rustc_codegen_llvm/

builder.rs

use std::borrow::{Borrow, Cow};
use std::ops::Deref;
use std::{iter, ptr};

use rustc_ast::expand::typetree::FncTree;
pub(crate) mod autodiff;
pub(crate) mod gpu_offload;

use libc::{c_char, c_uint};
use rustc_abi::{self as abi, Align, CanonAbi, Float, Size, WrappingRange};
use rustc_attr_ir::UnrollAttr;
use rustc_codegen_ssa::MemFlags;
use rustc_codegen_ssa::common::{IntPredicate, RealPredicate, SynchronizationScope};
use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
use rustc_codegen_ssa::mir::place::PlaceRef;
use rustc_codegen_ssa::traits::*;
use rustc_data_structures::small_c_str::SmallCStr;
use rustc_hir::def_id::DefId;
use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
use rustc_middle::ty::layout::{
    FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasTypingEnv, LayoutError, LayoutOfHelpers,
    TyAndLayout,
};
use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
use rustc_sanitizers::{cfi, kcfi};
use rustc_session::config::OptLevel;
use rustc_span::{Span, bug};
use rustc_target::callconv::{FnAbi, PassMode};
use rustc_target::spec::{Arch, HasTargetSpec, SanitizerSet, Target};
use smallvec::SmallVec;
use tracing::{debug, instrument};

use crate::abi::FnAbiLlvmExt;
use crate::attributes;
use crate::common::Funclet;
use crate::context::{CodegenCx, FullCx, GenericCx, SCx};
use crate::llvm::{
    self, AtomicOrdering, AtomicRmwBinOp, BasicBlock, FromGeneric, GEPNoWrapFlags, Metadata, TRUE,
    ToLlvmBool, Type, TypeKind, Value,
};
use crate::type_of::LayoutLlvmExt;

#[must_use]
pub(crate) struct GenericBuilder<'a, 'll, CX: Borrow<SCx<'ll>>> {
    /// Owning pointer that corresponds to an `LLVMBuilderRef`. Disposed on drop.
    llbuilder_raw: ptr::NonNull<llvm::Builder<'ll>>,
    pub cx: &'a GenericCx<'ll, CX>,
    pub span: rustc_span::Span,
}

pub(crate) type SBuilder<'a, 'll> = GenericBuilder<'a, 'll, SCx<'ll>>;
pub(crate) type Builder<'a, 'll, 'tcx> = GenericBuilder<'a, 'll, FullCx<'ll, 'tcx>>;

impl<'a, 'll, CX: Borrow<SCx<'ll>>> Drop for GenericBuilder<'a, 'll, CX> {
    fn drop(&mut self) {
        unsafe { llvm::LLVMDisposeBuilder(self.llbuilder_raw) };
    }
}

impl<'a, 'll> SBuilder<'a, 'll> {
    pub(crate) fn call(
        &mut self,
        llty: Type<'ll>,
        llfn: Value<'ll>,
        args: &[Value<'ll>],
        funclet: Option<&Funclet<'ll>>,
    ) -> Value<'ll> {
        debug!("call {:?} with args ({:?})", llfn, args);

        let args = self.check_call("call", llty, llfn, args);
        let funclet_bundle = funclet.map(|funclet| funclet.bundle());
        let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
        if let Some(funclet_bundle) = funclet_bundle {
            bundles.push(funclet_bundle);
        }

        unsafe {
            llvm::LLVMBuildCallWithOperandBundles(
                self.llbuilder(),
                llty,
                llfn,
                args.as_ptr(),
                args.len() as c_uint,
                bundles.as_ptr(),
                bundles.len() as c_uint,
                c"".as_ptr(),
            )
        }
    }
}

impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
    fn with_cx(scx: &'a GenericCx<'ll, CX>) -> Self {
        // Create a fresh builder from the simple context.
        let llbuilder_raw = unsafe { llvm::LLVMCreateBuilderForModule(scx.deref().borrow().llmod) };
        GenericBuilder { llbuilder_raw, cx: scx, span: rustc_span::DUMMY_SP }
    }

    pub(crate) fn append_block(
        cx: &'a GenericCx<'ll, CX>,
        llfn: Value<'ll>,
        name: &str,
    ) -> &'ll BasicBlock {
        unsafe {
            let name = SmallCStr::new(name);
            llvm::LLVMAppendBasicBlockInContext(cx.llcx(), llfn, name.as_ptr())
        }
    }

    /// Returns a reference to this builder's underlying [`llvm::Builder`],
    /// tied to the lifetime of self.
    pub(crate) fn llbuilder<'s>(&'s self) -> &'s llvm::Builder<'ll> {
        // SAFETY: We own this builder, and the reference lifetime is tied to `&self`
        // (not to 'll), so it won't be disposed by our Drop while the ref is live.
        unsafe { self.llbuilder_raw.as_ref() }
    }

    pub(crate) fn trunc(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildTrunc(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    pub(crate) fn bitcast(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildBitCast(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    pub(crate) fn ret_void(&mut self) {
        llvm::LLVMBuildRetVoid(self.llbuilder());
    }

    pub(crate) fn ret(&mut self, v: Value<'ll>) {
        unsafe {
            llvm::LLVMBuildRet(self.llbuilder(), v);
        }
    }

    pub(crate) fn build(cx: &'a GenericCx<'ll, CX>, llbb: &'ll BasicBlock) -> Self {
        let bx = Self::with_cx(cx);
        unsafe {
            llvm::LLVMPositionBuilderAtEnd(bx.llbuilder(), llbb);
        }
        bx
    }

