Codegen non-overloaded LLVM intrinsics using their name

Author: sayantn

compiler/rustc_codegen_gcc/src/type_of.rs

@@ -1,6 +1,6 @@
 use std::fmt::Write;
 
-use gccjit::{Struct, Type};
+use gccjit::{RValue, Struct, Type};
 use rustc_abi as abi;
 use rustc_abi::Primitive::*;
 use rustc_abi::{
@@ -373,7 +373,11 @@ impl<'gcc, 'tcx> LayoutTypeCodegenMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         unimplemented!();
     }
 
-    fn fn_decl_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> {
+    fn fn_decl_backend_type(
+        &self,
+        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
+        _fn_ptr: RValue<'gcc>,
+    ) -> Type<'gcc> {
         // FIXME(antoyo): Should we do something with `FnAbiGcc::fn_attributes`?
         let FnAbiGcc { return_type, arguments_type, is_c_variadic, .. } = fn_abi.gcc_type(self);
         self.context.new_function_pointer_type(None, return_type, &arguments_type, is_c_variadic)

compiler/rustc_codegen_llvm/messages.ftl

@@ -9,6 +9,10 @@ codegen_llvm_from_llvm_diag = {$message}
 
 codegen_llvm_from_llvm_optimization_diag = {$filename}:{$line}:{$column} {$pass_name} ({$kind}): {$message}
 
+codegen_llvm_intrinsic_signature_mismatch =
+    intrinsic signature mismatch for `{$name}`: expected signature `{$llvm_fn_ty}`, found `{$rust_fn_ty}`
+
+
 codegen_llvm_load_bitcode = failed to load bitcode of module "{$name}"
 codegen_llvm_load_bitcode_with_llvm_err = failed to load bitcode of module "{$name}": {$llvm_err}
 

compiler/rustc_codegen_llvm/src/abi.rs

@@ -1,4 +1,4 @@
-use std::cmp;
+use std::{cmp, iter};
 
 use libc::c_uint;
 use rustc_abi::{
@@ -307,8 +307,46 @@ impl<'ll, 'tcx> ArgAbiBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
     }
 }
 
+pub(crate) enum FunctionSignature<'ll> {
+    /// This is an LLVM intrinsic, the signature is obtained directly from LLVM, and **may not match the Rust signature**
+    LLVMSignature(llvm::Intrinsic, &'ll Type),
+    /// This is an LLVM intrinsic, but the signature is just the Rust signature.
+    /// FIXME: this should ideally not exist, we should be using the LLVM signature for all LLVM intrinsics
+    RustSignature(llvm::Intrinsic, &'ll Type),
+    /// The name starts with `llvm.`, but can't obtain the intrinsic ID. May be invalid or upgradable
+    MaybeInvalid(&'ll Type),
+    /// Just the Rust signature
+    NotIntrinsic(&'ll Type),
+}
+
+impl<'ll> FunctionSignature<'ll> {
+    pub(crate) fn fn_ty(&self) -> &'ll Type {
+        match self {
+            FunctionSignature::LLVMSignature(_, fn_ty)
+            | FunctionSignature::RustSignature(_, fn_ty)
+            | FunctionSignature::MaybeInvalid(fn_ty)
+            | FunctionSignature::NotIntrinsic(fn_ty) => fn_ty,
+        }
+    }
+
+    pub(crate) fn intrinsic(&self) -> Option<llvm::Intrinsic> {
+        match self {
+            FunctionSignature::RustSignature(intrinsic, _)
+            | FunctionSignature::LLVMSignature(intrinsic, _) => Some(*intrinsic),
+            _ => None,
+        }
+    }
+}
+
 pub(crate) trait FnAbiLlvmExt<'ll, 'tcx> {
-    fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
+    fn llvm_return_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
+    fn llvm_argument_types(&self, cx: &CodegenCx<'ll, 'tcx>) -> Vec<&'ll Type>;
+    fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>, name: &[u8]) -> FunctionSignature<'ll>;
+    /// The normal Rust signature for this
+    fn rust_signature(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
+    /// **If this function is an LLVM intrinsic** checks if the LLVM signature provided matches with this
+    fn verify_intrinsic_signature(&self, cx: &CodegenCx<'ll, 'tcx>, llvm_ty: &'ll Type) -> bool;
+
     fn ptr_to_llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
     fn llvm_cconv(&self, cx: &CodegenCx<'ll, 'tcx>) -> llvm::CallConv;
 
