about summary refs log tree commit diff
path: root/compiler
diff options
context:
space:
mode:
Diffstat (limited to 'compiler')
-rw-r--r--compiler/rustc_builtin_macros/messages.ftl1
-rw-r--r--compiler/rustc_builtin_macros/src/autodiff.rs67
-rw-r--r--compiler/rustc_builtin_macros/src/errors.rs8
-rw-r--r--compiler/rustc_builtin_macros/src/lib.rs5
-rw-r--r--compiler/rustc_codegen_gcc/src/intrinsic/mod.rs53
-rw-r--r--compiler/rustc_codegen_llvm/src/intrinsic.rs80
-rw-r--r--compiler/rustc_codegen_ssa/src/mir/block.rs329
-rw-r--r--compiler/rustc_codegen_ssa/src/mir/intrinsic.rs34
-rw-r--r--compiler/rustc_codegen_ssa/src/traits/intrinsic.rs7
-rw-r--r--compiler/rustc_expand/src/expand.rs6
-rw-r--r--compiler/rustc_mir_build/src/builder/scope.rs112
-rw-r--r--compiler/rustc_parse/src/parser/stmt.rs13
-rw-r--r--compiler/rustc_passes/src/check_attr.rs2
-rw-r--r--compiler/rustc_resolve/src/diagnostics.rs50
-rw-r--r--compiler/rustc_span/src/symbol.rs5
15 files changed, 408 insertions, 364 deletions
diff --git a/compiler/rustc_builtin_macros/messages.ftl b/compiler/rustc_builtin_macros/messages.ftl
index 628bdee1129..d32e6f1558e 100644
--- a/compiler/rustc_builtin_macros/messages.ftl
+++ b/compiler/rustc_builtin_macros/messages.ftl
@@ -56,7 +56,6 @@ builtin_macros_assert_requires_expression = macro requires an expression as an a
 
 builtin_macros_autodiff = autodiff must be applied to function
 builtin_macros_autodiff_missing_config = autodiff requires at least a name and mode
-builtin_macros_autodiff_mode = unknown Mode: `{$mode}`. Use `Forward` or `Reverse`
 builtin_macros_autodiff_mode_activity = {$act} can not be used in {$mode} Mode
 builtin_macros_autodiff_not_build = this rustc version does not support autodiff
 builtin_macros_autodiff_number_activities = expected {$expected} activities, but found {$found}
diff --git a/compiler/rustc_builtin_macros/src/autodiff.rs b/compiler/rustc_builtin_macros/src/autodiff.rs
index 1ff4fc6aaab..dc3bb8ab52a 100644
--- a/compiler/rustc_builtin_macros/src/autodiff.rs
+++ b/compiler/rustc_builtin_macros/src/autodiff.rs
@@ -86,27 +86,23 @@ mod llvm_enzyme {
         ecx: &mut ExtCtxt<'_>,
         meta_item: &ThinVec<MetaItemInner>,
         has_ret: bool,
+        mode: DiffMode,
     ) -> AutoDiffAttrs {
         let dcx = ecx.sess.dcx();
-        let mode = name(&meta_item[1]);
-        let Ok(mode) = DiffMode::from_str(&mode) else {
-            dcx.emit_err(errors::AutoDiffInvalidMode { span: meta_item[1].span(), mode });
-            return AutoDiffAttrs::error();
-        };
 
         // Now we check, whether the user wants autodiff in batch/vector mode, or scalar mode.
         // If he doesn't specify an integer (=width), we default to scalar mode, thus width=1.
-        let mut first_activity = 2;
+        let mut first_activity = 1;
 
-        let width = if let [_, _, x, ..] = &meta_item[..]
+        let width = if let [_, x, ..] = &meta_item[..]
             && let Some(x) = width(x)
         {
-            first_activity = 3;
+            first_activity = 2;
             match x.try_into() {
                 Ok(x) => x,
                 Err(_) => {
                     dcx.emit_err(errors::AutoDiffInvalidWidth {
-                        span: meta_item[2].span(),
+                        span: meta_item[1].span(),
                         width: x,
                     });
                     return AutoDiffAttrs::error();
@@ -165,6 +161,24 @@ mod llvm_enzyme {
         ts.push(TokenTree::Token(comma.clone(), Spacing::Alone));
     }
 
+    pub(crate) fn expand_forward(
+        ecx: &mut ExtCtxt<'_>,
+        expand_span: Span,
+        meta_item: &ast::MetaItem,
+        item: Annotatable,
+    ) -> Vec<Annotatable> {
+        expand_with_mode(ecx, expand_span, meta_item, item, DiffMode::Forward)
+    }
+
+    pub(crate) fn expand_reverse(
+        ecx: &mut ExtCtxt<'_>,
+        expand_span: Span,
+        meta_item: &ast::MetaItem,
+        item: Annotatable,
+    ) -> Vec<Annotatable> {
+        expand_with_mode(ecx, expand_span, meta_item, item, DiffMode::Reverse)
+    }
+
     /// We expand the autodiff macro to generate a new placeholder function which passes
     /// type-checking and can be called by users. The function body of the placeholder function will
     /// later be replaced on LLVM-IR level, so the design of the body is less important and for now
@@ -198,11 +212,12 @@ mod llvm_enzyme {
     /// ```
     /// FIXME(ZuseZ4): Once autodiff is enabled by default, make this a doc comment which is checked
     /// in CI.
-    pub(crate) fn expand(
+    pub(crate) fn expand_with_mode(
         ecx: &mut ExtCtxt<'_>,
         expand_span: Span,
         meta_item: &ast::MetaItem,
         mut item: Annotatable,
+        mode: DiffMode,
     ) -> Vec<Annotatable> {
         if cfg!(not(llvm_enzyme)) {
             ecx.sess.dcx().emit_err(errors::AutoDiffSupportNotBuild { span: meta_item.span });
@@ -245,29 +260,41 @@ mod llvm_enzyme {
         // create TokenStream from vec elemtents:
         // meta_item doesn't have a .tokens field
         let mut ts: Vec<TokenTree> = vec![];
-        if meta_item_vec.len() < 2 {
-            // At the bare minimum, we need a fnc name and a mode, even for a dummy function with no
-            // input and output args.
+        if meta_item_vec.len() < 1 {
+            // At the bare minimum, we need a fnc name.
             dcx.emit_err(errors::AutoDiffMissingConfig { span: item.span() });
             return vec![item];
         }
 
-        meta_item_inner_to_ts(&meta_item_vec[1], &mut ts);
+        let mode_symbol = match mode {
+            DiffMode::Forward => sym::Forward,
+            DiffMode::Reverse => sym::Reverse,
+            _ => unreachable!("Unsupported mode: {:?}", mode),
+        };
+
+        // Insert mode token
+        let mode_token = Token::new(TokenKind::Ident(mode_symbol, false.into()), Span::default());
+        ts.insert(0, TokenTree::Token(mode_token, Spacing::Joint));
+        ts.insert(
+            1,
+            TokenTree::Token(Token::new(TokenKind::Comma, Span::default()), Spacing::Alone),
+        );
 
         // Now, if the user gave a width (vector aka batch-mode ad), then we copy it.
         // If it is not given, we default to 1 (scalar mode).
         let start_position;
         let kind: LitKind = LitKind::Integer;
         let symbol;
-        if meta_item_vec.len() >= 3
-            && let Some(width) = width(&meta_item_vec[2])
+        if meta_item_vec.len() >= 2
+            && let Some(width) = width(&meta_item_vec[1])
         {
-            start_position = 3;
+            start_position = 2;
             symbol = Symbol::intern(&width.to_string());
         } else {
-            start_position = 2;
+            start_position = 1;
             symbol = sym::integer(1);
         }
+
         let l: Lit = Lit { kind, symbol, suffix: None };
         let t = Token::new(TokenKind::Literal(l), Span::default());
         let comma = Token::new(TokenKind::Comma, Span::default());
@@ -289,7 +316,7 @@ mod llvm_enzyme {
         ts.pop();
         let ts: TokenStream = TokenStream::from_iter(ts);
 
-        let x: AutoDiffAttrs = from_ast(ecx, &meta_item_vec, has_ret);
+        let x: AutoDiffAttrs = from_ast(ecx, &meta_item_vec, has_ret, mode);
         if !x.is_active() {
             // We encountered an error, so we return the original item.
             // This allows us to potentially parse other attributes.
@@ -1017,4 +1044,4 @@ mod llvm_enzyme {
     }
 }
 