    // The generic builder has less functionality and thus (unlike the other alloca) we can not
    // easily jump to the beginning of the function to place our allocas there. We trust the user
    // to manually do that. FIXME(offload): improve the genericCx and add more llvm wrappers to
    // handle this.
    pub(crate) fn direct_alloca(&mut self, size: Size, align: Align, name: &str) -> Value<'ll> {
        let ty = self.cx.type_array(self.cx.type_i8(), size.bytes());
        let val = unsafe {
            let alloca = llvm::LLVMBuildAlloca(self.llbuilder(), ty, UNNAMED);
            llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
            // Cast to default addrspace if necessary
            llvm::LLVMBuildPointerCast(self.llbuilder(), alloca, self.cx.type_ptr(), UNNAMED)
        };
        if !name.is_empty() {
            llvm::set_value_name(val, name.as_bytes());
        }
        val
    }

    fn build_gep(
        &mut self,
        stride: Size,
        ptr: Value<'ll>,
        index: Value<'ll>,
        flags: GEPNoWrapFlags,
    ) -> Value<'ll> {
        let ty = if stride.bytes() == 1 {
            // LLVM prefers `i8` as the canonical form for a single-byte GEP.
            self.cx.type_i8()
        } else {
            self.cx.type_array(self.cx.type_i8(), stride.bytes())
        };
        let indices = &[index];
        unsafe {
            llvm::LLVMBuildGEPWithNoWrapFlags(
                self.llbuilder(),
                ty,
                ptr,
                indices.as_ptr(),
                indices.len() as c_uint,
                UNNAMED,
                flags,
            )
        }
    }

    pub(crate) fn store(&mut self, val: Value<'ll>, ptr: Value<'ll>, align: Align) -> Value<'ll> {
        debug!("Store {:?} -> {:?}", val, ptr);
        assert_eq!(self.cx.type_kind(self.cx.val_ty(ptr)), TypeKind::Pointer);
        unsafe {
            let store = llvm::LLVMBuildStore(self.llbuilder(), val, ptr);
            llvm::LLVMSetAlignment(store, align.bytes() as c_uint);
            store
        }
    }

    pub(crate) fn load(&mut self, ty: Type<'ll>, ptr: Value<'ll>, align: Align) -> Value<'ll> {
        unsafe {
            let load = llvm::LLVMBuildLoad2(self.llbuilder(), ty, ptr, UNNAMED);
            llvm::LLVMSetAlignment(load, align.bytes() as c_uint);
            load
        }
    }
}

/// Empty string, to be used where LLVM expects an instruction name, indicating
/// that the instruction is to be left unnamed (i.e. numbered, in textual IR).
// FIXME(eddyb) pass `&CStr` directly to FFI once it's a thin pointer.
pub(crate) const UNNAMED: *const c_char = c"".as_ptr();

impl<'ll, CX: Borrow<SCx<'ll>>> BackendTypes for GenericBuilder<'_, 'll, CX> {
    type Function = <GenericCx<'ll, CX> as BackendTypes>::Function;
    type BasicBlock = <GenericCx<'ll, CX> as BackendTypes>::BasicBlock;
    type Funclet = <GenericCx<'ll, CX> as BackendTypes>::Funclet;

    type Value = <GenericCx<'ll, CX> as BackendTypes>::Value;
    type Type = <GenericCx<'ll, CX> as BackendTypes>::Type;
    type FunctionSignature = <GenericCx<'ll, CX> as BackendTypes>::FunctionSignature;

    type DIScope = <GenericCx<'ll, CX> as BackendTypes>::DIScope;
    type DILocation = <GenericCx<'ll, CX> as BackendTypes>::DILocation;
    type DIVariable = <GenericCx<'ll, CX> as BackendTypes>::DIVariable;
}

impl abi::HasDataLayout for Builder<'_, '_, '_> {
    fn data_layout(&self) -> &abi::TargetDataLayout {
        self.cx.data_layout()
    }
}

impl<'tcx> ty::layout::HasTyCtxt<'tcx> for Builder<'_, '_, 'tcx> {
    #[inline]
    fn tcx(&self) -> TyCtxt<'tcx> {
        self.cx.tcx
    }
}

impl<'tcx> ty::layout::HasTypingEnv<'tcx> for Builder<'_, '_, 'tcx> {
    fn typing_env(&self) -> ty::TypingEnv<'tcx> {
        self.cx.typing_env()
    }
}

impl HasTargetSpec for Builder<'_, '_, '_> {
    #[inline]
    fn target_spec(&self) -> &Target {
        self.cx.target_spec()
    }
}

impl<'tcx> LayoutOfHelpers<'tcx> for Builder<'_, '_, 'tcx> {
    #[inline]
    fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! {
        self.cx.handle_layout_err(err, span, ty)
    }
}

impl<'tcx> FnAbiOfHelpers<'tcx> for Builder<'_, '_, 'tcx> {
    #[inline]
    fn handle_fn_abi_err(
        &self,
        err: FnAbiError<'tcx>,
        span: Span,
        fn_abi_request: FnAbiRequest<'tcx>,
    ) -> ! {
        self.cx.handle_fn_abi_err(err, span, fn_abi_request)
    }
}

impl<'ll, 'tcx> Deref for Builder<'_, 'll, 'tcx> {
    type Target = CodegenCx<'ll, 'tcx>;

    #[inline]
    fn deref(&self) -> &Self::Target {
        self.cx
    }
}

macro_rules! math_builder_methods {
    ($($name:ident($($arg:ident),*) => $llvm_capi:ident),+ $(,)?) => {
        $(fn $name(&mut self, $($arg: Value<'ll>),*) -> Value<'ll> {
            unsafe {
                llvm::$llvm_capi(self.llbuilder(), $($arg,)* UNNAMED)
            }
        })+
    }
}

macro_rules! set_math_builder_methods {
    ($($name:ident($($arg:ident),*) => ($llvm_capi:ident, $llvm_set_math:ident)),+ $(,)?) => {
        $(fn $name(&mut self, $($arg: Value<'ll>),*) -> Value<'ll> {
            unsafe {
                let instr = llvm::$llvm_capi(self.llbuilder(), $($arg,)* UNNAMED);
                llvm::$llvm_set_math(instr);
                instr
            }
        })+
    }
}

impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
    type CodegenCx = CodegenCx<'ll, 'tcx>;

