Auto merge of #86873 - nikic:opaque-ptrs, r=nagisa

Improve opaque pointers support

Opaque pointers are coming, and rustc is not ready.

This adds partial support by passing an explicit load type to LLVM. Two issues I've encountered:
 * The necessary type was not available at the point where non-temporal copies were generated. I've pushed the code for that upwards out of the memcpy implementation and moved the position of a cast to make do with the types we have available. (I'm not sure that cast is needed at all, but have retained it in the interest of conservativeness.)
 * The `PlaceRef::project_deref()` function used during debuginfo generation seems to be buggy in some way -- though I haven't figured out specifically what it does wrong. Replacing it with `load_operand().deref()` did the trick, but I don't really know what I'm doing here.

compiler/rustc_codegen_llvm/src/builder.rs

@@ -410,17 +410,17 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn load(&mut self, ptr: &'ll Value, align: Align) -> &'ll Value {
+    fn load(&mut self, ty: &'ll Type, ptr: &'ll Value, align: Align) -> &'ll Value {
         unsafe {
-            let load = llvm::LLVMBuildLoad(self.llbuilder, ptr, UNNAMED);
+            let load = llvm::LLVMBuildLoad2(self.llbuilder, ty, ptr, UNNAMED);
             llvm::LLVMSetAlignment(load, align.bytes() as c_uint);
             load
         }
     }
 
-    fn volatile_load(&mut self, ptr: &'ll Value) -> &'ll Value {
+    fn volatile_load(&mut self, ty: &'ll Type, ptr: &'ll Value) -> &'ll Value {
         unsafe {
-            let load = llvm::LLVMBuildLoad(self.llbuilder, ptr, UNNAMED);
+            let load = llvm::LLVMBuildLoad2(self.llbuilder, ty, ptr, UNNAMED);
             llvm::LLVMSetVolatile(load, llvm::True);
             load
         }
@@ -428,6 +428,7 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
 
     fn atomic_load(
         &mut self,
+        ty: &'ll Type,
         ptr: &'ll Value,
         order: rustc_codegen_ssa::common::AtomicOrdering,
         size: Size,
@@ -435,6 +436,7 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         unsafe {
             let load = llvm::LLVMRustBuildAtomicLoad(
                 self.llbuilder,
+                ty,
                 ptr,
                 UNNAMED,
                 AtomicOrdering::from_generic(order),
@@ -486,7 +488,7 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
                 }
             }
             let llval = const_llval.unwrap_or_else(|| {
-                let load = self.load(place.llval, place.align);
+                let load = self.load(place.layout.llvm_type(self), place.llval, place.align);
                 if let abi::Abi::Scalar(ref scalar) = place.layout.abi {
                     scalar_load_metadata(self, load, scalar);
                 }
@@ -498,7 +500,8 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
 
             let mut load = |i, scalar: &abi::Scalar, align| {
                 let llptr = self.struct_gep(place.llval, i as u64);
-                let load = self.load(llptr, align);
+                let llty = place.layout.scalar_pair_element_llvm_type(self, i, false);
+                let load = self.load(llty, llptr, align);
                 scalar_load_metadata(self, load, scalar);
                 self.to_immediate_scalar(load, scalar)
             };
@@ -815,13 +818,7 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         size: &'ll Value,
         flags: MemFlags,
     ) {
-        if flags.contains(MemFlags::NONTEMPORAL) {
+        assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memcpy not supported");
-            // HACK(nox): This is inefficient but there is no nontemporal memcpy.
-            let val = self.load(src, src_align);
-            let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
-            self.store_with_flags(val, ptr, dst_align, flags);
-            return;
-        }
         let size = self.intcast(size, self.type_isize(), false);
         let is_volatile = flags.contains(MemFlags::VOLATILE);
         let dst = self.pointercast(dst, self.type_i8p());
@@ -848,13 +845,7 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         size: &'ll Value,
         flags: MemFlags,
     ) {
-        if flags.contains(MemFlags::NONTEMPORAL) {
+        assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memmove not supported");
-            // HACK(nox): This is inefficient but there is no nontemporal memmove.
-            let val = self.load(src, src_align);
-            let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
-            self.store_with_flags(val, ptr, dst_align, flags);
-            return;
-        }
         let size = self.intcast(size, self.type_isize(), false);
         let is_volatile = flags.contains(MemFlags::VOLATILE);
         let dst = self.pointercast(dst, self.type_i8p());

compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs

@@ -20,7 +20,7 @@ pub fn insert_reference_to_gdb_debug_scripts_section_global(bx: &mut Builder<'_,
         // LLVM to keep around the reference to the global.
         let indices = [bx.const_i32(0), bx.const_i32(0)];
         let element = bx.inbounds_gep(gdb_debug_scripts_section, &indices);
-        let volative_load_instruction = bx.volatile_load(element);
+        let volative_load_instruction = bx.volatile_load(bx.type_i8(), element);
         unsafe {
             llvm::LLVMSetAlignment(volative_load_instruction, 1);
         }

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -162,11 +162,14 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
 
             sym::volatile_load | sym::unaligned_volatile_load => {
                 let tp_ty = substs.type_at(0);
-                let mut ptr = args[0].immediate();
+                let ptr = args[0].immediate();
-                if let PassMode::Cast(ty) = fn_abi.ret.mode {
+                let load = if let PassMode::Cast(ty) = fn_abi.ret.mode {
-                    ptr = self.pointercast(ptr, self.type_ptr_to(ty.llvm_type(self)));
+                    let llty = ty.llvm_type(self);
-                }
+                    let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
-                let load = self.volatile_load(ptr);
+                    self.volatile_load(llty, ptr)
+                } else {
+                    self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
+                };
                 let align = if name == sym::unaligned_volatile_load {
                     1
                 } else {
@@ -319,9 +322,9 @@ impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
                     let integer_ty = self.type_ix(layout.size.bits());
                     let ptr_ty = self.type_ptr_to(integer_ty);
                     let a_ptr = self.bitcast(a, ptr_ty);
-                    let a_val = self.load(a_ptr, layout.align.abi);
+                    let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
                     let b_ptr = self.bitcast(b, ptr_ty);
-                    let b_val = self.load(b_ptr, layout.align.abi);
+                    let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
                     self.icmp(IntPredicate::IntEQ, a_val, b_val)
                 } else {
                     let i8p_ty = self.type_i8p();
@@ -540,7 +543,7 @@ fn codegen_msvc_try(
         // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
         let flags = bx.const_i32(8);
         let funclet = catchpad_rust.catch_pad(cs, &[tydesc, flags, slot]);
-        let ptr = catchpad_rust.load(slot, ptr_align);
+        let ptr = catchpad_rust.load(bx.type_i8p(), slot, ptr_align);
         catchpad_rust.call(catch_func, &[data, ptr], Some(&funclet));
         catchpad_rust.catch_ret(&funclet, caught.llbb());
 

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -1385,7 +1385,12 @@ extern "C" {
         Val: &'a Value,
         Name: *const c_char,
     ) -> &'a Value;
-    pub fn LLVMBuildLoad(B: &Builder<'a>, PointerVal: &'a Value, Name: *const c_char) -> &'a Value;
+    pub fn LLVMBuildLoad2(
+        B: &Builder<'a>,
+        Ty: &'a Type,
+        PointerVal: &'a Value,
+        Name: *const c_char,
+    ) -> &'a Value;
 
     pub fn LLVMBuildStore(B: &Builder<'a>, Val: &'a Value, Ptr: &'a Value) -> &'a Value;
 
@@ -1631,6 +1636,7 @@ extern "C" {
     // Atomic Operations
     pub fn LLVMRustBuildAtomicLoad(
         B: &Builder<'a>,
+        ElementType: &'a Type,
         PointerVal: &'a Value,
         Name: *const c_char,
         Order: AtomicOrdering,

compiler/rustc_codegen_llvm/src/va_arg.rs

@@ -32,14 +32,15 @@ fn emit_direct_ptr_va_arg(
     slot_size: Align,
     allow_higher_align: bool,
 ) -> (&'ll Value, Align) {
-    let va_list_ptr_ty = bx.cx().type_ptr_to(bx.cx.type_i8p());
+    let va_list_ty = bx.type_i8p();
+    let va_list_ptr_ty = bx.type_ptr_to(va_list_ty);
     let va_list_addr = if list.layout.llvm_type(bx.cx) != va_list_ptr_ty {
         bx.bitcast(list.immediate(), va_list_ptr_ty)
     } else {
         list.immediate()
     };
 
-    let ptr = bx.load(va_list_addr, bx.tcx().data_layout.pointer_align.abi);
+    let ptr = bx.load(va_list_ty, va_list_addr, bx.tcx().data_layout.pointer_align.abi);
 
     let (addr, addr_align) = if allow_higher_align && align > slot_size {
         (round_pointer_up_to_alignment(bx, ptr, align, bx.cx().type_i8p()), align)
@@ -82,10 +83,10 @@ fn emit_ptr_va_arg(
     let (addr, addr_align) =
         emit_direct_ptr_va_arg(bx, list, llty, size, align.abi, slot_size, allow_higher_align);
     if indirect {
-        let tmp_ret = bx.load(addr, addr_align);
+        let tmp_ret = bx.load(llty, addr, addr_align);
-        bx.load(tmp_ret, align.abi)
+        bx.load(bx.cx.layout_of(target_ty).llvm_type(bx.cx), tmp_ret, align.abi)
     } else {
-        bx.load(addr, addr_align)
+        bx.load(llty, addr, addr_align)
     }
 }
 
@@ -118,7 +119,7 @@ fn emit_aapcs_va_arg(
     };
 
     // if the offset >= 0 then the value will be on the stack
-    let mut reg_off_v = bx.load(reg_off, offset_align);
+    let mut reg_off_v = bx.load(bx.type_i32(), reg_off, offset_align);
     let use_stack = bx.icmp(IntPredicate::IntSGE, reg_off_v, zero);
     bx.cond_br(use_stack, &on_stack.llbb(), &maybe_reg.llbb());
 
@@ -139,8 +140,9 @@ fn emit_aapcs_va_arg(
     let use_stack = maybe_reg.icmp(IntPredicate::IntSGT, new_reg_off_v, zero);
     maybe_reg.cond_br(use_stack, &on_stack.llbb(), &in_reg.llbb());
 
+    let top_type = bx.type_i8p();
     let top = in_reg.struct_gep(va_list_addr, reg_top_index);
-    let top = in_reg.load(top, bx.tcx().data_layout.pointer_align.abi);
+    let top = in_reg.load(top_type, top, bx.tcx().data_layout.pointer_align.abi);
 
     // reg_value = *(@top + reg_off_v);
     let mut reg_addr = in_reg.gep(top, &[reg_off_v]);
@@ -149,8 +151,9 @@ fn emit_aapcs_va_arg(
         let offset = bx.const_i32((slot_size - layout.size.bytes()) as i32);
         reg_addr = in_reg.gep(reg_addr, &[offset]);
     }
-    let reg_addr = in_reg.bitcast(reg_addr, bx.cx.type_ptr_to(layout.llvm_type(bx)));
+    let reg_type = layout.llvm_type(bx);
-    let reg_value = in_reg.load(reg_addr, layout.align.abi);
+    let reg_addr = in_reg.bitcast(reg_addr, bx.cx.type_ptr_to(reg_type));
+    let reg_value = in_reg.load(reg_type, reg_addr, layout.align.abi);
     in_reg.br(&end.llbb());
 
     // On Stack block

compiler/rustc_codegen_ssa/src/meth.rs

@@ -20,10 +20,11 @@ impl<'a, 'tcx> VirtualIndex {
         // Load the data pointer from the object.
         debug!("get_fn({:?}, {:?})", llvtable, self);
 
-        let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(bx.fn_ptr_backend_type(fn_abi)));
+        let llty = bx.fn_ptr_backend_type(fn_abi);
+        let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(llty));
         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
         let gep = bx.inbounds_gep(llvtable, &[bx.const_usize(self.0)]);
-        let ptr = bx.load(gep, ptr_align);
+        let ptr = bx.load(llty, gep, ptr_align);
         bx.nonnull_metadata(ptr);
         // Vtable loads are invariant.
         bx.set_invariant_load(ptr);
@@ -38,10 +39,11 @@ impl<'a, 'tcx> VirtualIndex {
         // Load the data pointer from the object.
         debug!("get_int({:?}, {:?})", llvtable, self);
 
-        let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(bx.type_isize()));
+        let llty = bx.type_isize();
+        let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(llty));
         let usize_align = bx.tcx().data_layout.pointer_align.abi;
         let gep = bx.inbounds_gep(llvtable, &[bx.const_usize(self.0)]);
-        let ptr = bx.load(gep, usize_align);
+        let ptr = bx.load(llty, gep, usize_align);
         // Vtable loads are invariant.
         bx.set_invariant_load(ptr);
         ptr

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -260,7 +260,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             PassMode::Direct(_) | PassMode::Pair(..) => {
                 let op = self.codegen_consume(&mut bx, mir::Place::return_place().as_ref());
                 if let Ref(llval, _, align) = op.val {
-                    bx.load(llval, align)
+                    bx.load(bx.backend_type(op.layout), llval, align)
                 } else {
                     op.immediate_or_packed_pair(&mut bx)
                 }
@@ -287,8 +287,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         llval
                     }
                 };
-                let addr = bx.pointercast(llslot, bx.type_ptr_to(bx.cast_backend_type(&cast_ty)));
+                let ty = bx.cast_backend_type(&cast_ty);
-                bx.load(addr, self.fn_abi.ret.layout.align.abi)
+                let addr = bx.pointercast(llslot, bx.type_ptr_to(ty));
+                bx.load(ty, addr, self.fn_abi.ret.layout.align.abi)
             }
         };
         bx.ret(llval);
@@ -1086,15 +1087,16 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         if by_ref && !arg.is_indirect() {
             // Have to load the argument, maybe while casting it.
             if let PassMode::Cast(ty) = arg.mode {
-                let addr = bx.pointercast(llval, bx.type_ptr_to(bx.cast_backend_type(&ty)));
+                let llty = bx.cast_backend_type(&ty);
-                llval = bx.load(addr, align.min(arg.layout.align.abi));
+                let addr = bx.pointercast(llval, bx.type_ptr_to(llty));
+                llval = bx.load(llty, addr, align.min(arg.layout.align.abi));
             } else {
                 // We can't use `PlaceRef::load` here because the argument
                 // may have a type we don't treat as immediate, but the ABI
                 // used for this call is passing it by-value. In that case,
                 // the load would just produce `OperandValue::Ref` instead
                 // of the `OperandValue::Immediate` we need for the call.
-                llval = bx.load(llval, align);
+                llval = bx.load(bx.backend_type(arg.layout), llval, align);
                 if let abi::Abi::Scalar(ref scalar) = arg.layout.abi {
                     if scalar.is_bool() {
                         bx.range_metadata(llval, 0..2);

compiler/rustc_codegen_ssa/src/mir/debuginfo.rs

@@ -274,7 +274,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 match *elem {
                     mir::ProjectionElem::Deref => {
                         indirect_offsets.push(Size::ZERO);
-                        place = place.project_deref(bx);
+                        place = bx.load_operand(place).deref(bx.cx());
                     }
                     mir::ProjectionElem::Field(field, _) => {
                         let i = field.index();

compiler/rustc_codegen_ssa/src/mir/intrinsic.rs

@@ -448,15 +448,14 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                             if ty.is_unsafe_ptr() {
                                 // Some platforms do not support atomic operations on pointers,
                                 // so we cast to integer first...
-                                let ptr_llty = bx.type_ptr_to(bx.type_isize());
+                                let llty = bx.type_isize();
+                                let ptr_llty = bx.type_ptr_to(llty);
                                 source = bx.pointercast(source, ptr_llty);
-                            }
+                                let result = bx.atomic_load(llty, source, order, size);
-                            let result = bx.atomic_load(source, order, size);
-                            if ty.is_unsafe_ptr() {
                                 // ... and then cast the result back to a pointer
                                 bx.inttoptr(result, bx.backend_type(layout))
                             } else {
-                                result
+                                bx.atomic_load(bx.backend_type(layout), source, order, size)
                             }
                         } else {
                             return invalid_monomorphization(ty);