-pub(crate) use llvm_enzyme::expand;
+pub(crate) use llvm_enzyme::{expand_forward, expand_reverse};
diff --git a/compiler/rustc_builtin_macros/src/errors.rs b/compiler/rustc_builtin_macros/src/errors.rs
index 75d06a8df14..3a2e96a5e5a 100644
--- a/compiler/rustc_builtin_macros/src/errors.rs
+++ b/compiler/rustc_builtin_macros/src/errors.rs
@@ -181,14 +181,6 @@ mod autodiff {
     }
 
     #[derive(Diagnostic)]
-    #[diag(builtin_macros_autodiff_mode)]
-    pub(crate) struct AutoDiffInvalidMode {
-        #[primary_span]
-        pub(crate) span: Span,
-        pub(crate) mode: String,
-    }
-
-    #[derive(Diagnostic)]
     #[diag(builtin_macros_autodiff_width)]
     pub(crate) struct AutoDiffInvalidWidth {
         #[primary_span]
diff --git a/compiler/rustc_builtin_macros/src/lib.rs b/compiler/rustc_builtin_macros/src/lib.rs
index 9cd4d17059a..b16f3cff5cf 100644
--- a/compiler/rustc_builtin_macros/src/lib.rs
+++ b/compiler/rustc_builtin_macros/src/lib.rs
@@ -5,10 +5,10 @@
 #![allow(internal_features)]
 #![allow(rustc::diagnostic_outside_of_impl)]
 #![allow(rustc::untranslatable_diagnostic)]
+#![cfg_attr(not(bootstrap), feature(autodiff))]
 #![doc(html_root_url = "https://doc.rust-lang.org/nightly/nightly-rustc/")]
 #![doc(rust_logo)]
 #![feature(assert_matches)]
-#![feature(autodiff)]
 #![feature(box_patterns)]
 #![feature(decl_macro)]
 #![feature(if_let_guard)]
@@ -112,7 +112,8 @@ pub fn register_builtin_macros(resolver: &mut dyn ResolverExpand) {
 
     register_attr! {
         alloc_error_handler: alloc_error_handler::expand,
-        autodiff: autodiff::expand,
+        autodiff_forward: autodiff::expand_forward,
+        autodiff_reverse: autodiff::expand_reverse,
         bench: test::expand_bench,
         cfg_accessible: cfg_accessible::Expander,
         cfg_eval: cfg_eval::expand,
diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
index ba65c8205a5..1bcb891a250 100644
--- a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
+++ b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs
@@ -22,11 +22,11 @@ use rustc_codegen_ssa::traits::{
 };
 use rustc_middle::bug;
 #[cfg(feature = "master")]
-use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt};
+use rustc_middle::ty::layout::FnAbiOf;
 use rustc_middle::ty::layout::{HasTypingEnv, LayoutOf};
 use rustc_middle::ty::{self, Instance, Ty};
 use rustc_span::{Span, Symbol, sym};
-use rustc_target::callconv::{ArgAbi, FnAbi, PassMode};
+use rustc_target::callconv::{ArgAbi, PassMode};
 use rustc_target::spec::PanicStrategy;
 
 #[cfg(feature = "master")]
@@ -200,9 +200,8 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
     fn codegen_intrinsic_call(
         &mut self,
         instance: Instance<'tcx>,
-        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         args: &[OperandRef<'tcx, RValue<'gcc>>],
-        llresult: RValue<'gcc>,
+        result: PlaceRef<'tcx, RValue<'gcc>>,
         span: Span,
     ) -> Result<(), Instance<'tcx>> {
         let tcx = self.tcx;
@@ -221,7 +220,6 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
         let name_str = name.as_str();
 
         let llret_ty = self.layout_of(ret_ty).gcc_type(self);
-        let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 
         let simple = get_simple_intrinsic(self, name);
         let simple_func = get_simple_function(self, name);
@@ -271,7 +269,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
                     args[0].immediate(),
                     args[1].immediate(),
                     args[2].immediate(),
-                    llresult,
+                    result,
                 );
                 return Ok(());
             }
@@ -286,17 +284,10 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
             }
 
             sym::volatile_load | sym::unaligned_volatile_load => {
-                let tp_ty = fn_args.type_at(0);
                 let ptr = args[0].immediate();
-                let layout = self.layout_of(tp_ty);
-                let load = if let PassMode::Cast { cast: ref ty, pad_i32: _ } = fn_abi.ret.mode {
-                    let gcc_ty = ty.gcc_type(self);
-                    self.volatile_load(gcc_ty, ptr)
-                } else {
-                    self.volatile_load(layout.gcc_type(self), ptr)
-                };
+                let load = self.volatile_load(result.layout.gcc_type(self), ptr);
                 // TODO(antoyo): set alignment.
-                if let BackendRepr::Scalar(scalar) = layout.backend_repr {
+                if let BackendRepr::Scalar(scalar) = result.layout.backend_repr {
                     self.to_immediate_scalar(load, scalar)
                 } else {
                     load
@@ -511,16 +502,14 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tc
             _ => return Err(Instance::new_raw(instance.def_id(), instance.args)),
         };
 
-        if !fn_abi.ret.is_ignore() {
-            if let PassMode::Cast { cast: ref ty, .. } = fn_abi.ret.mode {
-                let ptr_llty = self.type_ptr_to(ty.gcc_type(self));
-                let ptr = self.pointercast(result.val.llval, ptr_llty);
-                self.store(value, ptr, result.val.align);
-            } else {
-                OperandRef::from_immediate_or_packed_pair(self, value, result.layout)
-                    .val
-                    .store(self, result);
-            }
+        if result.layout.ty.is_bool() {
+            OperandRef::from_immediate_or_packed_pair(self, value, result.layout)
+                .val
+                .store(self, result);
+        } else if !result.layout.ty.is_unit() {
+            let ptr_llty = self.type_ptr_to(result.layout.gcc_type(self));
+            let ptr = self.pointercast(result.val.llval, ptr_llty);
+            self.store(value, ptr, result.val.align);
         }
         Ok(())
     }
@@ -1230,14 +1219,13 @@ fn try_intrinsic<'a, 'b, 'gcc, 'tcx>(
     try_func: RValue<'gcc>,
     data: RValue<'gcc>,
     _catch_func: RValue<'gcc>,
-    dest: RValue<'gcc>,
+    dest: PlaceRef<'tcx, RValue<'gcc>>,
 ) {
     if bx.sess().panic_strategy() == PanicStrategy::Abort {
         bx.call(bx.type_void(), None, None, try_func, &[data], None, None);
         // Return 0 unconditionally from the intrinsic call;
         // we can never unwind.
-        let ret_align = bx.tcx.data_layout.i32_align.abi;
-        bx.store(bx.const_i32(0), dest, ret_align);
+        OperandValue::Immediate(bx.const_i32(0)).store(bx, dest);
     } else {
         if wants_msvc_seh(bx.sess()) {
             unimplemented!();
@@ -1261,12 +1249,12 @@ fn try_intrinsic<'a, 'b, 'gcc, 'tcx>(
 // functions in play. By calling a shim we're guaranteed that our shim will have
 // the right personality function.
 #[cfg(feature = "master")]
-fn codegen_gnu_try<'gcc>(
-    bx: &mut Builder<'_, 'gcc, '_>,
+fn codegen_gnu_try<'gcc, 'tcx>(
+    bx: &mut Builder<'_, 'gcc, 'tcx>,
     try_func: RValue<'gcc>,
     data: RValue<'gcc>,
     catch_func: RValue<'gcc>,
-    dest: RValue<'gcc>,
+    dest: PlaceRef<'tcx, RValue<'gcc>>,
 ) {
     let cx: &CodegenCx<'gcc, '_> = bx.cx;
     let (llty, func) = get_rust_try_fn(cx, &mut |mut bx| {
@@ -1322,8 +1310,7 @@ fn codegen_gnu_try<'gcc>(
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
     let ret = bx.call(llty, None, None, func, &[try_func, data, catch_func], None, None);
-    let i32_align = bx.tcx().data_layout.i32_align.abi;
-    bx.store(ret, dest, i32_align);
+    OperandValue::Immediate(ret).store(bx, dest);
 }
 
 // Helper function used to get a handle to the `__rust_try` function used to
diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs
index 5ca57375292..e8629aeebb9 100644
--- a/compiler/rustc_codegen_llvm/src/intrinsic.rs
+++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs
@@ -15,11 +15,10 @@ use rustc_middle::ty::{self, GenericArgsRef, Ty};
 use rustc_middle::{bug, span_bug};
 use rustc_span::{Span, Symbol, sym};
 use rustc_symbol_mangling::mangle_internal_symbol;
-use rustc_target::callconv::{FnAbi, PassMode};
 use rustc_target::spec::{HasTargetSpec, PanicStrategy};
 use tracing::debug;
 
-use crate::abi::{FnAbiLlvmExt, LlvmType};
+use crate::abi::FnAbiLlvmExt;
 use crate::builder::Builder;
 use crate::context::CodegenCx;
 use crate::llvm::{self, Metadata};
@@ -165,9 +164,8 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
     fn codegen_intrinsic_call(
         &mut self,
         instance: ty::Instance<'tcx>,
-        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         args: &[OperandRef<'tcx, &'ll Value>],
-        llresult: &'ll Value,
+        result: PlaceRef<'tcx, &'ll Value>,
         span: Span,
     ) -> Result<(), ty::Instance<'tcx>> {
         let tcx = self.tcx;
@@ -184,7 +182,6 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
         let name = tcx.item_name(def_id);
 
         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
-        let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 
         let simple = get_simple_intrinsic(self, name);
         let llval = match name {
@@ -255,7 +252,7 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
                     args[0].immediate(),
                     args[1].immediate(),
                     args[2].immediate(),
-                    llresult,
+                    result,
                 );
                 return Ok(());
             }
@@ -264,7 +261,7 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
                 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
             }
             sym::va_arg => {
-                match fn_abi.ret.layout.backend_repr {
+                match result.layout.backend_repr {
                     BackendRepr::Scalar(scalar) => {
                         match scalar.primitive() {
                             Primitive::Int(..) => {
@@ -299,18 +296,12 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
             }
 
             sym::volatile_load | sym::unaligned_volatile_load => {
-                let tp_ty = fn_args.type_at(0);
                 let ptr = args[0].immediate();
-                let load = if let PassMode::Cast { cast: ty, pad_i32: _ } = &fn_abi.ret.mode {
-                    let llty = ty.llvm_type(self);
-                    self.volatile_load(llty, ptr)
-                } else {
-                    self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
-                };
+                let load = self.volatile_load(result.layout.llvm_type(self), ptr);
                 let align = if name == sym::unaligned_volatile_load {
                     1
                 } else {
-                    self.align_of(tp_ty).bytes() as u32
+                    result.layout.align.abi.bytes() as u32
                 };
                 unsafe {
                     llvm::LLVMSetAlignment(load, align);
@@ -629,14 +620,12 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
             }
         };
 
-        if !fn_abi.ret.is_ignore() {
-            if let PassMode::Cast { .. } = &fn_abi.ret.mode {
-                self.store(llval, result.val.llval, result.val.align);
-            } else {
-                OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
-                    .val
-                    .store(self, result);
-            }
+        if result.layout.ty.is_bool() {
+            OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
+                .val
+                .store(self, result);
+        } else if !result.layout.ty.is_unit() {
+            self.store_to_place(llval, result.val);
         }
         Ok(())
     }
@@ -688,20 +677,19 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
     }
 }
 
-fn catch_unwind_intrinsic<'ll>(
-    bx: &mut Builder<'_, 'll, '_>,
+fn catch_unwind_intrinsic<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
     try_func: &'ll Value,
     data: &'ll Value,
     catch_func: &'ll Value,
-    dest: &'ll Value,
+    dest: PlaceRef<'tcx, &'ll Value>,
 ) {
     if bx.sess().panic_strategy() == PanicStrategy::Abort {
         let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
         bx.call(try_func_ty, None, None, try_func, &[data], None, None);
         // Return 0 unconditionally from the intrinsic call;
         // we can never unwind.
-        let ret_align = bx.tcx().data_layout.i32_align.abi;
-        bx.store(bx.const_i32(0), dest, ret_align);
+        OperandValue::Immediate(bx.const_i32(0)).store(bx, dest);
     } else if wants_msvc_seh(bx.sess()) {
         codegen_msvc_try(bx, try_func, data, catch_func, dest);
     } else if wants_wasm_eh(bx.sess()) {
@@ -720,12 +708,12 @@ fn catch_unwind_intrinsic<'ll>(
 // instructions are meant to work for all targets, as of the time of this
 // writing, however, LLVM does not recommend the usage of these new instructions
 // as the old ones are still more optimized.
-fn codegen_msvc_try<'ll>(
-    bx: &mut Builder<'_, 'll, '_>,
+fn codegen_msvc_try<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
     try_func: &'ll Value,
     data: &'ll Value,
     catch_func: &'ll Value,
-    dest: &'ll Value,
+    dest: PlaceRef<'tcx, &'ll Value>,
 ) {
     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
         bx.set_personality_fn(bx.eh_personality());
@@ -865,17 +853,16 @@ fn codegen_msvc_try<'ll>(
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
     let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
-    let i32_align = bx.tcx().data_layout.i32_align.abi;
-    bx.store(ret, dest, i32_align);
+    OperandValue::Immediate(ret).store(bx, dest);
 }
 
 // WASM's definition of the `rust_try` function.
-fn codegen_wasm_try<'ll>(
-    bx: &mut Builder<'_, 'll, '_>,
+fn codegen_wasm_try<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
     try_func: &'ll Value,
     data: &'ll Value,
     catch_func: &'ll Value,
-    dest: &'ll Value,
+    dest: PlaceRef<'tcx, &'ll Value>,
 ) {
     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
         bx.set_personality_fn(bx.eh_personality());
@@ -939,8 +926,7 @@ fn codegen_wasm_try<'ll>(
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
     let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
-    let i32_align = bx.tcx().data_layout.i32_align.abi;
-    bx.store(ret, dest, i32_align);
+    OperandValue::Immediate(ret).store(bx, dest);
 }
 
 // Definition of the standard `try` function for Rust using the GNU-like model
@@ -954,12 +940,12 @@ fn codegen_wasm_try<'ll>(
 // function calling it, and that function may already have other personality
 // functions in play. By calling a shim we're guaranteed that our shim will have
 // the right personality function.
-fn codegen_gnu_try<'ll>(
-    bx: &mut Builder<'_, 'll, '_>,
+fn codegen_gnu_try<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
     try_func: &'ll Value,
     data: &'ll Value,
     catch_func: &'ll Value,
-    dest: &'ll Value,
+    dest: PlaceRef<'tcx, &'ll Value>,
 ) {
     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
         // Codegens the shims described above:
@@ -1006,19 +992,18 @@ fn codegen_gnu_try<'ll>(
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
     let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
-    let i32_align = bx.tcx().data_layout.i32_align.abi;
-    bx.store(ret, dest, i32_align);
+    OperandValue::Immediate(ret).store(bx, dest);
 }
 
 // Variant of codegen_gnu_try used for emscripten where Rust panics are
 // implemented using C++ exceptions. Here we use exceptions of a specific type
 // (`struct rust_panic`) to represent Rust panics.
-fn codegen_emcc_try<'ll>(
-    bx: &mut Builder<'_, 'll, '_>,
+fn codegen_emcc_try<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
     try_func: &'ll Value,
     data: &'ll Value,
     catch_func: &'ll Value,
-    dest: &'ll Value,
+    dest: PlaceRef<'tcx, &'ll Value>,
 ) {
     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
         // Codegens the shims described above:
@@ -1089,8 +1074,7 @@ fn codegen_emcc_try<'ll>(
     // Note that no invoke is used here because by definition this function
     // can't panic (that's what it's catching).
     let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
-    let i32_align = bx.tcx().data_layout.i32_align.abi;
-    bx.store(ret, dest, i32_align);
+    OperandValue::Immediate(ret).store(bx, dest);
 }
 
 // Helper function to give a Block to a closure to codegen a shim function.
diff --git a/compiler/rustc_codegen_ssa/src/mir/block.rs b/compiler/rustc_codegen_ssa/src/mir/block.rs
index 922b8a5824b..1baab62ae43 100644
--- a/compiler/rustc_codegen_ssa/src/mir/block.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/block.rs
@@ -11,8 +11,8 @@ use rustc_middle::ty::print::{with_no_trimmed_paths, with_no_visible_paths};
 use rustc_middle::ty::{self, Instance, Ty};
 use rustc_middle::{bug, span_bug};
 use rustc_session::config::OptLevel;
+use rustc_span::Span;
 use rustc_span::source_map::Spanned;
-use rustc_span::{Span, sym};
 use rustc_target::callconv::{ArgAbi, FnAbi, PassMode};
 use tracing::{debug, info};
 
@@ -827,7 +827,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         helper: &TerminatorCodegenHelper<'tcx>,
         bx: &mut Bx,
         intrinsic: ty::IntrinsicDef,
-        instance: Option<Instance<'tcx>>,
+        instance: Instance<'tcx>,
         source_info: mir::SourceInfo,
         target: Option<mir::BasicBlock>,
         unwind: mir::UnwindAction,
@@ -836,58 +836,56 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         // Emit a panic or a no-op for `assert_*` intrinsics.
         // These are intrinsics that compile to panics so that we can get a message
         // which mentions the offending type, even from a const context.
-        if let Some(requirement) = ValidityRequirement::from_intrinsic(intrinsic.name) {
-            let ty = instance.unwrap().args.type_at(0);
-
-            let do_panic = !bx
-                .tcx()
-                .check_validity_requirement((requirement, bx.typing_env().as_query_input(ty)))
-                .expect("expect to have layout during codegen");
-
-            let layout = bx.layout_of(ty);
-
-            Some(if do_panic {
-                let msg_str = with_no_visible_paths!({
-                    with_no_trimmed_paths!({
-                        if layout.is_uninhabited() {
-                            // Use this error even for the other intrinsics as it is more precise.
-                            format!("attempted to instantiate uninhabited type `{ty}`")
-                        } else if requirement == ValidityRequirement::Zero {
-                            format!("attempted to zero-initialize type `{ty}`, which is invalid")
-                        } else {
-                            format!(
-                                "attempted to leave type `{ty}` uninitialized, which is invalid"
-                            )
-                        }
-                    })
-                });
-                let msg = bx.const_str(&msg_str);
+        let Some(requirement) = ValidityRequirement::from_intrinsic(intrinsic.name) else {
+            return None;
+        };
 
-                // Obtain the panic entry point.
-                let (fn_abi, llfn, instance) =
-                    common::build_langcall(bx, Some(source_info.span), LangItem::PanicNounwind);
+        let ty = instance.args.type_at(0);
 
-                // Codegen the actual panic invoke/call.
-                helper.do_call(
-                    self,
-                    bx,
-                    fn_abi,
-                    llfn,
-                    &[msg.0, msg.1],
-                    target.as_ref().map(|bb| (ReturnDest::Nothing, *bb)),
-                    unwind,
-                    &[],
-                    Some(instance),
-                    mergeable_succ,
-                )
-            } else {
-                // a NOP
-                let target = target.unwrap();
-                helper.funclet_br(self, bx, target, mergeable_succ)
-            })
-        } else {
-            None
+        let is_valid = bx
+            .tcx()
+            .check_validity_requirement((requirement, bx.typing_env().as_query_input(ty)))
+            .expect("expect to have layout during codegen");
+
+        if is_valid {
+            // a NOP
+            let target = target.unwrap();
+            return Some(helper.funclet_br(self, bx, target, mergeable_succ));
         }
+
+        let layout = bx.layout_of(ty);
+
+        let msg_str = with_no_visible_paths!({
+            with_no_trimmed_paths!({
+                if layout.is_uninhabited() {
+                    // Use this error even for the other intrinsics as it is more precise.
+                    format!("attempted to instantiate uninhabited type `{ty}`")
+                } else if requirement == ValidityRequirement::Zero {
+                    format!("attempted to zero-initialize type `{ty}`, which is invalid")
+                } else {
+                    format!("attempted to leave type `{ty}` uninitialized, which is invalid")
+                }
+            })
+        });
+        let msg = bx.const_str(&msg_str);
+
+        // Obtain the panic entry point.
+        let (fn_abi, llfn, instance) =
+            common::build_langcall(bx, Some(source_info.span), LangItem::PanicNounwind);
+
+        // Codegen the actual panic invoke/call.
+        Some(helper.do_call(
+            self,
+            bx,
+            fn_abi,
+            llfn,
+            &[msg.0, msg.1],
+            target.as_ref().map(|bb| (ReturnDest::Nothing, *bb)),
+            unwind,
+            &[],
+            Some(instance),
+            mergeable_succ,
+        ))
     }
 
     fn codegen_call_terminator(
@@ -903,42 +901,127 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         fn_span: Span,
         mergeable_succ: bool,
     ) -> MergingSucc {
-        let source_info = terminator.source_info;
-        let span = source_info.span;
+        let source_info = mir::SourceInfo { span: fn_span, ..terminator.source_info };
 
         // Create the callee. This is a fn ptr or zero-sized and hence a kind of scalar.
         let callee = self.codegen_operand(bx, func);
 
         let (instance, mut llfn) = match *callee.layout.ty.kind() {
-            ty::FnDef(def_id, args) => (
-                Some(ty::Instance::expect_resolve(
+            ty::FnDef(def_id, generic_args) => {
+                let instance = ty::Instance::expect_resolve(
                     bx.tcx(),
                     bx.typing_env(),
                     def_id,
-                    args,
+                    generic_args,
                     fn_span,
-                )),
-                None,
-            ),
+                );
+
+                let instance = match instance.def {
+                    // We don't need AsyncDropGlueCtorShim here because it is not `noop func`,
+                    // it is `func returning noop future`
+                    ty::InstanceKind::DropGlue(_, None) => {
+                        // Empty drop glue; a no-op.
+                        let target = target.unwrap();
+                        return helper.funclet_br(self, bx, target, mergeable_succ);
+                    }
+                    ty::InstanceKind::Intrinsic(def_id) => {
+                        let intrinsic = bx.tcx().intrinsic(def_id).unwrap();
+                        if let Some(merging_succ) = self.codegen_panic_intrinsic(
+                            &helper,
+                            bx,
+                            intrinsic,
+                            instance,
+                            source_info,
+                            target,
+                            unwind,
+                            mergeable_succ,
+                        ) {
+                            return merging_succ;
+                        }
+
+                        let result_layout =
+                            self.cx.layout_of(self.monomorphized_place_ty(destination.as_ref()));
+
+                        let (result, store_in_local) = if result_layout.is_zst() {
+                            (
+                                PlaceRef::new_sized(bx.const_undef(bx.type_ptr()), result_layout),
+                                None,
+                            )
+                        } else if let Some(local) = destination.as_local() {
+                            match self.locals[local] {
+                                LocalRef::Place(dest) => (dest, None),
+                                LocalRef::UnsizedPlace(_) => bug!("return type must be sized"),
+                                LocalRef::PendingOperand => {
+                                    // Currently, intrinsics always need a location to store
+                                    // the result, so we create a temporary `alloca` for the
+                                    // result.
+                                    let tmp = PlaceRef::alloca(bx, result_layout);
+                                    tmp.storage_live(bx);
+                                    (tmp, Some(local))
+                                }
+                                LocalRef::Operand(_) => {
+                                    bug!("place local already assigned to");
+                                }
+                            }
+                        } else {
+                            (self.codegen_place(bx, destination.as_ref()), None)
+                        };
+
+                        if result.val.align < result.layout.align.abi {
+                            // Currently, MIR code generation does not create calls
+                            // that store directly to fields of packed structs (in
+                            // fact, the calls it creates write only to temps).
+                            //
+                            // If someone changes that, please update this code path
+                            // to create a temporary.
+                            span_bug!(self.mir.span, "can't directly store to unaligned value");
+                        }
+
+                        let args: Vec<_> =
+                            args.iter().map(|arg| self.codegen_operand(bx, &arg.node)).collect();
+
+                        match self.codegen_intrinsic_call(bx, instance, &args, result, source_info)
+                        {
+                            Ok(()) => {
+                                if let Some(local) = store_in_local {
+                                    let op = bx.load_operand(result);
+                                    result.storage_dead(bx);
+                                    self.overwrite_local(local, LocalRef::Operand(op));
+                                    self.debug_introduce_local(bx, local);
+                                }
+
+                                return if let Some(target) = target {
+                                    helper.funclet_br(self, bx, target, mergeable_succ)
+                                } else {
+                                    bx.unreachable();
+                                    MergingSucc::False
+                                };
+                            }
+                            Err(instance) => {
+                                if intrinsic.must_be_overridden {
+                                    span_bug!(
+                                        fn_span,
+                                        "intrinsic {} must be overridden by codegen backend, but isn't",
+                                        intrinsic.name,
+                                    );
+                                }
+                                instance
+                            }
+                        }
+                    }
+                    _ => instance,
+                };
+
+                (Some(instance), None)
+            }
             ty::FnPtr(..) => (None, Some(callee.immediate())),
             _ => bug!("{} is not callable", callee.layout.ty),
         };
 
-        let def = instance.map(|i| i.def);
-
-        // We don't need AsyncDropGlueCtorShim here because it is not `noop func`,
-        // it is `func returning noop future`
-        if let Some(ty::InstanceKind::DropGlue(_, None)) = def {
-            // Empty drop glue; a no-op.
-            let target = target.unwrap();
-            return helper.funclet_br(self, bx, target, mergeable_succ);
-        }
-
         // FIXME(eddyb) avoid computing this if possible, when `instance` is
         // available - right now `sig` is only needed for getting the `abi`
         // and figuring out how many extra args were passed to a C-variadic `fn`.
         let sig = callee.layout.ty.fn_sig(bx.tcx());
-        let abi = sig.abi();
 
         let extra_args = &args[sig.inputs().skip_binder().len()..];
         let extra_args = bx.tcx().mk_type_list_from_iter(extra_args.iter().map(|op_arg| {
@@ -954,93 +1037,16 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         // The arguments we'll be passing. Plus one to account for outptr, if used.
         let arg_count = fn_abi.args.len() + fn_abi.ret.is_indirect() as usize;
 
-        let instance = match def {
-            Some(ty::InstanceKind::Intrinsic(def_id)) => {
-                let intrinsic = bx.tcx().intrinsic(def_id).unwrap();
-                if let Some(merging_succ) = self.codegen_panic_intrinsic(
-                    &helper,
-                    bx,
-                    intrinsic,
-                    instance,
-                    source_info,
-                    target,
-                    unwind,
-                    mergeable_succ,
-                ) {
-                    return merging_succ;
-                }
-
-                let mut llargs = Vec::with_capacity(1);
-                let ret_dest = self.make_return_dest(
-                    bx,
-                    destination,
-                    &fn_abi.ret,
-                    &mut llargs,
-                    Some(intrinsic),
-                );
-                let dest = match ret_dest {
-                    _ if fn_abi.ret.is_indirect() => llargs[0],
-                    ReturnDest::Nothing => bx.const_undef(bx.type_ptr()),
-                    ReturnDest::IndirectOperand(dst, _) | ReturnDest::Store(dst) => dst.val.llval,
-                    ReturnDest::DirectOperand(_) => {
-                        bug!("Cannot use direct operand with an intrinsic call")
-                    }
-                };
-
-                let args: Vec<_> =
-                    args.iter().map(|arg| self.codegen_operand(bx, &arg.node)).collect();
-
-                if matches!(intrinsic, ty::IntrinsicDef { name: sym::caller_location, .. }) {
-                    let location = self
-                        .get_caller_location(bx, mir::SourceInfo { span: fn_span, ..source_info });
-
-                    assert_eq!(llargs, []);
-                    if let ReturnDest::IndirectOperand(tmp, _) = ret_dest {
-                        location.val.store(bx, tmp);
-                    }
-                    self.store_return(bx, ret_dest, &fn_abi.ret, location.immediate());
-                    return helper.funclet_br(self, bx, target.unwrap(), mergeable_succ);
-                }
-
-                let instance = *instance.as_ref().unwrap();
-                match Self::codegen_intrinsic_call(bx, instance, fn_abi, &args, dest, span) {
-                    Ok(()) => {
-                        if let ReturnDest::IndirectOperand(dst, _) = ret_dest {
-                            self.store_return(bx, ret_dest, &fn_abi.ret, dst.val.llval);
-                        }
-
-                        return if let Some(target) = target {
-                            helper.funclet_br(self, bx, target, mergeable_succ)
-                        } else {
-                            bx.unreachable();
-                            MergingSucc::False
-                        };
-                    }
-                    Err(instance) => {
-                        if intrinsic.must_be_overridden {
-                            span_bug!(
-                                span,
-                                "intrinsic {} must be overridden by codegen backend, but isn't",
-                                intrinsic.name,
-                            );
-                        }
-                        Some(instance)
-                    }
-                }
-            }
-            _ => instance,
-        };
-
         let mut llargs = Vec::with_capacity(arg_count);
 
         // We still need to call `make_return_dest` even if there's no `target`, since
         // `fn_abi.ret` could be `PassMode::Indirect`, even if it is uninhabited,
         // and `make_return_dest` adds the return-place indirect pointer to `llargs`.
-        let return_dest = self.make_return_dest(bx, destination, &fn_abi.ret, &mut llargs, None);
+        let return_dest = self.make_return_dest(bx, destination, &fn_abi.ret, &mut llargs);
         let destination = target.map(|target| (return_dest, target));
 
         // Split the rust-call tupled arguments off.
-        let (first_args, untuple) = if abi == ExternAbi::RustCall
+        let (first_args, untuple) = if sig.abi() == ExternAbi::RustCall
             && let Some((tup, args)) = args.split_last()
         {
             (args, Some(tup))
@@ -1055,7 +1061,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         'make_args: for (i, arg) in first_args.iter().enumerate() {
             let mut op = self.codegen_operand(bx, &arg.node);
 
-            if let (0, Some(ty::InstanceKind::Virtual(_, idx))) = (i, def) {
+            if let (0, Some(ty::InstanceKind::Virtual(_, idx))) = (i, instance.map(|i| i.def)) {
                 match op.val {
                     Pair(data_ptr, meta) => {
                         // In the case of Rc<Self>, we need to explicitly pass a
@@ -1109,7 +1115,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         // Make sure that we've actually unwrapped the rcvr down
                         // to a pointer or ref to `dyn* Trait`.
                         if !op.layout.ty.builtin_deref(true).unwrap().is_dyn_star() {
-                            span_bug!(span, "can't codegen a virtual call on {:#?}", op);
+                            span_bug!(fn_span, "can't codegen a virtual call on {:#?}", op);
                         }
                         let place = op.deref(bx.cx());
                         let data_place = place.project_field(bx, 0);
@@ -1125,7 +1131,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         continue;
                     }
                     _ => {
-                        span_bug!(span, "can't codegen a virtual call on {:#?}", op);
+                        span_bug!(fn_span, "can't codegen a virtual call on {:#?}", op);
                     }
                 }
             }
@@ -1175,8 +1181,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 mir_args + 1,
                 "#[track_caller] fn's must have 1 more argument in their ABI than in their MIR: {instance:?} {fn_span:?} {fn_abi:?}",
             );
-            let location =
-                self.get_caller_location(bx, mir::SourceInfo { span: fn_span, ..source_info });
+            let location = self.get_caller_location(bx, source_info);
             debug!(
                 "codegen_call_terminator({:?}): location={:?} (fn_span {:?})",
                 terminator, location, fn_span
@@ -1195,9 +1200,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         let fn_ptr = match (instance, llfn) {
             (Some(instance), None) => bx.get_fn_addr(instance),
             (_, Some(llfn)) => llfn,
-            _ => span_bug!(span, "no instance or llfn for call"),
+            _ => span_bug!(fn_span, "no instance or llfn for call"),
         };
-        self.set_debug_loc(bx, mir::SourceInfo { span: fn_span, ..source_info });
+        self.set_debug_loc(bx, source_info);
         helper.do_call(
             self,
             bx,
@@ -1667,7 +1672,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         tuple.layout.fields.count()
     }
 
-    fn get_caller_location(
+    pub(super) fn get_caller_location(
         &mut self,
         bx: &mut Bx,
         source_info: mir::SourceInfo,
@@ -1868,7 +1873,6 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         dest: mir::Place<'tcx>,
         fn_ret: &ArgAbi<'tcx, Ty<'tcx>>,
         llargs: &mut Vec<Bx::Value>,
-        intrinsic: Option<ty::IntrinsicDef>,
     ) -> ReturnDest<'tcx, Bx::Value> {
         // If the return is ignored, we can just return a do-nothing `ReturnDest`.
         if fn_ret.is_ignore() {
@@ -1888,13 +1892,6 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         tmp.storage_live(bx);
                         llargs.push(tmp.val.llval);
                         ReturnDest::IndirectOperand(tmp, index)
-                    } else if intrinsic.is_some() {
-                        // Currently, intrinsics always need a location to store
-                        // the result, so we create a temporary `alloca` for the
-                        // result.
-                        let tmp = PlaceRef::alloca(bx, fn_ret.layout);
-                        tmp.storage_live(bx);
-                        ReturnDest::IndirectOperand(tmp, index)
                     } else {
                         ReturnDest::DirectOperand(index)
                     };
@@ -1904,7 +1901,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 }
             }
         } else {
-            self.codegen_place(bx, mir::PlaceRef { local: dest.local, projection: dest.projection })
+            self.codegen_place(bx, dest.as_ref())
         };
         if fn_ret.is_indirect() {
             if dest.val.align < dest.layout.align.abi {
diff --git a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs
index b0fcfee2adf..a6d159c51e1 100644
--- a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs
@@ -1,9 +1,9 @@
 use rustc_abi::WrappingRange;
+use rustc_middle::mir::SourceInfo;
 use rustc_middle::ty::{self, Ty, TyCtxt};
 use rustc_middle::{bug, span_bug};
 use rustc_session::config::OptLevel;
-use rustc_span::{Span, sym};
-use rustc_target::callconv::{FnAbi, PassMode};
+use rustc_span::sym;
 
 use super::FunctionCx;
 use super::operand::OperandRef;
@@ -52,13 +52,14 @@ fn memset_intrinsic<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
     /// In the `Err` case, returns the instance that should be called instead.
     pub fn codegen_intrinsic_call(
+        &mut self,
         bx: &mut Bx,
         instance: ty::Instance<'tcx>,
-        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         args: &[OperandRef<'tcx, Bx::Value>],
-        llresult: Bx::Value,
-        span: Span,
+        result: PlaceRef<'tcx, Bx::Value>,
+        source_info: SourceInfo,
     ) -> Result<(), ty::Instance<'tcx>> {
+        let span = source_info.span;
         let callee_ty = instance.ty(bx.tcx(), bx.typing_env());
 
         let ty::FnDef(def_id, fn_args) = *callee_ty.kind() else {
@@ -97,7 +98,6 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         }
 
         let llret_ty = bx.backend_type(bx.layout_of(ret_ty));
-        let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 
         let llval = match name {
             sym::abort => {
@@ -105,6 +105,12 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 return Ok(());
             }
 
+            sym::caller_location => {
+                let location = self.get_caller_location(bx, source_info);
+                location.val.store(bx, result);
+                return Ok(());
+            }
+
             sym::va_start => bx.va_start(args[0].immediate()),
             sym::va_end => bx.va_end(args[0].immediate()),
             sym::size_of_val => {
@@ -528,18 +534,16 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
 
             _ => {
                 // Need to use backend-specific things in the implementation.
-                return bx.codegen_intrinsic_call(instance, fn_abi, args, llresult, span);
+                return bx.codegen_intrinsic_call(instance, args, result, span);
             }
         };
 
-        if !fn_abi.ret.is_ignore() {
-            if let PassMode::Cast { .. } = &fn_abi.ret.mode {
-                bx.store_to_place(llval, result.val);
-            } else {
-                OperandRef::from_immediate_or_packed_pair(bx, llval, result.layout)
-                    .val
-                    .store(bx, result);
-            }
+        if result.layout.ty.is_bool() {
+            OperandRef::from_immediate_or_packed_pair(bx, llval, result.layout)
+                .val
+                .store(bx, result);
+        } else if !result.layout.ty.is_unit() {
+            bx.store_to_place(llval, result.val);
         }
         Ok(())
     }
diff --git a/compiler/rustc_codegen_ssa/src/traits/intrinsic.rs b/compiler/rustc_codegen_ssa/src/traits/intrinsic.rs
index 88cf8dbf0c5..a07c569a032 100644
--- a/compiler/rustc_codegen_ssa/src/traits/intrinsic.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/intrinsic.rs
@@ -1,9 +1,9 @@
-use rustc_middle::ty::{self, Ty};
+use rustc_middle::ty;
 use rustc_span::Span;
-use rustc_target::callconv::FnAbi;
 
 use super::BackendTypes;
 use crate::mir::operand::OperandRef;
+use crate::mir::place::PlaceRef;
 
 pub trait IntrinsicCallBuilderMethods<'tcx>: BackendTypes {
     /// Remember to add all intrinsics here, in `compiler/rustc_hir_analysis/src/check/mod.rs`,
@@ -14,9 +14,8 @@ pub trait IntrinsicCallBuilderMethods<'tcx>: BackendTypes {
     fn codegen_intrinsic_call(
         &mut self,
         instance: ty::Instance<'tcx>,
-        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
         args: &[OperandRef<'tcx, Self::Value>],
-        llresult: Self::Value,
+        result: PlaceRef<'tcx, Self::Value>,
         span: Span,
     ) -> Result<(), ty::Instance<'tcx>>;
 
diff --git a/compiler/rustc_expand/src/expand.rs b/compiler/rustc_expand/src/expand.rs
index 81d4d59ee04..e5749ba96a6 100644
--- a/compiler/rustc_expand/src/expand.rs
+++ b/compiler/rustc_expand/src/expand.rs
@@ -1319,10 +1319,10 @@ impl InvocationCollectorNode for P<ast::Item> {
 
             let mut idents = Vec::new();
             collect_use_tree_leaves(ut, &mut idents);
-            return idents;
+            idents
+        } else {
+            self.kind.ident().into_iter().collect()
         }
-
-        if let Some(ident) = self.kind.ident() { vec![ident] } else { vec![] }
     }
 }
 
diff --git a/compiler/rustc_mir_build/src/builder/scope.rs b/compiler/rustc_mir_build/src/builder/scope.rs
index 2a30777e98c..67988f1fcbc 100644
--- a/compiler/rustc_mir_build/src/builder/scope.rs
+++ b/compiler/rustc_mir_build/src/builder/scope.rs
@@ -209,7 +209,7 @@ const ROOT_NODE: DropIdx = DropIdx::ZERO;
 #[derive(Debug)]
 struct DropTree {
     /// Nodes in the drop tree, containing drop data and a link to the next node.
-    drops: IndexVec<DropIdx, DropNode>,
+    drop_nodes: IndexVec<DropIdx, DropNode>,
     /// Map for finding the index of an existing node, given its contents.
     existing_drops_map: FxHashMap<DropNodeKey, DropIdx>,
     /// Edges into the `DropTree` that need to be added once it's lowered.
@@ -230,7 +230,6 @@ struct DropNode {
 struct DropNodeKey {
     next: DropIdx,
     local: Local,
-    kind: DropKind,
 }
 
 impl Scope {
@@ -278,8 +277,8 @@ impl DropTree {
         let fake_source_info = SourceInfo::outermost(DUMMY_SP);
         let fake_data =
             DropData { source_info: fake_source_info, local: Local::MAX, kind: DropKind::Storage };
-        let drops = IndexVec::from_raw(vec![DropNode { data: fake_data, next: DropIdx::MAX }]);
-        Self { drops, entry_points: Vec::new(), existing_drops_map: FxHashMap::default() }
+        let drop_nodes = IndexVec::from_raw(vec![DropNode { data: fake_data, next: DropIdx::MAX }]);
+        Self { drop_nodes, entry_points: Vec::new(), existing_drops_map: FxHashMap::default() }
     }
 
     /// Adds a node to the drop tree, consisting of drop data and the index of
@@ -288,12 +287,12 @@ impl DropTree {
     /// If there is already an equivalent node in the tree, nothing is added, and
     /// that node's index is returned. Otherwise, the new node's index is returned.
     fn add_drop(&mut self, data: DropData, next: DropIdx) -> DropIdx {
-        let drops = &mut self.drops;
+        let drop_nodes = &mut self.drop_nodes;
         *self
             .existing_drops_map
-            .entry(DropNodeKey { next, local: data.local, kind: data.kind })
+            .entry(DropNodeKey { next, local: data.local })
             // Create a new node, and also add its index to the map.
-            .or_insert_with(|| drops.push(DropNode { data, next }))
+            .or_insert_with(|| drop_nodes.push(DropNode { data, next }))
     }
 