    fn build(cx: &'a CodegenCx<'ll, 'tcx>, llbb: &'ll BasicBlock) -> Self {
        let bx = Builder::with_cx(cx);
        unsafe {
            llvm::LLVMPositionBuilderAtEnd(bx.llbuilder(), llbb);
        }
        bx
    }

    fn cx(&self) -> &CodegenCx<'ll, 'tcx> {
        self.cx
    }

    fn llbb(&self) -> &'ll BasicBlock {
        unsafe { llvm::LLVMGetInsertBlock(self.llbuilder()) }
    }

    fn set_span(&mut self, span: rustc_span::Span) {
        self.span = span;
    }

    fn append_block(cx: &'a CodegenCx<'ll, 'tcx>, llfn: Value<'ll>, name: &str) -> &'ll BasicBlock {
        unsafe {
            let name = SmallCStr::new(name);
            llvm::LLVMAppendBasicBlockInContext(cx.llcx, llfn, name.as_ptr())
        }
    }

    fn append_sibling_block(&mut self, name: &str) -> &'ll BasicBlock {
        Self::append_block(self.cx, self.llfn(), name)
    }

    fn switch_to_block(&mut self, llbb: Self::BasicBlock) {
        *self = Self::build(self.cx, llbb)
    }

    fn ret_void(&mut self) {
        llvm::LLVMBuildRetVoid(self.llbuilder());
    }

    fn ret(&mut self, v: Value<'ll>) {
        unsafe {
            llvm::LLVMBuildRet(self.llbuilder(), v);
        }
    }

    fn br(&mut self, dest: &'ll BasicBlock) {
        unsafe {
            llvm::LLVMBuildBr(self.llbuilder(), dest);
        }
    }

    fn br_with_attrs(&mut self, dest: &'ll BasicBlock, loop_hint_attrs: &[UnrollAttr]) {
        unsafe {
            let val = llvm::LLVMBuildBr(self.llbuilder(), dest);

            let mut nodes = Vec::new();

            for unroll in loop_hint_attrs {
                // UnrollAttr::Count needs a second operand, the provided count, but the other
                // unroll hints do not.
                let md_node = if let UnrollAttr::Count(count) = unroll {
                    let unroll_meta = self.create_metadata("llvm.loop.unroll.count".as_bytes());
                    let count = llvm::LLVMValueAsMetadata(self.get_const_i32(u64::from(*count)));
                    self.md_node_in_context(&[unroll_meta, count])
                } else {
                    let metadata_str = match unroll {
                        UnrollAttr::Hint => "llvm.loop.unroll.enable",
                        UnrollAttr::Full => "llvm.loop.unroll.full",
                        UnrollAttr::Never => "llvm.loop.unroll.disable",
                        UnrollAttr::Count(_) => unreachable!(),
                    };
                    let unroll_meta = self.create_metadata(metadata_str.as_bytes());
                    self.md_node_in_context(&[unroll_meta])
                };
                nodes.push(md_node);
            }

            if let [first, ..] = nodes[..] {
                nodes.insert(0, first);

                // Create the loop metadata node
                let loop_meta_mdnode = self.set_metadata_node(val, llvm::MD_loop, &nodes);

                // Look up the metadata node as a value
                let loop_meta_val = llvm::LLVMGetMetadata(val, llvm::MD_loop).unwrap();

                // Replace the first entry with a reference to itself
                // This is required by LLVM. See the LangRef page for llvm.loop metadata.
                llvm::LLVMReplaceMDNodeOperandWith(loop_meta_val, 0, loop_meta_mdnode);
            }
        }
    }

    fn cond_br(
        &mut self,
        cond: Value<'ll>,
        then_llbb: &'ll BasicBlock,
        else_llbb: &'ll BasicBlock,
    ) {
        unsafe {
            llvm::LLVMBuildCondBr(self.llbuilder(), cond, then_llbb, else_llbb);
        }
    }

    fn switch(
        &mut self,
        v: Value<'ll>,
        else_llbb: &'ll BasicBlock,
        cases: impl ExactSizeIterator<Item = (u128, &'ll BasicBlock)>,
    ) {
        let switch =
            unsafe { llvm::LLVMBuildSwitch(self.llbuilder(), v, else_llbb, cases.len() as c_uint) };
        for (on_val, dest) in cases {
            let on_val = self.const_uint_big(self.val_ty(v), on_val);
            unsafe { llvm::LLVMAddCase(switch, on_val, dest) }
        }
    }

    fn switch_with_weights(
        &mut self,
        v: Self::Value,
        else_llbb: Self::BasicBlock,
        else_is_cold: bool,
        cases: impl ExactSizeIterator<Item = (u128, Self::BasicBlock, bool)>,
    ) {
        if self.cx.sess().opts.optimize == rustc_session::config::OptLevel::No {
            self.switch(v, else_llbb, cases.map(|(val, dest, _)| (val, dest)));
            return;
        }

        let id = self.cx.create_metadata(b"branch_weights");

        // For switch instructions with 2 targets, the `llvm.expect` intrinsic is used.
        // This function handles switch instructions with more than 2 targets and it needs to
        // emit branch weights metadata instead of using the intrinsic.
        // The values 1 and 2000 are the same as the values used by the `llvm.expect` intrinsic.
        let cold_weight = llvm::LLVMValueAsMetadata(self.cx.const_u32(1));
        let hot_weight = llvm::LLVMValueAsMetadata(self.cx.const_u32(2000));
        let weight =
            |is_cold: bool| -> &Metadata { if is_cold { cold_weight } else { hot_weight } };

        let mut md: SmallVec<[&Metadata; 16]> = SmallVec::with_capacity(cases.len() + 2);
        md.push(id);
        md.push(weight(else_is_cold));

        let switch =
            unsafe { llvm::LLVMBuildSwitch(self.llbuilder(), v, else_llbb, cases.len() as c_uint) };
        for (on_val, dest, is_cold) in cases {
            let on_val = self.const_uint_big(self.val_ty(v), on_val);
            unsafe { llvm::LLVMAddCase(switch, on_val, dest) }
            md.push(weight(is_cold));
        }

        self.cx.set_metadata_node(switch, llvm::MD_prof, &md);
    }

    fn invoke(
        &mut self,
        llty: Type<'ll>,
        fn_attrs: Option<&CodegenFnAttrs>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
        llfn: Value<'ll>,
        return_slot: ReturnSlot<Value<'ll>>,
        args: &[Value<'ll>],
        then: &'ll BasicBlock,
        catch: &'ll BasicBlock,
        funclet: Option<&Funclet<'ll>>,
        instance: Option<Instance<'tcx>>,
    ) -> Value<'ll> {
        // If this function returns indirectly (`PassMode::Indirect`),
        // the `return_slot` should be the first argument.
        let args = match return_slot {
            ReturnSlot::Direct => args.to_vec(),
            ReturnSlot::Indirect(sret_ptr) => {
                let mut args = args.to_vec();
                args.insert(0, sret_ptr);
                args
            }
        };
        debug!("invoke {:?} with args ({:?})", llfn, args);
        let args = self.check_call("invoke", llty, llfn, &args);
        let funclet_bundle = funclet.map(|funclet| funclet.bundle());
        let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
        if let Some(funclet_bundle) = funclet_bundle {
            bundles.push(funclet_bundle);
        }

        // Emit CFI pointer type membership test
        self.cfi_type_test(fn_attrs, fn_abi, instance, llfn);

        // Emit KCFI operand bundle
        let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn);
        if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.as_ref()) {
            bundles.push(kcfi_bundle);
        }

        let pauth = self.ptrauth_operand_bundle(llfn, fn_abi);
        if let Some(p) = pauth.as_ref().map(|b| b.as_ref()) {
            bundles.push(p);
        }

        let invoke = unsafe {
            llvm::LLVMBuildInvokeWithOperandBundles(
                self.llbuilder(),
                llty,
                llfn,
                args.as_ptr(),
                args.len() as c_uint,
                then,
                catch,
                bundles.as_ptr(),
                bundles.len() as c_uint,
                UNNAMED,
            )
        };
        if let Some(fn_abi) = fn_abi {
            fn_abi.apply_attrs_callsite(self, invoke);
        }
        invoke
    }

    fn unreachable(&mut self) {
        unsafe {
            llvm::LLVMBuildUnreachable(self.llbuilder());
        }
    }

    math_builder_methods! {
        add(a, b) => LLVMBuildAdd,
        fadd(a, b) => LLVMBuildFAdd,
        sub(a, b) => LLVMBuildSub,
        fsub(a, b) => LLVMBuildFSub,
        mul(a, b) => LLVMBuildMul,
        fmul(a, b) => LLVMBuildFMul,
        udiv(a, b) => LLVMBuildUDiv,
        exactudiv(a, b) => LLVMBuildExactUDiv,
        sdiv(a, b) => LLVMBuildSDiv,
        exactsdiv(a, b) => LLVMBuildExactSDiv,
        fdiv(a, b) => LLVMBuildFDiv,
        urem(a, b) => LLVMBuildURem,
        srem(a, b) => LLVMBuildSRem,
        frem(a, b) => LLVMBuildFRem,
        shl(a, b) => LLVMBuildShl,
        lshr(a, b) => LLVMBuildLShr,
        ashr(a, b) => LLVMBuildAShr,
        and(a, b) => LLVMBuildAnd,
        or(a, b) => LLVMBuildOr,
        xor(a, b) => LLVMBuildXor,
        neg(x) => LLVMBuildNeg,
        fneg(x) => LLVMBuildFNeg,
        not(x) => LLVMBuildNot,
        unchecked_sadd(x, y) => LLVMBuildNSWAdd,
        unchecked_uadd(x, y) => LLVMBuildNUWAdd,
        unchecked_ssub(x, y) => LLVMBuildNSWSub,
        unchecked_usub(x, y) => LLVMBuildNUWSub,
        unchecked_smul(x, y) => LLVMBuildNSWMul,
        unchecked_umul(x, y) => LLVMBuildNUWMul,
    }

    fn unchecked_suadd(&mut self, a: Value<'ll>, b: Value<'ll>) -> Value<'ll> {
        unsafe {
            let add = llvm::LLVMBuildAdd(self.llbuilder(), a, b, UNNAMED);
            if llvm::LLVMIsAInstruction(add).is_some() {
                llvm::LLVMSetNUW(add, TRUE);
                llvm::LLVMSetNSW(add, TRUE);
            }
            add
        }
    }
    fn unchecked_susub(&mut self, a: Value<'ll>, b: Value<'ll>) -> Value<'ll> {
        unsafe {
            let sub = llvm::LLVMBuildSub(self.llbuilder(), a, b, UNNAMED);
            if llvm::LLVMIsAInstruction(sub).is_some() {
                llvm::LLVMSetNUW(sub, TRUE);
                llvm::LLVMSetNSW(sub, TRUE);
            }
            sub
        }
    }
    fn unchecked_sumul(&mut self, a: Value<'ll>, b: Value<'ll>) -> Value<'ll> {
        unsafe {
            let mul = llvm::LLVMBuildMul(self.llbuilder(), a, b, UNNAMED);
            if llvm::LLVMIsAInstruction(mul).is_some() {
                llvm::LLVMSetNUW(mul, TRUE);
                llvm::LLVMSetNSW(mul, TRUE);
            }
            mul
        }
    }

    fn or_disjoint(&mut self, a: Value<'ll>, b: Value<'ll>) -> Value<'ll> {
        unsafe {
            let or = llvm::LLVMBuildOr(self.llbuilder(), a, b, UNNAMED);

            // If a and b are both values, then `or` is a value, rather than
            // an instruction, so we need to check before setting the flag.
            // (See also `LLVMBuildNUWNeg` which also needs a check.)
            if llvm::LLVMIsAInstruction(or).is_some() {
                llvm::LLVMSetIsDisjoint(or, TRUE);
            }
            or
        }
    }

    set_math_builder_methods! {
        fadd_fast(x, y) => (LLVMBuildFAdd, LLVMRustSetFastMath),
        fsub_fast(x, y) => (LLVMBuildFSub, LLVMRustSetFastMath),
        fmul_fast(x, y) => (LLVMBuildFMul, LLVMRustSetFastMath),
        fdiv_fast(x, y) => (LLVMBuildFDiv, LLVMRustSetFastMath),
        frem_fast(x, y) => (LLVMBuildFRem, LLVMRustSetFastMath),
        fadd_algebraic(x, y) => (LLVMBuildFAdd, LLVMRustSetAlgebraicMath),
        fsub_algebraic(x, y) => (LLVMBuildFSub, LLVMRustSetAlgebraicMath),
        fmul_algebraic(x, y) => (LLVMBuildFMul, LLVMRustSetAlgebraicMath),
        fdiv_algebraic(x, y) => (LLVMBuildFDiv, LLVMRustSetAlgebraicMath),
        frem_algebraic(x, y) => (LLVMBuildFRem, LLVMRustSetAlgebraicMath),
    }

    fn checked_binop(
        &mut self,
        oop: OverflowOp,
        ty: Ty<'tcx>,
        lhs: Self::Value,
        rhs: Self::Value,
    ) -> (Self::Value, Self::Value) {
        let (size, signed) = ty.int_size_and_signed(self.tcx);
        let width = size.bits();

        if !signed {
            match oop {
                OverflowOp::Sub => {
                    // Emit sub and icmp instead of llvm.usub.with.overflow. LLVM considers these
                    // to be the canonical form. It will attempt to reform llvm.usub.with.overflow
                    // in the backend if profitable.
                    let sub = self.sub(lhs, rhs);
                    let cmp = self.icmp(IntPredicate::IntULT, lhs, rhs);
                    return (sub, cmp);
                }
                OverflowOp::Add => {
                    // Like with sub above, using icmp is the preferred form. See
                    // <https://rust-lang.zulipchat.com/#narrow/channel/187780-t-compiler.2Fllvm/topic/.60uadd.2Ewith.2Eoverflow.60.20.28again.29/near/533041085>
                    let add = self.add(lhs, rhs);
                    let cmp = self.icmp(IntPredicate::IntULT, add, lhs);
                    return (add, cmp);
                }
                OverflowOp::Mul => {}
            }
        }

        let oop_str = match oop {
            OverflowOp::Add => "add",
            OverflowOp::Sub => "sub",
            OverflowOp::Mul => "mul",
        };

        let name = format!("llvm.{}{oop_str}.with.overflow", if signed { 's' } else { 'u' });

        let res = self.call_intrinsic(name, &[self.type_ix(width)], &[lhs, rhs]);
        (self.extract_value(res, 0), self.extract_value(res, 1))
    }

    fn from_immediate(&mut self, val: Self::Value) -> Self::Value {
        if self.cx().val_ty(val) == self.cx().type_i1() {
            self.zext(val, self.cx().type_i8())
        } else {
            val
        }
    }

    fn to_immediate_scalar(&mut self, val: Self::Value, scalar: abi::Scalar) -> Self::Value {
        if scalar.is_bool() {
            return self.unchecked_utrunc(val, self.cx().type_i1());
        }
        val
    }

    fn alloca(&mut self, size: Size, align: Align) -> Value<'ll> {
        let mut bx = Builder::with_cx(self.cx);
        bx.position_at_start(unsafe { llvm::LLVMGetFirstBasicBlock(self.llfn()) });
        bx.direct_alloca(size, align, "")
    }

    fn alloca_with_ty(&mut self, layout: TyAndLayout<'tcx>) -> Self::Value {
        let mut bx = Builder::with_cx(self.cx);
        bx.position_at_start(unsafe { llvm::LLVMGetFirstBasicBlock(self.llfn()) });
        let scalable_vector_ty = layout.llvm_type(self.cx);

        unsafe {
            let alloca = llvm::LLVMBuildAlloca(bx.llbuilder(), scalable_vector_ty, UNNAMED);
            llvm::LLVMSetAlignment(alloca, layout.align.bytes() as c_uint);
            alloca
        }
    }

    fn load(&mut self, ty: Type<'ll>, ptr: Value<'ll>, align: Align) -> Value<'ll> {
        unsafe {
            let load = llvm::LLVMBuildLoad2(self.llbuilder(), ty, ptr, UNNAMED);
            let align = align.min(self.cx().tcx.sess.target.max_reliable_alignment());
            llvm::LLVMSetAlignment(load, align.bytes() as c_uint);
            load
        }
    }

    fn volatile_load(&mut self, ty: Type<'ll>, ptr: Value<'ll>, align: Align) -> Value<'ll> {
        unsafe {
            let load = self.load(ty, ptr, align);
            llvm::LLVMSetVolatile(load, llvm::TRUE);
            load
        }
    }

    fn atomic_load(
        &mut self,
        ty: Type<'ll>,
        ptr: Value<'ll>,
        order: rustc_middle::ty::AtomicOrdering,
        volatile: bool,
        size: Size,
    ) -> Value<'ll> {
        unsafe {
            let load = llvm::LLVMBuildLoad2(self.llbuilder(), ty, ptr, UNNAMED);
            // Set atomic ordering
            llvm::LLVMSetOrdering(load, AtomicOrdering::from_generic(order));
            if volatile {
                llvm::LLVMSetVolatile(load, llvm::TRUE);
            }
            // LLVM requires the alignment of atomic loads to be at least the size of the type.
            llvm::LLVMSetAlignment(load, size.bytes() as c_uint);
            load
        }
    }

    #[instrument(level = "trace", skip(self))]
    fn load_operand(&mut self, place: PlaceRef<'tcx, Value<'ll>>) -> OperandRef<'tcx, Value<'ll>> {
        if place.layout.is_unsized() {
            let tail = self.tcx.struct_tail_for_codegen(place.layout.ty, self.typing_env());
            if matches!(tail.kind(), ty::Foreign(..)) {
                // Unsized locals and, at least conceptually, even unsized arguments must be copied
                // around, which requires dynamically determining their size. Therefore, we cannot
                // allow `extern` types here. Consult t-opsem before removing this check.
                panic!("unsized locals must not be `extern` types");
            }
        }
        assert_eq!(place.val.llextra.is_some(), place.layout.is_unsized());

        if place.layout.is_zst() {
            return OperandRef::zero_sized(place.layout);
        }

        #[instrument(level = "trace", skip(bx))]
        fn scalar_load_metadata<'a, 'll, 'tcx>(
            bx: &mut Builder<'a, 'll, 'tcx>,
            load: Value<'ll>,
            scalar: abi::Scalar,
            layout: TyAndLayout<'tcx>,
            offset: Size,
        ) {
            if bx.cx.sess().opts.optimize == OptLevel::No {
                // Don't emit metadata we're not going to use
                return;
            }

            if !scalar.is_uninit_valid() {
                bx.noundef_metadata(load);
            }

            match scalar.primitive() {
                abi::Primitive::Int(..) => {
                    if !scalar.is_always_valid(bx) {
                        bx.range_metadata(load, scalar.valid_range(bx));
                    }
                }
                abi::Primitive::Pointer(_) => {
                    if !scalar.valid_range(bx).contains(0) {
                        bx.nonnull_metadata(load);
                    }

                    if let Some(pointee) = layout.pointee_info_at(bx, offset)
                        && pointee.align > Align::ONE
                    {
                        bx.align_metadata(load, pointee.align);
                    }
                }
                abi::Primitive::Float(_) => {}
            }
        }

        let val = if let Some(_) = place.val.llextra {
            // FIXME: Merge with the `else` below?
            OperandValue::Ref(place.val)
        } else if place.layout.backend_repr.is_scalar_or_simd() {
            let mut const_llval = None;
            let llty = place.layout.llvm_type(self);
            if let Some(global) = llvm::LLVMIsAGlobalVariable(place.val.llval) {
                if llvm::LLVMIsGlobalConstant(global).is_true() {
                    if let Some(init) = llvm::LLVMGetInitializer(global) {
                        if self.val_ty(init) == llty {
                            const_llval = Some(init);
                        }
                    }
                }
            }

            let llval = const_llval.unwrap_or_else(|| {
                let load = self.load(llty, place.val.llval, place.val.align);
                if let abi::BackendRepr::Scalar(scalar) = place.layout.backend_repr {
                    scalar_load_metadata(self, load, scalar, place.layout, Size::ZERO);
                    self.to_immediate_scalar(load, scalar)
                } else {
                    load
                }
            });
            OperandValue::Immediate(llval)
        } else if let abi::BackendRepr::ScalarPair { a, b, b_offset } = place.layout.backend_repr {
            let mut load = |i, scalar: abi::Scalar, layout, align, offset| {
                let llptr = if i == 0 {
                    place.val.llval
                } else {
                    self.inbounds_ptradd(place.val.llval, self.const_usize(b_offset.bytes()))
                };
                let llty = place.layout.scalar_pair_element_llvm_type(self, i, false);
                let load = self.load(llty, llptr, align);
                scalar_load_metadata(self, load, scalar, layout, offset);
                self.to_immediate_scalar(load, scalar)
            };

            OperandValue::Pair(
                load(0, a, place.layout, place.val.align, Size::ZERO),
                load(1, b, place.layout, place.val.align.restrict_for_offset(b_offset), b_offset),
            )
        } else {
            OperandValue::Ref(place.val)
        };

        OperandRef { val, layout: place.layout, move_annotation: None }
    }

    fn write_operand_repeatedly(
        &mut self,
        cg_elem: OperandRef<'tcx, Value<'ll>>,
        count: u64,
        dest: PlaceRef<'tcx, Value<'ll>>,
    ) {
        if self.cx.sess().opts.optimize == OptLevel::No {
            // To let debuggers single-step over lines like
            //
            //     let foo = ["bar"; 42];
            //
            // we need the debugger-friendly LLVM IR that `_unoptimized()`
            // provides. The `_optimized()` version generates trickier LLVM IR.
            // See PR #148058 for a failed attempt at handling that.
            self.write_operand_repeatedly_unoptimized(cg_elem, count, dest);
        } else {
            self.write_operand_repeatedly_optimized(cg_elem, count, dest);
        }
    }

    fn range_metadata(&mut self, load: Value<'ll>, range: WrappingRange) {
        if self.cx.sess().opts.optimize == OptLevel::No {
            // Don't emit metadata we're not going to use
            return;
        }

        let llty = self.cx.val_ty(load);
        let md = [
            llvm::LLVMValueAsMetadata(self.cx.const_uint_big(llty, range.start)),
            llvm::LLVMValueAsMetadata(self.cx.const_uint_big(llty, range.end.wrapping_add(1))),
        ];
        self.set_metadata_node(load, llvm::MD_range, &md);
    }

    fn nonnull_metadata(&mut self, load: Value<'ll>) {
        self.set_metadata_node(load, llvm::MD_nonnull, &[]);
    }

