Merge commit 'b7091eca6d8eb0fe88b58cc9a7aec405d8de5b85' into subtree-update_cg_gcc_2025-06-28

6 files changed, 76 insertions, 67 deletions

diff --git a/compiler/rustc_codegen_gcc/src/builder.rs b/compiler/rustc_codegen_gcc/src/builder.rs
index 7852aebe0c2..a4454cd9c73 100644
--- a/compiler/rustc_codegen_gcc/src/builder.rs
+++ b/compiler/rustc_codegen_gcc/src/builder.rs
@@ -520,8 +520,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         self.block
     }
 
-    fn append_block(cx: &'a CodegenCx<'gcc, 'tcx>, func: RValue<'gcc>, name: &str) -> Block<'gcc> {
-        let func = cx.rvalue_as_function(func);
+    fn append_block(_: &'a CodegenCx<'gcc, 'tcx>, func: Function<'gcc>, name: &str) -> Block<'gcc> {
         func.new_block(name)
     }
 
@@ -782,6 +781,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
                 return self.context.new_call(self.location, fmod, &[a, b]);
             }
             TypeKind::FP128 => {
+                // TODO(antoyo): use get_simple_function_f128_2args.
                 let f128_type = self.type_f128();
                 let fmodf128 = self.context.new_function(
                     None,
@@ -938,22 +938,36 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> {
         let block = self.llbb();
         let function = block.get_function();
+        // NOTE(FractalFir): In some cases, we *should* skip the call to get_aligned.
+        // For example, calling `get_aligned` on a i8 is pointless(since it can only be 1 aligned)
+        // Calling get_aligned on a `u128`/`i128` causes the attribute to become "stacked"
+        //
+        // From GCCs perspective:
+        // __int128_t  __attribute__((aligned(16)))  __attribute__((aligned(16)))
+        // and:
+        // __int128_t  __attribute__((aligned(16)))
+        // are 2 distinct, incompatible types.
+        //
+        // So, we skip the call to `get_aligned` in such a case. *Ideally*, we could do this for all the types,
+        // but the GCC APIs to facilitate this just aren't quite there yet.
+
+        // This checks that we only skip `get_aligned` on 128 bit ints if they have the correct alignment.
+        // Otherwise, this may be an under-aligned load, so we will still call get_aligned.
+        let mut can_skip_align = (pointee_ty == self.cx.u128_type
+            || pointee_ty == self.cx.i128_type)
+            && align == self.int128_align;
+        // We can skip the call to `get_aligned` for byte-sized types with alignment of 1.
+        can_skip_align = can_skip_align
+            || (pointee_ty == self.cx.u8_type || pointee_ty == self.cx.i8_type)
+                && align.bytes() == 1;
+        // Skip the call to `get_aligned` when possible.
+        let aligned_type =
+            if can_skip_align { pointee_ty } else { pointee_ty.get_aligned(align.bytes()) };
+
+        let ptr = self.context.new_cast(self.location, ptr, aligned_type.make_pointer());
         // NOTE: instead of returning the dereference here, we have to assign it to a variable in
         // the current basic block. Otherwise, it could be used in another basic block, causing a
         // dereference after a drop, for instance.
-        // FIXME(antoyo): this check that we don't call get_aligned() a second time on a type.
-        // Ideally, we shouldn't need to do this check.
-        // FractalFir: the `align == self.int128_align` check ensures we *do* call `get_aligned` if
-        // the alignment of a `u128`/`i128` is not the one mandated by the ABI. This ensures we handle
-        // under-aligned loads correctly.
-        let aligned_type = if (pointee_ty == self.cx.u128_type || pointee_ty == self.cx.i128_type)
-            && align == self.int128_align
-        {
-            pointee_ty
-        } else {
-            pointee_ty.get_aligned(align.bytes())
-        };
-        let ptr = self.context.new_cast(self.location, ptr, aligned_type.make_pointer());
         let deref = ptr.dereference(self.location).to_rvalue();
         let loaded_value = function.new_local(
             self.location,
@@ -1105,7 +1119,13 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         // TODO(antoyo)
     }
 
-    fn store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> {
+    fn store(&mut self, mut val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> {
+        if self.structs_as_pointer.borrow().contains(&val) {
+            // NOTE: hack to workaround a limitation of the rustc API: see comment on
+            // CodegenCx.structs_as_pointer
+            val = val.dereference(self.location).to_rvalue();
+        }
+
         self.store_with_flags(val, ptr, align, MemFlags::empty())
     }
 
@@ -1551,16 +1571,13 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         aggregate_value
     }
 