@@ -321,30 +359,39 @@ pub(crate) trait FnAbiLlvmExt<'ll, 'tcx> {
     );
 
     /// Apply attributes to a function call.
-    fn apply_attrs_callsite(&self, bx: &mut Builder<'_, 'll, 'tcx>, callsite: &'ll Value);
+    fn apply_attrs_callsite(
+        &self,
+        bx: &mut Builder<'_, 'll, 'tcx>,
+        callsite: &'ll Value,
+        llfn: &'ll Value,
+    );
 }
 
 impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
-    fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
+    fn llvm_return_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
+        match &self.ret.mode {
+            PassMode::Ignore => cx.type_void(),
+            PassMode::Direct(_) | PassMode::Pair(..) => self.ret.layout.immediate_llvm_type(cx),
+            PassMode::Cast { cast, pad_i32: _ } => cast.llvm_type(cx),
+            PassMode::Indirect { .. } => cx.type_void(),
+        }
+    }
+
+    fn llvm_argument_types(&self, cx: &CodegenCx<'ll, 'tcx>) -> Vec<&'ll Type> {
+        let indirect_return = matches!(self.ret.mode, PassMode::Indirect { .. });
+
         // Ignore "extra" args from the call site for C variadic functions.
         // Only the "fixed" args are part of the LLVM function signature.
         let args =
             if self.c_variadic { &self.args[..self.fixed_count as usize] } else { &self.args };
 
         // This capacity calculation is approximate.
-        let mut llargument_tys = Vec::with_capacity(
-            self.args.len() + if let PassMode::Indirect { .. } = self.ret.mode { 1 } else { 0 },
-        );
+        let mut llargument_tys =
+            Vec::with_capacity(args.len() + if indirect_return { 1 } else { 0 });
 
-        let llreturn_ty = match &self.ret.mode {
-            PassMode::Ignore => cx.type_void(),
-            PassMode::Direct(_) | PassMode::Pair(..) => self.ret.layout.immediate_llvm_type(cx),
-            PassMode::Cast { cast, pad_i32: _ } => cast.llvm_type(cx),
-            PassMode::Indirect { .. } => {
-                llargument_tys.push(cx.type_ptr());
-                cx.type_void()
-            }
-        };
+        if indirect_return {
+            llargument_tys.push(cx.type_ptr());
+        }
 
         for arg in args {
             // Note that the exact number of arguments pushed here is carefully synchronized with
@@ -391,10 +438,56 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
             llargument_tys.push(llarg_ty);
         }
 
+        llargument_tys
+    }
+
+    fn verify_intrinsic_signature(&self, cx: &CodegenCx<'ll, 'tcx>, llvm_fn_ty: &'ll Type) -> bool {
+        let rust_return_ty = self.llvm_return_type(cx);
+        let rust_argument_tys = self.llvm_argument_types(cx);
+
+        let llvm_return_ty = cx.get_return_type(llvm_fn_ty);
+        let llvm_argument_tys = cx.func_params_types(llvm_fn_ty);
+        let llvm_is_variadic = cx.func_is_variadic(llvm_fn_ty);
+
+        if self.c_variadic != llvm_is_variadic || rust_argument_tys.len() != llvm_argument_tys.len()
+        {
+            return false;
+        }
+
+        iter::once((rust_return_ty, llvm_return_ty))
+            .chain(iter::zip(rust_argument_tys, llvm_argument_tys))
+            .all(|(rust_ty, llvm_ty)| rust_ty == llvm_ty)
+    }
+
+    fn rust_signature(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
+        let return_ty = self.llvm_return_type(cx);
+        let argument_tys = self.llvm_argument_types(cx);
+
         if self.c_variadic {
-            cx.type_variadic_func(&llargument_tys, llreturn_ty)
+            cx.type_variadic_func(&argument_tys, return_ty)
         } else {
-            cx.type_func(&llargument_tys, llreturn_ty)
+            cx.type_func(&argument_tys, return_ty)
+        }
+    }
+
+    fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>, name: &[u8]) -> FunctionSignature<'ll> {
+        if name.starts_with(b"llvm.") {
+            if let Some(intrinsic) = llvm::Intrinsic::lookup(name) {
+                if !intrinsic.is_overloaded() {
+                    // FIXME: also do this for overloaded intrinsics
+                    FunctionSignature::LLVMSignature(intrinsic, intrinsic.get_type(cx.llcx, &[]))
+                } else {
+                    FunctionSignature::RustSignature(intrinsic, self.rust_signature(cx))
+                }
+            } else {
+                // it's one of 2 cases,
+                // - either the base name is invalid
+                // - it has been superseded by something else, so the intrinsic was removed entirely
+                // to check for upgrades, we need the `llfn`, so we defer it for now
+                FunctionSignature::MaybeInvalid(self.rust_signature(cx))
+            }
+        } else {
+            FunctionSignature::NotIntrinsic(self.rust_signature(cx))
         }
     }
 