     /// Registers `from` as an entry point to this drop tree, at `to`.
@@ -301,7 +300,7 @@ impl DropTree {
     /// During [`Self::build_mir`], `from` will be linked to the corresponding
     /// block within the drop tree.
     fn add_entry_point(&mut self, from: BasicBlock, to: DropIdx) {
-        debug_assert!(to < self.drops.next_index());
+        debug_assert!(to < self.drop_nodes.next_index());
         self.entry_points.push((to, from));
     }
 
@@ -341,10 +340,10 @@ impl DropTree {
             Own,
         }
 
-        let mut blocks = IndexVec::from_elem(None, &self.drops);
+        let mut blocks = IndexVec::from_elem(None, &self.drop_nodes);
         blocks[ROOT_NODE] = root_node;
 
-        let mut needs_block = IndexVec::from_elem(Block::None, &self.drops);
+        let mut needs_block = IndexVec::from_elem(Block::None, &self.drop_nodes);
         if root_node.is_some() {
             // In some cases (such as drops for `continue`) the root node
             // already has a block. In this case, make sure that we don't
@@ -356,7 +355,7 @@ impl DropTree {
         let entry_points = &mut self.entry_points;
         entry_points.sort();
 
-        for (drop_idx, drop_node) in self.drops.iter_enumerated().rev() {
+        for (drop_idx, drop_node) in self.drop_nodes.iter_enumerated().rev() {
             if entry_points.last().is_some_and(|entry_point| entry_point.0 == drop_idx) {
                 let block = *blocks[drop_idx].get_or_insert_with(|| T::make_block(cfg));
                 needs_block[drop_idx] = Block::Own;
@@ -396,7 +395,7 @@ impl DropTree {
         cfg: &mut CFG<'tcx>,
         blocks: &IndexSlice<DropIdx, Option<BasicBlock>>,
     ) {
-        for (drop_idx, drop_node) in self.drops.iter_enumerated().rev() {
+        for (drop_idx, drop_node) in self.drop_nodes.iter_enumerated().rev() {
             let Some(block) = blocks[drop_idx] else { continue };
             match drop_node.data.kind {
                 DropKind::Value => {
@@ -726,11 +725,12 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
             drops
         };
 
-        let drop_idx = self.scopes.scopes[scope_index + 1..]
-            .iter()
-            .flat_map(|scope| &scope.drops)
-            .fold(ROOT_NODE, |drop_idx, &drop| drops.add_drop(drop, drop_idx));
-
+        let mut drop_idx = ROOT_NODE;
+        for scope in &self.scopes.scopes[scope_index + 1..] {
+            for drop in &scope.drops {
+                drop_idx = drops.add_drop(*drop, drop_idx);
+            }
+        }
         drops.add_entry_point(block, drop_idx);
 
         // `build_drop_trees` doesn't have access to our source_info, so we
@@ -829,9 +829,15 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                         // `unwind_to` should drop the value that we're about to
                         // schedule. If dropping this value panics, then we continue
                         // with the *next* value on the unwind path.
-                        debug_assert_eq!(unwind_drops.drops[unwind_to].data.local, drop_data.local);
-                        debug_assert_eq!(unwind_drops.drops[unwind_to].data.kind, drop_data.kind);
-                        unwind_to = unwind_drops.drops[unwind_to].next;
+                        debug_assert_eq!(
+                            unwind_drops.drop_nodes[unwind_to].data.local,
+                            drop_data.local
+                        );
+                        debug_assert_eq!(
+                            unwind_drops.drop_nodes[unwind_to].data.kind,
+                            drop_data.kind
+                        );
+                        unwind_to = unwind_drops.drop_nodes[unwind_to].next;
 
                         let mut unwind_entry_point = unwind_to;
 
@@ -1551,14 +1557,14 @@ where
                 //
                 // We adjust this BEFORE we create the drop (e.g., `drops[n]`)
                 // because `drops[n]` should unwind to `drops[n-1]`.
-                debug_assert_eq!(unwind_drops.drops[unwind_to].data.local, drop_data.local);
-                debug_assert_eq!(unwind_drops.drops[unwind_to].data.kind, drop_data.kind);
-                unwind_to = unwind_drops.drops[unwind_to].next;
+                debug_assert_eq!(unwind_drops.drop_nodes[unwind_to].data.local, drop_data.local);
+                debug_assert_eq!(unwind_drops.drop_nodes[unwind_to].data.kind, drop_data.kind);
+                unwind_to = unwind_drops.drop_nodes[unwind_to].next;
 
                 if let Some(idx) = dropline_to {
-                    debug_assert_eq!(coroutine_drops.drops[idx].data.local, drop_data.local);
-                    debug_assert_eq!(coroutine_drops.drops[idx].data.kind, drop_data.kind);
-                    dropline_to = Some(coroutine_drops.drops[idx].next);
+                    debug_assert_eq!(coroutine_drops.drop_nodes[idx].data.local, drop_data.local);
+                    debug_assert_eq!(coroutine_drops.drop_nodes[idx].data.kind, drop_data.kind);
+                    dropline_to = Some(coroutine_drops.drop_nodes[idx].next);
                 }
 
                 // If the operand has been moved, and we are not on an unwind
@@ -1598,9 +1604,12 @@ where
                 // cases we emit things ALSO on the unwind path, so we need to adjust
                 // `unwind_to` in that case.
                 if storage_dead_on_unwind {
-                    debug_assert_eq!(unwind_drops.drops[unwind_to].data.local, drop_data.local);
-                    debug_assert_eq!(unwind_drops.drops[unwind_to].data.kind, drop_data.kind);
-                    unwind_to = unwind_drops.drops[unwind_to].next;
+                    debug_assert_eq!(
+                        unwind_drops.drop_nodes[unwind_to].data.local,
+                        drop_data.local
+                    );
+                    debug_assert_eq!(unwind_drops.drop_nodes[unwind_to].data.kind, drop_data.kind);
+                    unwind_to = unwind_drops.drop_nodes[unwind_to].next;
                 }
 
                 // If the operand has been moved, and we are not on an unwind
@@ -1629,14 +1638,17 @@ where
                 // the storage-dead has completed, we need to adjust the `unwind_to` pointer
                 // so that any future drops we emit will not register storage-dead.
                 if storage_dead_on_unwind {
-                    debug_assert_eq!(unwind_drops.drops[unwind_to].data.local, drop_data.local);
-                    debug_assert_eq!(unwind_drops.drops[unwind_to].data.kind, drop_data.kind);
-                    unwind_to = unwind_drops.drops[unwind_to].next;
+                    debug_assert_eq!(
+                        unwind_drops.drop_nodes[unwind_to].data.local,
+                        drop_data.local
+                    );
+                    debug_assert_eq!(unwind_drops.drop_nodes[unwind_to].data.kind, drop_data.kind);
+                    unwind_to = unwind_drops.drop_nodes[unwind_to].next;
                 }
                 if let Some(idx) = dropline_to {
-                    debug_assert_eq!(coroutine_drops.drops[idx].data.local, drop_data.local);
-                    debug_assert_eq!(coroutine_drops.drops[idx].data.kind, drop_data.kind);
-                    dropline_to = Some(coroutine_drops.drops[idx].next);
+                    debug_assert_eq!(coroutine_drops.drop_nodes[idx].data.local, drop_data.local);
+                    debug_assert_eq!(coroutine_drops.drop_nodes[idx].data.kind, drop_data.kind);
+                    dropline_to = Some(coroutine_drops.drop_nodes[idx].next);
                 }
                 // Only temps and vars need their storage dead.
                 assert!(local.index() > arg_count);
@@ -1663,10 +1675,10 @@ impl<'a, 'tcx: 'a> Builder<'a, 'tcx> {
         let is_coroutine = self.coroutine.is_some();
 
