi i64 [ 1, %L4 ], [ %5, %L22 ]
-  ; CHECK: %value_phi5 = phi i64 [ 1, %L4 ], [ %5, %L22 ]
-  ; CHECK-NEXT %4 = bitcast {} addrspace(10)* %3 to i64 addrspace(10)*
+  ; CHECK: %value_phi5 = phi i64 [ 1, %L4 ], [ %6, %L22 ]
+  ; CHECK-NEXT %5 = bitcast {} addrspace(10)* %3 to i64 addrspace(10)*
   %3 = call noalias nonnull {} addrspace(10)* @julia.gc_alloc_obj({}** nonnull %current_task1, i64 8, {} addrspace(10)* @tag) #1
   %4 = bitcast {} addrspace(10)* %3 to i64 addrspace(10)*
   store i64 %value_phi5, i64 addrspace(10)* %4, align 8, !tbaa !2
@@ -55,6 +57,8 @@ top:
 ; CHECK: preheader:
 preheader:
 ; CHECK-NEXT: %alloc = call noalias nonnull {} addrspace(10)* @julia.gc_alloc_obj({}** nonnull %current_task, i64 8, {} addrspace(10)* @tag)
+; CHECK-NEXT: [[casted:%.*]] = bitcast {} addrspace(10)* %alloc to i8 addrspace(10)*
+; CHECK-NEXT: call void @llvm.memset.p10i8.i64(i8 addrspace(10)* align {{[0-9]+}} [[casted]], i8 0, i64 8, i1 false)
 ; CHECK-NEXT: br label %loop
   br label %loop
 loop: