Add create_stack_slot helper to FunctionCx

This will allow centrally handling a workaround for the lack of stack
alignment specifier in cranelift.

Author: bjorn3

src/abi/mod.rs

@@ -120,32 +120,25 @@ impl<'tcx> FunctionCx<'_, '_, 'tcx> {
         args: &[Value],
     ) -> Cow<'_, [Value]> {
         if self.tcx.sess.target.is_like_windows {
-            let (mut params, mut args): (Vec<_>, Vec<_>) =
-                params
-                    .into_iter()
-                    .zip(args)
-                    .map(|(param, &arg)| {
-                        if param.value_type == types::I128 {
-                            let arg_ptr = Pointer::stack_slot(self.bcx.create_sized_stack_slot(
-                                StackSlotData { kind: StackSlotKind::ExplicitSlot, size: 16 },
-                            ));
-                            arg_ptr.store(self, arg, MemFlags::trusted());
-                            (AbiParam::new(self.pointer_type), arg_ptr.get_addr(self))
-                        } else {
-                            (param, arg)
-                        }
-                    })
-                    .unzip();
+            let (mut params, mut args): (Vec<_>, Vec<_>) = params
+                .into_iter()
+                .zip(args)
+                .map(|(param, &arg)| {
+                    if param.value_type == types::I128 {
+                        let arg_ptr = self.create_stack_slot(16, 16);
+                        arg_ptr.store(self, arg, MemFlags::trusted());
+                        (AbiParam::new(self.pointer_type), arg_ptr.get_addr(self))
+                    } else {
+                        (param, arg)
+                    }
+                })
+                .unzip();
 
             let indirect_ret_val = returns.len() == 1 && returns[0].value_type == types::I128;
 
             if indirect_ret_val {
                 params.insert(0, AbiParam::new(self.pointer_type));
-                let ret_ptr =
-                    Pointer::stack_slot(self.bcx.create_sized_stack_slot(StackSlotData {
-                        kind: StackSlotKind::ExplicitSlot,
-                        size: 16,
-                    }));
+                let ret_ptr = self.create_stack_slot(16, 16);
                 args.insert(0, ret_ptr.get_addr(self));
                 self.lib_call_unadjusted(name, params, vec![], &args);
                 return Cow::Owned(vec![ret_ptr.load(self, types::I128, MemFlags::trusted())]);

src/abi/pass_mode.rs

@@ -189,16 +189,13 @@ pub(super) fn from_casted_value<'tcx>(
     let abi_params = cast_target_to_abi_params(cast);
     let abi_param_size: u32 = abi_params.iter().map(|param| param.value_type.bytes()).sum();
     let layout_size = u32::try_from(layout.size.bytes()).unwrap();
-    let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-        kind: StackSlotKind::ExplicitSlot,
-        // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
-        // specify stack slot alignment.
+    let ptr = fx.create_stack_slot(
         // Stack slot size may be bigger for example `[u8; 3]` which is packed into an `i32`.
         // It may also be smaller for example when the type is a wrapper around an integer with a
         // larger alignment than the integer.
-        size: (std::cmp::max(abi_param_size, layout_size) + 15) / 16 * 16,
-    });
-    let ptr = Pointer::stack_slot(stack_slot);
+        std::cmp::max(abi_param_size, layout_size),
+        u32::try_from(layout.align.pref.bytes()).unwrap(),
+    );
     let mut offset = 0;
     let mut block_params_iter = block_params.iter().copied();
     for param in abi_params {

src/cast.rs

@@ -104,11 +104,7 @@ pub(crate) fn clif_int_or_float_cast(
                     &[from],
                 )[0];
                 // FIXME(bytecodealliance/wasmtime#6104) use bitcast instead of store to get from i64x2 to i128
-                let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-                    kind: StackSlotKind::ExplicitSlot,
-                    size: 16,
-                });
-                let ret_ptr = Pointer::stack_slot(stack_slot);
+                let ret_ptr = fx.create_stack_slot(16, 16);
                 ret_ptr.store(fx, ret, MemFlags::trusted());
                 ret_ptr.load(fx, types::I128, MemFlags::trusted())
             } else {

src/common.rs

@@ -383,6 +383,16 @@ impl<'tcx> FunctionCx<'_, '_, 'tcx> {
         })
     }
 
+    pub(crate) fn create_stack_slot(&mut self, size: u32, align: u32) -> Pointer {
+        let stack_slot = self.bcx.create_sized_stack_slot(StackSlotData {
+            kind: StackSlotKind::ExplicitSlot,
+            // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
+            // specify stack slot alignment.
+            size: (size + 15) / 16 * 16,
+        });
+        Pointer::stack_slot(stack_slot)
+    }
+
     pub(crate) fn set_debug_loc(&mut self, source_info: mir::SourceInfo) {
         if let Some(debug_context) = &mut self.cx.debug_context {
             let (file, line, column) =