    fn store_with_flags(
        &mut self,
        val: Value<'ll>,
        ptr: Value<'ll>,
        align: Align,
        flags: MemFlags,
    ) {
        debug!("Store {:?} -> {:?} ({:?})", val, ptr, flags);
        assert_eq!(self.cx.type_kind(self.cx.val_ty(ptr)), TypeKind::Pointer);
        unsafe {
            let store = llvm::LLVMBuildStore(self.llbuilder(), val, ptr);
            let align = align.min(self.cx().tcx.sess.target.max_reliable_alignment());
            let align = align.bytes() as c_uint;
            llvm::LLVMSetAlignment(store, align);
            if flags.contains(MemFlags::VOLATILE) {
                llvm::LLVMSetVolatile(store, llvm::TRUE);
            }
            if flags.contains(MemFlags::NONTEMPORAL) {
                // Make sure that the current target architectures supports "sane" non-temporal
                // stores, i.e., non-temporal stores that are equivalent to regular stores except
                // for performance. LLVM doesn't seem to care about this, and will happily treat
                // `!nontemporal` stores as-if they were normal stores (for reordering optimizations
                // etc) even on x86, despite later lowering them to MOVNT which do *not* behave like
                // regular stores but require special fences. So we keep a list of architectures
                // where `!nontemporal` is known to be truly just a hint, and use regular stores
                // everywhere else. (In the future, we could alternatively ensure that an sfence
                // gets emitted after a sequence of movnt before any kind of synchronizing
                // operation. But it's not clear how to do that with LLVM.)
                // For more context, see <https://github.com/rust-lang/rust/issues/114582> and
                // <https://github.com/llvm/llvm-project/issues/64521>.
                let use_nontemporal = matches!(
                    self.cx.tcx.sess.target.arch,
                    Arch::AArch64 | Arch::Arm | Arch::RiscV32 | Arch::RiscV64
                );
                if use_nontemporal {
                    // According to LLVM [1] building a nontemporal store must
                    // *always* point to a metadata value of the integer 1.
                    //
                    // [1]: https://llvm.org/docs/LangRef.html#store-instruction
                    let one = llvm::LLVMValueAsMetadata(self.cx.const_i32(1));
                    self.set_metadata_node(store, llvm::MD_nontemporal, &[one]);
                }
            }
            if flags.contains(MemFlags::CAPTURES_READ_ONLY) {
                assert!(
                    self.type_kind(self.val_ty(val)) == TypeKind::Pointer,
                    "CAPTURED_READ_ONLY is only supported on pointer stores"
                );
                let args = [
                    self.cx.create_metadata(b"address"),
                    self.cx.create_metadata(b"read_provenance"),
                ];
                // FIXME: Switch this to use MD_captures once LLVM 22 is the minimum.
                let id = self.get_md_kind_id("captures");
                let md = llvm::LLVMMDNodeInContext2(self.cx.llcx, args.as_ptr(), args.len());
                self.set_metadata(store, id, md);
            }
        }
    }

    fn atomic_store(
        &mut self,
        val: Value<'ll>,
        ptr: Value<'ll>,
        order: rustc_middle::ty::AtomicOrdering,
        volatile: bool,
        size: Size,
    ) {
        debug!("Store {:?} -> {:?}", val, ptr);
        assert_eq!(self.cx.type_kind(self.cx.val_ty(ptr)), TypeKind::Pointer);
        unsafe {
            let store = llvm::LLVMBuildStore(self.llbuilder(), val, ptr);
            // Set atomic ordering
            llvm::LLVMSetOrdering(store, AtomicOrdering::from_generic(order));
            if volatile {
                llvm::LLVMSetVolatile(store, llvm::TRUE);
            }
            // LLVM requires the alignment of atomic stores to be at least the size of the type.
            llvm::LLVMSetAlignment(store, size.bytes() as c_uint);
        }
    }

    fn gep(&mut self, stride: Size, ptr: Value<'ll>, index: Value<'ll>) -> Value<'ll> {
        self.build_gep(stride, ptr, index, GEPNoWrapFlags::default())
    }

    fn inbounds_gep(&mut self, stride: Size, ptr: Value<'ll>, index: Value<'ll>) -> Value<'ll> {
        self.build_gep(stride, ptr, index, GEPNoWrapFlags::InBounds)
    }

    fn inbounds_nuw_gep(&mut self, stride: Size, ptr: Value<'ll>, index: Value<'ll>) -> Value<'ll> {
        self.build_gep(stride, ptr, index, GEPNoWrapFlags::InBounds | GEPNoWrapFlags::NUW)
    }

    /* Casts */
    fn trunc(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildTrunc(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn unchecked_utrunc(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        debug_assert_ne!(self.val_ty(val), dest_ty);

        let trunc = self.trunc(val, dest_ty);
        unsafe {
            if llvm::LLVMIsAInstruction(trunc).is_some() {
                llvm::LLVMSetNUW(trunc, TRUE);
            }
        }
        trunc
    }

    fn unchecked_strunc(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        debug_assert_ne!(self.val_ty(val), dest_ty);

        let trunc = self.trunc(val, dest_ty);
        unsafe {
            if llvm::LLVMIsAInstruction(trunc).is_some() {
                llvm::LLVMSetNSW(trunc, TRUE);
            }
        }
        trunc
    }

    fn sext(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildSExt(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn fptoui_sat(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.fptoui.sat", &[dest_ty, self.val_ty(val)], &[val])
    }

    fn fptosi_sat(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.fptosi.sat", &[dest_ty, self.val_ty(val)], &[val])
    }

    fn fptoui(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        // On WebAssembly the `fptoui` and `fptosi` instructions currently have
        // poor codegen. The reason for this is that the corresponding wasm
        // instructions, `i32.trunc_f32_s` for example, will trap when the float
        // is out-of-bounds, infinity, or nan. This means that LLVM
        // automatically inserts control flow around `fptoui` and `fptosi`
        // because the LLVM instruction `fptoui` is defined as producing a
        // poison value, not having UB on out-of-bounds values.
        //
        // This method, however, is only used with non-saturating casts that
        // have UB on out-of-bounds values. This means that it's ok if we use
        // the raw wasm instruction since out-of-bounds values can do whatever
        // we like. To ensure that LLVM picks the right instruction we choose
        // the raw wasm intrinsic functions which avoid LLVM inserting all the
        // other control flow automatically.
        if self.sess().target.is_like_wasm {
            let src_ty = self.cx.val_ty(val);
            if let Some(float_type) = self.cx.float_type(src_ty) {
                let int_width = self.cx.int_width(dest_ty).expect("dest was not an integer");
                if matches!((int_width, float_type), (32 | 64, Float::F32 | Float::F64)) {
                    return self.call_intrinsic(
                        "llvm.wasm.trunc.unsigned",
                        &[dest_ty, src_ty],
                        &[val],
                    );
                }
            }
        }
        unsafe { llvm::LLVMBuildFPToUI(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn fptosi(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        // see `fptoui` above for why wasm is different here
        if self.sess().target.is_like_wasm {
            let src_ty = self.cx.val_ty(val);
            if let Some(float_type) = self.cx.float_type(src_ty) {
                let int_width = self.cx.int_width(dest_ty).expect("dest ty was not an integer");
                if matches!((int_width, float_type), (32 | 64, Float::F32 | Float::F64)) {
                    return self.call_intrinsic(
                        "llvm.wasm.trunc.signed",
                        &[dest_ty, src_ty],
                        &[val],
                    );
                }
            }
        }
        unsafe { llvm::LLVMBuildFPToSI(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn uitofp(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildUIToFP(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn sitofp(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildSIToFP(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn fptrunc(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildFPTrunc(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn fpext(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildFPExt(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn ptrtoint(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildPtrToInt(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn inttoptr(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildIntToPtr(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn bitcast(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildBitCast(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn intcast(&mut self, val: Value<'ll>, dest_ty: Type<'ll>, is_signed: bool) -> Value<'ll> {
        unsafe {
            llvm::LLVMBuildIntCast2(
                self.llbuilder(),
                val,
                dest_ty,
                is_signed.to_llvm_bool(),
                UNNAMED,
            )
        }
    }