@@ -574,7 +667,17 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
         }
     }
 
-    fn apply_attrs_callsite(&self, bx: &mut Builder<'_, 'll, 'tcx>, callsite: &'ll Value) {
+    fn apply_attrs_callsite(
+        &self,
+        bx: &mut Builder<'_, 'll, 'tcx>,
+        callsite: &'ll Value,
+        llfn: &'ll Value,
+    ) {
+        // Don't apply any attributes to LLVM intrinsics, they will be applied by AutoUpgrade
+        if llvm::get_value_name(llfn).starts_with(b"llvm.") {
+            return;
+        }
+
         let mut func_attrs = SmallVec::<[_; 2]>::new();
         if self.ret.layout.is_uninhabited() {
             func_attrs.push(llvm::AttributeKind::NoReturn.create_attr(bx.cx.llcx));

compiler/rustc_codegen_llvm/src/builder.rs

@@ -442,7 +442,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
             )
         };
         if let Some(fn_abi) = fn_abi {
-            fn_abi.apply_attrs_callsite(self, invoke);
+            fn_abi.apply_attrs_callsite(self, invoke, llfn);
         }
         invoke
     }
@@ -1436,7 +1436,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
 
         if let Some(fn_abi) = fn_abi {
-            fn_abi.apply_attrs_callsite(self, call);
+            fn_abi.apply_attrs_callsite(self, call, llfn);
         }
         call
     }
@@ -1779,7 +1779,7 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
             )
         };
         if let Some(fn_abi) = fn_abi {
-            fn_abi.apply_attrs_callsite(self, callbr);
+            fn_abi.apply_attrs_callsite(self, callbr, llfn);
         }
         callbr
     }

compiler/rustc_codegen_llvm/src/declare.rs

@@ -22,12 +22,12 @@ use rustc_target::callconv::FnAbi;
 use smallvec::SmallVec;
 use tracing::debug;
 
-use crate::abi::FnAbiLlvmExt;
-use crate::attributes;
+use crate::abi::{FnAbiLlvmExt, FunctionSignature};
 use crate::common::AsCCharPtr;
 use crate::context::{CodegenCx, GenericCx, SCx, SimpleCx};
 use crate::llvm::AttributePlace::Function;
 use crate::llvm::{self, FromGeneric, Type, Value, Visibility};
+use crate::{attributes, errors};
 
 /// Declare a function with a SimpleCx.
 ///
@@ -148,17 +148,39 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
     ) -> &'ll Value {
         debug!("declare_rust_fn(name={:?}, fn_abi={:?})", name, fn_abi);
 
-        // Function addresses in Rust are never significant, allowing functions to
-        // be merged.
+        let signature = fn_abi.llvm_type(self, name.as_bytes());
+
+        let span = || instance.map(|instance| self.tcx.def_span(instance.def_id()));
+
+        if let FunctionSignature::LLVMSignature(_, llvm_fn_ty) = signature {
+            // check if the intrinsic signatures match
+            if !fn_abi.verify_intrinsic_signature(self, llvm_fn_ty) {
+                self.tcx.dcx().emit_fatal(errors::IntrinsicSignatureMismatch {
+                name,
+                    llvm_fn_ty: &format!("{llvm_fn_ty:?}"),
+                    rust_fn_ty: &format!("{:?}", fn_abi.rust_signature(self)),
+                    span: span(),
+                });
+            }
+        }
+
+            // Function addresses in Rust are never significant, allowing functions to
+            // be merged.
         let llfn = declare_raw_fn(
-            self,
-            name,
-            fn_abi.llvm_cconv(self),
-            llvm::UnnamedAddr::Global,
-            llvm::Visibility::Default,
-            fn_abi.llvm_type(self),
-        );
-        fn_abi.apply_attrs_llfn(self, llfn, instance);
+                self,
+                name,
+                fn_abi.llvm_cconv(self),
+                llvm::UnnamedAddr::Global,
+                llvm::Visibility::Default,
+                signature.fn_ty(),
+            );
+
+        if signature.intrinsic().is_none() {
+            // Don't apply any attributes to intrinsics, they will be applied by AutoUpgrade
+            fn_abi.apply_attrs_llfn(self, llfn, instance);
+        }
+
+        // todo: check for upgrades, and emit error if not upgradable
 
