Merge pull request #23564 from JuliaLang/kf/gcmemcpy

Don't ignore memcpy in GC root placement

Author: Keno

src/llvm-late-gc-lowering.cpp

@@ -729,8 +729,16 @@ State LateLowerGCFrame::LocalScan(Function &F) {
             Instruction &I = *it;
             if (CallInst *CI = dyn_cast<CallInst>(&I)) {
                 if (isa<IntrinsicInst>(CI)) {
-                    // Intrinsics are never GC uses/defs
-                    continue;
+                    // Most intrinsics are not gc uses/defs, however some have
+                    // memory operands and could thus be GC uses. To be conservative,
+                    // we only skip processing for those that we know we emit often
+                    // and cannot possibly be GC uses.
+                    IntrinsicInst *II = cast<IntrinsicInst>(CI);
+                    if (isa<DbgInfoIntrinsic>(CI) ||
+                        II->getIntrinsicID() == Intrinsic::lifetime_start ||
+                        II->getIntrinsicID() == Intrinsic::lifetime_end) {
+                        continue;
+                    }
                 }
                 MaybeNoteDef(S, BBS, CI, BBS.Safepoints);
                 NoteOperandUses(S, BBS, I, BBS.UpExposedUses);
@@ -744,6 +752,10 @@ State LateLowerGCFrame::LocalScan(Function &F) {
                 if (CI->canReturnTwice()) {
                     S.ReturnsTwice.push_back(CI);
                 }
+                if (isa<IntrinsicInst>(CI)) {
+                    // Intrinsics are never safepoints.
+                    continue;
+                }
                 if (auto callee = CI->getCalledFunction()) {
                     // Known functions emitted in codegen that are not safepoints
                     if (callee == pointer_from_objref_func || callee->getName() == "memcmp") {

test/llvmpasses/gcroots.ll

@@ -192,7 +192,7 @@ define %jl_value_t addrspace(10)* @no_redundant_rerooting(i64 %a, i1 %cond) {
 top:
     %ptls = call %jl_value_t*** @jl_get_ptls_states()
     %aboxed = call %jl_value_t addrspace(10)* @jl_box_int64(i64 signext %a)
-; CHECK: store %jl_value_t addrspace(10)* %aboxed 
+; CHECK: store %jl_value_t addrspace(10)* %aboxed
 ; CHECK-NEXT: call void @jl_safepoint()
     call void @jl_safepoint()
     br i1 %cond, label %blocka, label %blockb
@@ -207,3 +207,19 @@ blockb:
     call void @jl_safepoint()
     ret %jl_value_t addrspace(10)* %aboxed
 }
+
+declare void @llvm.memcpy.p064.p10i8.i64(i64*, i8 addrspace(10)*, i64, i32, i1)
+
+define void @memcpy_use(i64 %a, i64 *%aptr) {
+; CHECK-LABEL: @memcpy_use
+; CHECK: %gcframe = alloca %jl_value_t addrspace(10)*, i32 3
+top:
+    %ptls = call %jl_value_t*** @jl_get_ptls_states()
+    %aboxed = call %jl_value_t addrspace(10)* @jl_box_int64(i64 signext %a)
+; CHECK: store %jl_value_t addrspace(10)* %aboxed
+    call void @jl_safepoint()
+    %acast = bitcast %jl_value_t addrspace(10)* %aboxed to i8 addrspace(10)*
+    call void @llvm.memcpy.p064.p10i8.i64(i64* %aptr, i8 addrspace(10)* %acast, i64 8, i32 1, i1 false)
+    ret void
+}
+