src/inline_asm.rs

@@ -878,13 +878,7 @@ fn call_inline_asm<'tcx>(
     inputs: Vec<(Size, Value)>,
     outputs: Vec<(Size, CPlace<'tcx>)>,
 ) {
-    let stack_slot = fx.bcx.func.create_sized_stack_slot(StackSlotData {
-        kind: StackSlotKind::ExplicitSlot,
-        size: u32::try_from(slot_size.bytes()).unwrap(),
-    });
-    if fx.clif_comments.enabled() {
-        fx.add_comment(stack_slot, "inline asm scratch slot");
-    }
+    let stack_slot = fx.create_stack_slot(u32::try_from(slot_size.bytes()).unwrap(), 16);
 
     let inline_asm_func = fx
         .module
@@ -904,15 +898,23 @@ fn call_inline_asm<'tcx>(
     }
 
     for (offset, value) in inputs {
-        fx.bcx.ins().stack_store(value, stack_slot, i32::try_from(offset.bytes()).unwrap());
+        stack_slot.offset(fx, i32::try_from(offset.bytes()).unwrap().into()).store(
+            fx,
+            value,
+            MemFlags::trusted(),
+        );
     }
 
-    let stack_slot_addr = fx.bcx.ins().stack_addr(fx.pointer_type, stack_slot, 0);
+    let stack_slot_addr = stack_slot.get_addr(fx);
     fx.bcx.ins().call(inline_asm_func, &[stack_slot_addr]);
 
     for (offset, place) in outputs {
         let ty = fx.clif_type(place.layout().ty).unwrap();
-        let value = fx.bcx.ins().stack_load(ty, stack_slot, i32::try_from(offset.bytes()).unwrap());
+        let value = stack_slot.offset(fx, i32::try_from(offset.bytes()).unwrap().into()).load(
+            fx,
+            ty,
+            MemFlags::trusted(),
+        );
         place.write_cvalue(fx, CValue::by_val(value, place.layout()));
     }
 }

src/value_and_place.rs

@@ -132,18 +132,11 @@ impl<'tcx> CValue<'tcx> {
                 (ptr.get_addr(fx), vtable)
             }
             CValueInner::ByValPair(data, vtable) => {
-                let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-                    kind: StackSlotKind::ExplicitSlot,
-                    // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
-                    // specify stack slot alignment.
-                    size: (u32::try_from(fx.target_config.pointer_type().bytes()).unwrap() + 15)
-                        / 16
-                        * 16,
-                });
-                let data_ptr = Pointer::stack_slot(stack_slot);
-                let mut flags = MemFlags::new();
-                flags.set_notrap();
-                data_ptr.store(fx, data, flags);
+                let data_ptr = fx.create_stack_slot(
+                    u32::try_from(fx.target_config.pointer_type().bytes()).unwrap(),
+                    u32::try_from(fx.target_config.pointer_type().bytes()).unwrap(),
+                );
+                data_ptr.store(fx, data, MemFlags::trusted());
 
                 (data_ptr.get_addr(fx), vtable)
             }
@@ -372,13 +365,11 @@ impl<'tcx> CPlace<'tcx> {
                 .fatal(format!("values of type {} are too big to store on the stack", layout.ty));
         }
 
-        let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-            kind: StackSlotKind::ExplicitSlot,
-            // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
-            // specify stack slot alignment.
-            size: (u32::try_from(layout.size.bytes()).unwrap() + 15) / 16 * 16,
-        });
-        CPlace { inner: CPlaceInner::Addr(Pointer::stack_slot(stack_slot), None), layout }
+        let stack_slot = fx.create_stack_slot(
+            u32::try_from(layout.size.bytes()).unwrap(),
+            u32::try_from(layout.align.pref.bytes()).unwrap(),
+        );
+        CPlace { inner: CPlaceInner::Addr(stack_slot, None), layout }
     }
 
     pub(crate) fn new_var(
@@ -543,13 +534,7 @@ impl<'tcx> CPlace<'tcx> {
                 _ if src_ty.is_vector() && dst_ty.is_vector() => codegen_bitcast(fx, dst_ty, data),
                 _ if src_ty.is_vector() || dst_ty.is_vector() => {
                     // FIXME(bytecodealliance/wasmtime#6104) do something more efficient for transmutes between vectors and integers.
-                    let stack_slot = fx.bcx.create_sized_stack_slot(StackSlotData {
-                        kind: StackSlotKind::ExplicitSlot,
-                        // FIXME Don't force the size to a multiple of 16 bytes once Cranelift gets a way to
-                        // specify stack slot alignment.
-                        size: (src_ty.bytes() + 15) / 16 * 16,
-                    });
-                    let ptr = Pointer::stack_slot(stack_slot);
+                    let ptr = fx.create_stack_slot(src_ty.bytes(), src_ty.bytes());
                     ptr.store(fx, data, MemFlags::trusted());
                     ptr.load(fx, dst_ty, MemFlags::trusted())
                 }