         // Link the exit drop tree to unwind drop tree.
-        if drops.drops.iter().any(|drop_node| drop_node.data.kind == DropKind::Value) {
+        if drops.drop_nodes.iter().any(|drop_node| drop_node.data.kind == DropKind::Value) {
             let unwind_target = self.diverge_cleanup_target(else_scope, span);
             let mut unwind_indices = IndexVec::from_elem_n(unwind_target, 1);
-            for (drop_idx, drop_node) in drops.drops.iter_enumerated().skip(1) {
+            for (drop_idx, drop_node) in drops.drop_nodes.iter_enumerated().skip(1) {
                 match drop_node.data.kind {
                     DropKind::Storage | DropKind::ForLint => {
                         if is_coroutine {
@@ -1695,35 +1707,29 @@ impl<'a, 'tcx: 'a> Builder<'a, 'tcx> {
         }
         // Link the exit drop tree to dropline drop tree (coroutine drop path) for async drops
         if is_coroutine
-            && drops.drops.iter().any(|DropNode { data, next: _ }| {
+            && drops.drop_nodes.iter().any(|DropNode { data, next: _ }| {
                 data.kind == DropKind::Value && self.is_async_drop(data.local)
             })
         {
             let dropline_target = self.diverge_dropline_target(else_scope, span);
             let mut dropline_indices = IndexVec::from_elem_n(dropline_target, 1);
-            for (drop_idx, drop_data) in drops.drops.iter_enumerated().skip(1) {
+            for (drop_idx, drop_data) in drops.drop_nodes.iter_enumerated().skip(1) {
+                let coroutine_drop = self
+                    .scopes
+                    .coroutine_drops
+                    .add_drop(drop_data.data, dropline_indices[drop_data.next]);
                 match drop_data.data.kind {
-                    DropKind::Storage | DropKind::ForLint => {
-                        let coroutine_drop = self
-                            .scopes
-                            .coroutine_drops
-                            .add_drop(drop_data.data, dropline_indices[drop_data.next]);
-                        dropline_indices.push(coroutine_drop);
-                    }
+                    DropKind::Storage | DropKind::ForLint => {}
                     DropKind::Value => {
-                        let coroutine_drop = self
-                            .scopes
-                            .coroutine_drops
-                            .add_drop(drop_data.data, dropline_indices[drop_data.next]);
                         if self.is_async_drop(drop_data.data.local) {
                             self.scopes.coroutine_drops.add_entry_point(
                                 blocks[drop_idx].unwrap(),
                                 dropline_indices[drop_data.next],
                             );
                         }
-                        dropline_indices.push(coroutine_drop);
                     }
                 }
+                dropline_indices.push(coroutine_drop);
             }
         }
         blocks[ROOT_NODE].map(BasicBlock::unit)
@@ -1769,11 +1775,11 @@ impl<'a, 'tcx: 'a> Builder<'a, 'tcx> {
         // prevent drop elaboration from creating drop flags that would have
         // to be captured by the coroutine. I'm not sure how important this
         // optimization is, but it is here.
-        for (drop_idx, drop_node) in drops.drops.iter_enumerated() {
+        for (drop_idx, drop_node) in drops.drop_nodes.iter_enumerated() {
             if let DropKind::Value = drop_node.data.kind
                 && let Some(bb) = blocks[drop_idx]
             {
-                debug_assert!(drop_node.next < drops.drops.next_index());
+                debug_assert!(drop_node.next < drops.drop_nodes.next_index());
                 drops.entry_points.push((drop_node.next, bb));
             }
         }
diff --git a/compiler/rustc_parse/src/parser/stmt.rs b/compiler/rustc_parse/src/parser/stmt.rs
index 396ded96bde..ccc3410674b 100644
--- a/compiler/rustc_parse/src/parser/stmt.rs
+++ b/compiler/rustc_parse/src/parser/stmt.rs
@@ -515,8 +515,8 @@ impl<'a> Parser<'a> {
     fn error_block_no_opening_brace_msg(&mut self, msg: Cow<'static, str>) -> Diag<'a> {
         let prev = self.prev_token.span;
         let sp = self.token.span;
-        let mut e = self.dcx().struct_span_err(sp, msg);
-        self.label_expected_raw_ref(&mut e);
+        let mut err = self.dcx().struct_span_err(sp, msg);
+        self.label_expected_raw_ref(&mut err);
 
         let do_not_suggest_help = self.token.is_keyword(kw::In)
             || self.token == token::Colon
@@ -558,20 +558,19 @@ impl<'a> Parser<'a> {
                     stmt.span
                 };
                 self.suggest_fixes_misparsed_for_loop_head(
-                    &mut e,
+                    &mut err,
                     prev.between(sp),
                     stmt_span,
                     &stmt.kind,
                 );
             }
             Err(e) => {
-                self.recover_stmt_(SemiColonMode::Break, BlockMode::Ignore);
-                e.cancel();
+                e.delay_as_bug();
             }
             _ => {}
         }
-        e.span_label(sp, "expected `{`");
-        e
+        err.span_label(sp, "expected `{`");
+        err
     }
 
     fn suggest_fixes_misparsed_for_loop_head(
diff --git a/compiler/rustc_passes/src/check_attr.rs b/compiler/rustc_passes/src/check_attr.rs
index 1d024694049..df8b5a6b181 100644
--- a/compiler/rustc_passes/src/check_attr.rs
+++ b/compiler/rustc_passes/src/check_attr.rs
@@ -255,7 +255,7 @@ impl<'tcx> CheckAttrVisitor<'tcx> {
                             self.check_generic_attr(hir_id, attr, target, Target::Fn);
                             self.check_proc_macro(hir_id, target, ProcMacroKind::Derive)
                         }
-                        [sym::autodiff, ..] => {
+                        [sym::autodiff_forward, ..] | [sym::autodiff_reverse, ..] => {
                             self.check_autodiff(hir_id, attr, span, target)
                         }
                         [sym::coroutine, ..] => {
diff --git a/compiler/rustc_resolve/src/diagnostics.rs b/compiler/rustc_resolve/src/diagnostics.rs
index d09750fa281..201b1c0a493 100644
--- a/compiler/rustc_resolve/src/diagnostics.rs
+++ b/compiler/rustc_resolve/src/diagnostics.rs
@@ -6,7 +6,7 @@ use rustc_ast::{
 };
 use rustc_ast_pretty::pprust;
 use rustc_attr_data_structures::{self as attr, Stability};
-use rustc_data_structures::fx::FxHashSet;
+use rustc_data_structures::fx::{FxHashMap, FxHashSet};
 use rustc_data_structures::unord::UnordSet;
 use rustc_errors::codes::*;
 use rustc_errors::{
@@ -2623,7 +2623,53 @@ impl<'ra, 'tcx> Resolver<'ra, 'tcx> {
         };
 
         for &StrippedCfgItem { parent_module, ident, ref cfg } in symbols {
-            if parent_module != module || ident.name != *segment {
+            if ident.name != *segment {
+                continue;
+            }
+
+            fn comes_from_same_module_for_glob(
+                r: &Resolver<'_, '_>,
+                parent_module: DefId,
+                module: DefId,
+                visited: &mut FxHashMap<DefId, bool>,
+            ) -> bool {
+                if let Some(&cached) = visited.get(&parent_module) {
+                    // this branch is prevent from being called recursively infinity,
+                    // because there has some cycles in globs imports,
+                    // see more spec case at `tests/ui/cfg/diagnostics-reexport-2.rs#reexport32`
+                    return cached;
+                }
+                visited.insert(parent_module, false);
+                let res = r.module_map.get(&parent_module).is_some_and(|m| {
+                    for importer in m.glob_importers.borrow().iter() {
+                        if let Some(next_parent_module) = importer.parent_scope.module.opt_def_id()
+                        {
+                            if next_parent_module == module
+                                || comes_from_same_module_for_glob(
+                                    r,
+                                    next_parent_module,
+                                    module,
+                                    visited,
+                                )
+                            {
+                                return true;
+                            }
+                        }
+                    }
+                    false
+                });
+                visited.insert(parent_module, res);
+                res
+            }
+
+            let comes_from_same_module = parent_module == module
+                || comes_from_same_module_for_glob(
+                    self,
+                    parent_module,
+                    module,
+                    &mut Default::default(),
+                );
+            if !comes_from_same_module {
                 continue;
             }
 
diff --git a/compiler/rustc_span/src/symbol.rs b/compiler/rustc_span/src/symbol.rs
index 0447713fb05..8f7e72f0ae1 100644
--- a/compiler/rustc_span/src/symbol.rs
+++ b/compiler/rustc_span/src/symbol.rs
@@ -244,6 +244,7 @@ symbols! {
         FnMut,
         FnOnce,
         Formatter,
+        Forward,
         From,
         FromIterator,
         FromResidual,
@@ -339,6 +340,7 @@ symbols! {
         Result,
         ResumeTy,
         Return,
+        Reverse,
         Right,
         Rust,
         RustaceansAreAwesome,
@@ -522,7 +524,8 @@ symbols! {
         audit_that,
         augmented_assignments,
         auto_traits,
-        autodiff,
+        autodiff_forward,
+        autodiff_reverse,
         automatically_derived,
         avx,
         avx10_target_feature,