compiler/rustc_codegen_ssa/src/mir/operand.rs

@@ -289,6 +289,14 @@ impl<'a, 'tcx, V: CodegenObject> OperandValue<V> {
         }
         match self {
             OperandValue::Ref(r, None, source_align) => {
+                if flags.contains(MemFlags::NONTEMPORAL) {
+                    // HACK(nox): This is inefficient but there is no nontemporal memcpy.
+                    let ty = bx.backend_type(dest.layout);
+                    let ptr = bx.pointercast(r, bx.type_ptr_to(ty));
+                    let val = bx.load(ty, ptr, source_align);
+                    bx.store_with_flags(val, dest.llval, dest.align, flags);
+                    return;
+                }
                 base::memcpy_ty(bx, dest.llval, dest.align, r, source_align, dest.layout, flags)
             }
             OperandValue::Ref(_, Some(_), _) => {

compiler/rustc_codegen_ssa/src/mir/place.rs

@@ -402,18 +402,6 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
         downcast
     }
 
-    pub fn project_deref<Bx: BuilderMethods<'a, 'tcx, Value = V>>(&self, bx: &mut Bx) -> Self {
-        let target_ty = self.layout.ty.builtin_deref(true).expect("failed to deref");
-        let layout = bx.layout_of(target_ty.ty);
-
-        PlaceRef {
-            llval: bx.load(self.llval, self.align),
-            llextra: None,
-            layout,
-            align: layout.align.abi,
-        }
-    }
-
     pub fn storage_live<Bx: BuilderMethods<'a, 'tcx, Value = V>>(&self, bx: &mut Bx) {
         bx.lifetime_start(self.llval, self.layout.size);
     }

compiler/rustc_codegen_ssa/src/traits/builder.rs

@@ -137,9 +137,15 @@ pub trait BuilderMethods<'a, 'tcx>:
     fn dynamic_alloca(&mut self, ty: Self::Type, align: Align) -> Self::Value;
     fn array_alloca(&mut self, ty: Self::Type, len: Self::Value, align: Align) -> Self::Value;
 
-    fn load(&mut self, ptr: Self::Value, align: Align) -> Self::Value;
+    fn load(&mut self, ty: Self::Type, ptr: Self::Value, align: Align) -> Self::Value;
-    fn volatile_load(&mut self, ptr: Self::Value) -> Self::Value;
+    fn volatile_load(&mut self, ty: Self::Type, ptr: Self::Value) -> Self::Value;
-    fn atomic_load(&mut self, ptr: Self::Value, order: AtomicOrdering, size: Size) -> Self::Value;
+    fn atomic_load(
+        &mut self,
+        ty: Self::Type,
+        ptr: Self::Value,
+        order: AtomicOrdering,
+        size: Size,
+    ) -> Self::Value;
     fn load_operand(&mut self, place: PlaceRef<'tcx, Self::Value>)
     -> OperandRef<'tcx, Self::Value>;
 

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -349,11 +349,10 @@ extern "C" void LLVMRustSetFastMath(LLVMValueRef V) {
 }
 
 extern "C" LLVMValueRef
-LLVMRustBuildAtomicLoad(LLVMBuilderRef B, LLVMValueRef Source, const char *Name,
+LLVMRustBuildAtomicLoad(LLVMBuilderRef B, LLVMTypeRef Ty, LLVMValueRef Source,
-                        LLVMAtomicOrdering Order) {
+                        const char *Name, LLVMAtomicOrdering Order) {
   Value *Ptr = unwrap(Source);
-  Type *Ty = Ptr->getType()->getPointerElementType();
+  LoadInst *LI = unwrap(B)->CreateLoad(unwrap(Ty), Ptr, Name);
-  LoadInst *LI = unwrap(B)->CreateLoad(Ty, Ptr, Name);
   LI->setAtomic(fromRust(Order));
   return wrap(LI);
 }