         if self.tcx.sess.is_sanitizer_cfi_enabled() {
             if let Some(instance) = instance {

compiler/rustc_codegen_llvm/src/errors.rs

@@ -147,3 +147,13 @@ pub(crate) struct FixedX18InvalidArch<'a> {
 #[derive(Diagnostic)]
 #[diag(codegen_llvm_sanitizer_kcfi_arity_requires_llvm_21_0_0)]
 pub(crate) struct SanitizerKcfiArityRequiresLLVM2100;
+
+#[derive(Diagnostic)]
+#[diag(codegen_llvm_intrinsic_signature_mismatch)]
+pub(crate) struct IntrinsicSignatureMismatch<'a> {
+    pub name: &'a str,
+    pub llvm_fn_ty: &'a str,
+    pub rust_fn_ty: &'a str,
+    #[primary_span]
+    pub span: Option<Span>,
+}

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -1073,7 +1073,7 @@ fn gen_fn<'a, 'll, 'tcx>(
     codegen: &mut dyn FnMut(Builder<'a, 'll, 'tcx>),
 ) -> (&'ll Type, &'ll Value) {
     let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
-    let llty = fn_abi.llvm_type(cx);
+    let llty = fn_abi.llvm_type(cx, name.as_bytes()).fn_ty();
     let llfn = cx.declare_fn(name, fn_abi, None);
     cx.set_frame_pointer_type(llfn);
     cx.apply_target_cpu_attr(llfn);

compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs

@@ -73,7 +73,6 @@ unsafe extern "C" {
     pub(crate) fn LLVMDumpModule(M: &Module);
     pub(crate) fn LLVMDumpValue(V: &Value);
     pub(crate) fn LLVMGetFunctionCallConv(F: &Value) -> c_uint;
-    pub(crate) fn LLVMGetReturnType(T: &Type) -> &Type;
     pub(crate) fn LLVMGetParams(Fnc: &Value, params: *mut &Value);
     pub(crate) fn LLVMGetNamedFunction(M: &Module, Name: *const c_char) -> Option<&Value>;
 }

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -983,6 +983,8 @@ unsafe extern "C" {
     ) -> &'a Type;
     pub(crate) fn LLVMCountParamTypes(FunctionTy: &Type) -> c_uint;
     pub(crate) fn LLVMGetParamTypes<'a>(FunctionTy: &'a Type, Dest: *mut &'a Type);
+    pub(crate) fn LLVMGetReturnType(FunctionTy: &Type) -> &Type;
+    pub(crate) fn LLVMIsFunctionVarArg(FunctionTy: &Type) -> Bool;
 
     // Operations on struct types
     pub(crate) fn LLVMStructTypeInContext<'a>(
@@ -1131,6 +1133,13 @@ unsafe extern "C" {
 
     // Operations about llvm intrinsics
     pub(crate) fn LLVMLookupIntrinsicID(Name: *const c_char, NameLen: size_t) -> c_uint;
+    pub(crate) fn LLVMIntrinsicIsOverloaded(ID: NonZero<c_uint>) -> Bool;
+    pub(crate) fn LLVMIntrinsicGetType<'a>(
+        C: &'a Context,
+        ID: NonZero<c_uint>,
+        ParamTypes: *const &'a Type,
+        ParamCount: size_t,
+    ) -> &'a Type;
     pub(crate) fn LLVMGetIntrinsicDeclaration<'a>(
         Mod: &'a Module,
         ID: NonZero<c_uint>,

compiler/rustc_codegen_llvm/src/llvm/mod.rs

@@ -311,6 +311,14 @@ impl Intrinsic {
         NonZero::new(id).map(|id| Self { id })
     }
 
+    pub(crate) fn is_overloaded(self) -> bool {
+        unsafe { LLVMIntrinsicIsOverloaded(self.id).is_true() }
+    }
+
+    pub(crate) fn get_type<'ll>(self, llcx: &'ll Context, type_params: &[&'ll Type]) -> &'ll Type {
+        unsafe { LLVMIntrinsicGetType(llcx, self.id, type_params.as_ptr(), type_params.len()) }
+    }
+
     pub(crate) fn get_declaration<'ll>(
         self,
         llmod: &'ll Module,