-    fn set_personality_fn(&mut self, _personality: RValue<'gcc>) {
+    fn set_personality_fn(&mut self, _personality: Function<'gcc>) {
         #[cfg(feature = "master")]
-        {
-            let personality = self.rvalue_as_function(_personality);
-            self.current_func().set_personality_function(personality);
-        }
+        self.current_func().set_personality_function(_personality);
     }
 
     #[cfg(feature = "master")]
-    fn cleanup_landing_pad(&mut self, pers_fn: RValue<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) {
+    fn cleanup_landing_pad(&mut self, pers_fn: Function<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) {
         self.set_personality_fn(pers_fn);
 
         // NOTE: insert the current block in a variable so that a later call to invoke knows to
@@ -1581,7 +1598,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     #[cfg(not(feature = "master"))]
-    fn cleanup_landing_pad(&mut self, _pers_fn: RValue<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) {
+    fn cleanup_landing_pad(&mut self, _pers_fn: Function<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) {
         let value1 = self
             .current_func()
             .new_local(self.location, self.u8_type.make_pointer(), "landing_pad0")
@@ -1591,7 +1608,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         (value1, value2)
     }
 
-    fn filter_landing_pad(&mut self, pers_fn: RValue<'gcc>) {
+    fn filter_landing_pad(&mut self, pers_fn: Function<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) {
         // TODO(antoyo): generate the correct landing pad
         self.cleanup_landing_pad(pers_fn);
     }
diff --git a/compiler/rustc_codegen_gcc/src/common.rs b/compiler/rustc_codegen_gcc/src/common.rs
index 58ff2f1f8f0..fdd47821b51 100644
--- a/compiler/rustc_codegen_gcc/src/common.rs
+++ b/compiler/rustc_codegen_gcc/src/common.rs
@@ -234,19 +234,6 @@ impl<'gcc, 'tcx> ConstCodegenMethods for CodegenCx<'gcc, 'tcx> {
         match cv {
             Scalar::Int(int) => {
                 let data = int.to_bits(layout.size(self));
-
-                // FIXME(antoyo): there's some issues with using the u128 code that follows, so hard-code
-                // the paths for floating-point values.
-                // TODO: Remove this code?
-                /*if ty == self.float_type {
-                    return self
-                        .context
-                        .new_rvalue_from_double(ty, f32::from_bits(data as u32) as f64);
-                }
-                if ty == self.double_type {
-                    return self.context.new_rvalue_from_double(ty, f64::from_bits(data as u64));
-                }*/
-
                 let value = self.const_uint_big(self.type_ix(bitsize), data);
                 let bytesize = layout.size(self).bytes();
                 if bitsize > 1 && ty.is_integral() && bytesize as u32 == ty.get_size() {
diff --git a/compiler/rustc_codegen_gcc/src/context.rs b/compiler/rustc_codegen_gcc/src/context.rs
index ff141ad365b..1d029811dfe 100644
--- a/compiler/rustc_codegen_gcc/src/context.rs
+++ b/compiler/rustc_codegen_gcc/src/context.rs
@@ -118,14 +118,15 @@ pub struct CodegenCx<'gcc, 'tcx> {
     /// A counter that is used for generating local symbol names
     local_gen_sym_counter: Cell<usize>,
 
-    eh_personality: Cell<Option<RValue<'gcc>>>,
+    eh_personality: Cell<Option<Function<'gcc>>>,
     #[cfg(feature = "master")]
     pub rust_try_fn: Cell<Option<(Type<'gcc>, Function<'gcc>)>>,
 
     pub pointee_infos: RefCell<FxHashMap<(Ty<'tcx>, Size), Option<PointeeInfo>>>,
 
     /// NOTE: a hack is used because the rustc API is not suitable to libgccjit and as such,
-    /// `const_undef()` returns struct as pointer so that they can later be assigned a value.
+    /// `const_undef()` returns struct as pointer so that they can later be assigned a value (in
+    /// e.g. Builder::insert_value).
     /// As such, this set remembers which of these pointers were returned by this function so that
     /// they can be dereferenced later.
     /// FIXME(antoyo): fix the rustc API to avoid having this hack.
@@ -155,6 +156,13 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
                 .layout_of(ty::TypingEnv::fully_monomorphized().as_query_input(rust_type))
                 .unwrap();
             let align = layout.align.abi.bytes();
+            // For types with size 1, the alignment can be 1 and only 1
+            // So, we can skip the call to ``get_aligned`.
+            // In the future, we can add a GCC API to query the type align,
+            // and call `get_aligned` if and only if that differs from Rust's expectations.
+            if layout.size.bytes() == 1 {
+                return context.new_c_type(ctype);
+            }
             #[cfg(feature = "master")]
             {
                 context.new_c_type(ctype).get_aligned(align)
@@ -373,8 +381,7 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
 impl<'gcc, 'tcx> BackendTypes for CodegenCx<'gcc, 'tcx> {
     type Value = RValue<'gcc>;
     type Metadata = RValue<'gcc>;
-    // TODO(antoyo): change to Function<'gcc>.
-    type Function = RValue<'gcc>;
+    type Function = Function<'gcc>;
 
     type BasicBlock = Block<'gcc>;
     type Type = Type<'gcc>;
@@ -392,11 +399,10 @@ impl<'gcc, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         &self.vtables
     }
 
-    fn get_fn(&self, instance: Instance<'tcx>) -> RValue<'gcc> {
+    fn get_fn(&self, instance: Instance<'tcx>) -> Function<'gcc> {
         let func = get_fn(self, instance);
         *self.current_func.borrow_mut() = Some(func);
-        // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API.
-        unsafe { std::mem::transmute(func) }
+        func
     }
 
     fn get_fn_addr(&self, instance: Instance<'tcx>) -> RValue<'gcc> {
@@ -420,7 +426,7 @@ impl<'gcc, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         ptr
     }
 
-    fn eh_personality(&self) -> RValue<'gcc> {
+    fn eh_personality(&self) -> Function<'gcc> {
         // The exception handling personality function.
         //
         // If our compilation unit has the `eh_personality` lang item somewhere
@@ -458,9 +464,7 @@ impl<'gcc, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
                 let symbol_name = tcx.symbol_name(instance).name;
                 let fn_abi = self.fn_abi_of_instance(instance, ty::List::empty());
                 self.linkage.set(FunctionType::Extern);
-                let func = self.declare_fn(symbol_name, fn_abi);
-                let func: RValue<'gcc> = unsafe { std::mem::transmute(func) };
-                func
+                self.declare_fn(symbol_name, fn_abi)
             }
             _ => {
                 let name = if wants_msvc_seh(self.sess()) {
@@ -468,8 +472,7 @@ impl<'gcc, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
                 } else {
                     "rust_eh_personality"
                 };
-                let func = self.declare_func(name, self.type_i32(), &[], true);
-                unsafe { std::mem::transmute::<Function<'gcc>, RValue<'gcc>>(func) }
+                self.declare_func(name, self.type_i32(), &[], true)
             }
         };
         // TODO(antoyo): apply target cpu attributes.
@@ -481,11 +484,11 @@ impl<'gcc, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         self.tcx.sess
     }
 
-    fn set_frame_pointer_type(&self, _llfn: RValue<'gcc>) {
+    fn set_frame_pointer_type(&self, _llfn: Function<'gcc>) {
         // TODO(antoyo)
     }
 
-    fn apply_target_cpu_attr(&self, _llfn: RValue<'gcc>) {
+    fn apply_target_cpu_attr(&self, _llfn: Function<'gcc>) {
         // TODO(antoyo)
     }
 
diff --git a/compiler/rustc_codegen_gcc/src/debuginfo.rs b/compiler/rustc_codegen_gcc/src/debuginfo.rs
index 3a265fbc64f..4c8585192a1 100644
--- a/compiler/rustc_codegen_gcc/src/debuginfo.rs
+++ b/compiler/rustc_codegen_gcc/src/debuginfo.rs
@@ -1,7 +1,7 @@
 use std::ops::Range;
 use std::sync::Arc;
 
-use gccjit::{Location, RValue};
+use gccjit::{Function, Location, RValue};
 use rustc_abi::Size;
 use rustc_codegen_ssa::mir::debuginfo::{DebugScope, FunctionDebugContext, VariableKind};
 use rustc_codegen_ssa::traits::{DebugInfoBuilderMethods, DebugInfoCodegenMethods};
@@ -221,7 +221,7 @@ impl<'gcc, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         &self,
         instance: Instance<'tcx>,
         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
-        llfn: RValue<'gcc>,
+        llfn: Function<'gcc>,
         mir: &mir::Body<'tcx>,
     ) -> Option<FunctionDebugContext<'tcx, Self::DIScope, Self::DILocation>> {
         if self.sess().opts.debuginfo == DebugInfo::None {
@@ -272,7 +272,7 @@ impl<'gcc, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         &self,
         _instance: Instance<'tcx>,
         _fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
-        _maybe_definition_llfn: Option<RValue<'gcc>>,
+        _maybe_definition_llfn: Option<Function<'gcc>>,
     ) -> Self::DIScope {
         // TODO(antoyo): implement.
     }
diff --git a/compiler/rustc_codegen_gcc/src/declare.rs b/compiler/rustc_codegen_gcc/src/declare.rs
index bed82073e2c..691fd8729e3 100644
--- a/compiler/rustc_codegen_gcc/src/declare.rs
+++ b/compiler/rustc_codegen_gcc/src/declare.rs
@@ -94,7 +94,7 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         _fn_type: Type<'gcc>,
         #[cfg(feature = "master")] callconv: Option<FnAttribute<'gcc>>,
         #[cfg(not(feature = "master"))] callconv: Option<()>,
-    ) -> RValue<'gcc> {
+    ) -> Function<'gcc> {
         // TODO(antoyo): use the fn_type parameter.
         let const_string = self.context.new_type::<u8>().make_pointer().make_pointer();
         let return_type = self.type_i32();
@@ -111,8 +111,7 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         // NOTE: it is needed to set the current_func here as well, because get_fn() is not called
         // for the main function.
         *self.current_func.borrow_mut() = Some(func);
-        // FIXME(antoyo): this is a wrong cast. That requires changing the compiler API.
-        unsafe { std::mem::transmute(func) }
+        func
     }
 
     pub fn declare_fn(&self, name: &str, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Function<'gcc> {
diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
index 09132c34aae..4c10380fe74 100644
--- a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
+++ b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
@@ -4,7 +4,9 @@ mod simd;
 #[cfg(feature = "master")]
 use std::iter;
 
-use gccjit::{ComparisonOp, Function, FunctionType, RValue, ToRValue, Type, UnaryOp};
+#[cfg(feature = "master")]
+use gccjit::Type;
+use gccjit::{ComparisonOp, Function, FunctionType, RValue, ToRValue, UnaryOp};
 #[cfg(feature = "master")]
 use rustc_abi::ExternAbi;
 use rustc_abi::{BackendRepr, HasDataLayout};
@@ -300,6 +302,8 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
         let fn_args = instance.args;
 
         let simple = get_simple_intrinsic(self, name);
+        // TODO(antoyo): Only call get_simple_function_f128 and get_simple_function_f128_2args when
+        // it is the symbols for the supported f128 builtins.
         let simple_func = get_simple_function(self, name)
             .or_else(|| get_simple_function_f128(self, name))
             .or_else(|| get_simple_function_f128_2args(self, name));
@@ -441,7 +445,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
                 match int_type_width_signed(args[0].layout.ty, self) {
                     Some((width, signed)) => match name {
                         sym::ctlz | sym::cttz => {
-                            let func = self.current_func.borrow().expect("func");
+                            let func = self.current_func();
                             let then_block = func.new_block("then");
                             let else_block = func.new_block("else");
                             let after_block = func.new_block("after");
@@ -1109,7 +1113,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         // for (int counter = 0; value != 0; counter++) {
         //     value &= value - 1;
         // }
-        let func = self.current_func.borrow().expect("func");
+        let func = self.current_func();
         let loop_head = func.new_block("head");
         let loop_body = func.new_block("body");
         let loop_tail = func.new_block("tail");
@@ -1188,7 +1192,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let result_type = lhs.get_type();
         if signed {
             // Based on algorithm from: https://stackoverflow.com/a/56531252/389119
-            let func = self.current_func.borrow().expect("func");
+            let func = self.current_func();
             let res = func.new_local(self.location, result_type, "saturating_sum");
             let supports_native_type = self.is_native_int_type(result_type);
             let overflow = if supports_native_type {
@@ -1259,7 +1263,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let result_type = lhs.get_type();
         if signed {
             // Based on algorithm from: https://stackoverflow.com/a/56531252/389119
-            let func = self.current_func.borrow().expect("func");
+            let func = self.current_func();
             let res = func.new_local(self.location, result_type, "saturating_diff");
             let supports_native_type = self.is_native_int_type(result_type);
             let overflow = if supports_native_type {
@@ -1483,10 +1487,9 @@ fn gen_fn<'a, 'gcc, 'tcx>(
     // FIXME(eddyb) find a nicer way to do this.
     cx.linkage.set(FunctionType::Internal);
     let func = cx.declare_fn(name, fn_abi);
-    let func_val = unsafe { std::mem::transmute::<Function<'gcc>, RValue<'gcc>>(func) };
-    cx.set_frame_pointer_type(func_val);
-    cx.apply_target_cpu_attr(func_val);
-    let block = Builder::append_block(cx, func_val, "entry-block");
+    cx.set_frame_pointer_type(func);
+    cx.apply_target_cpu_attr(func);
+    let block = Builder::append_block(cx, func, "entry-block");
     let bx = Builder::build(cx, block);
     codegen(bx);
     (return_type, func)