    fn pointercast(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildPointerCast(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    /* Comparisons */
    fn icmp(&mut self, op: IntPredicate, lhs: Value<'ll>, rhs: Value<'ll>) -> Value<'ll> {
        let op = llvm::IntPredicate::from_generic(op);
        unsafe { llvm::LLVMBuildICmp(self.llbuilder(), op as c_uint, lhs, rhs, UNNAMED) }
    }

    fn fcmp(&mut self, op: RealPredicate, lhs: Value<'ll>, rhs: Value<'ll>) -> Value<'ll> {
        let op = llvm::RealPredicate::from_generic(op);
        unsafe { llvm::LLVMBuildFCmp(self.llbuilder(), op as c_uint, lhs, rhs, UNNAMED) }
    }

    fn three_way_compare(
        &mut self,
        ty: Ty<'tcx>,
        lhs: Self::Value,
        rhs: Self::Value,
    ) -> Self::Value {
        let size = ty.primitive_size(self.tcx);
        let name = if ty.is_signed() { "llvm.scmp" } else { "llvm.ucmp" };

        self.call_intrinsic(name, &[self.type_i8(), self.type_ix(size.bits())], &[lhs, rhs])
    }

    /* Miscellaneous instructions */
    fn memcpy(
        &mut self,
        dst: Value<'ll>,
        dst_align: Align,
        src: Value<'ll>,
        src_align: Align,
        size: Value<'ll>,
        flags: MemFlags,
        tt: Option<FncTree>,
    ) {
        assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memcpy not supported");
        let size = self.intcast(size, self.type_isize(), false);
        let is_volatile = flags.contains(MemFlags::VOLATILE);
        let memcpy = unsafe {
            llvm::LLVMRustBuildMemCpy(
                self.llbuilder(),
                dst,
                dst_align.bytes() as c_uint,
                src,
                src_align.bytes() as c_uint,
                size,
                is_volatile,
            )
        };

        // TypeTree metadata for memcpy is especially important: when Enzyme encounters
        // a memcpy during autodiff, it needs to know the structure of the data being
        // copied to properly track derivatives. For example, copying an array of floats
        // vs. copying a struct with mixed types requires different derivative handling.
        // The TypeTree tells Enzyme exactly what memory layout to expect.
        if let Some(tt) = tt {
            crate::typetree::add_tt(self, memcpy, tt);
        }
    }

    fn memmove(
        &mut self,
        dst: Value<'ll>,
        dst_align: Align,
        src: Value<'ll>,
        src_align: Align,
        size: Value<'ll>,
        flags: MemFlags,
    ) {
        assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memmove not supported");
        let size = self.intcast(size, self.type_isize(), false);
        let is_volatile = flags.contains(MemFlags::VOLATILE);
        unsafe {
            llvm::LLVMRustBuildMemMove(
                self.llbuilder(),
                dst,
                dst_align.bytes() as c_uint,
                src,
                src_align.bytes() as c_uint,
                size,
                is_volatile,
            );
        }
    }

    fn memset(
        &mut self,
        ptr: Value<'ll>,
        fill_byte: Value<'ll>,
        size: Value<'ll>,
        align: Align,
        flags: MemFlags,
    ) {
        assert!(!flags.contains(MemFlags::NONTEMPORAL), "non-temporal memset not supported");
        let is_volatile = flags.contains(MemFlags::VOLATILE);
        unsafe {
            llvm::LLVMRustBuildMemSet(
                self.llbuilder(),
                ptr,
                align.bytes() as c_uint,
                fill_byte,
                size,
                is_volatile,
            );
        }
    }

    fn vscale(&mut self, ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMRustBuildVScale(self.llbuilder(), ty) }
    }

    fn select(
        &mut self,
        cond: Value<'ll>,
        then_val: Value<'ll>,
        else_val: Value<'ll>,
    ) -> Value<'ll> {
        unsafe { llvm::LLVMBuildSelect(self.llbuilder(), cond, then_val, else_val, UNNAMED) }
    }

    fn va_arg(&mut self, list: Value<'ll>, ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildVAArg(self.llbuilder(), list, ty, UNNAMED) }
    }

    fn extract_element(&mut self, vec: Value<'ll>, idx: Value<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildExtractElement(self.llbuilder(), vec, idx, UNNAMED) }
    }

    fn vector_splat(&mut self, num_elts: u64, elt: Value<'ll>) -> Value<'ll> {
        unsafe {
            let elt_ty = self.cx.val_ty(elt);
            let undef = llvm::LLVMGetUndef(self.type_vector(elt_ty, num_elts));
            let vec = self.insert_element(undef, elt, self.cx.const_i32(0));
            let vec_i32_ty = self.type_vector(self.type_i32(), num_elts);
            self.shuffle_vector(vec, undef, self.const_null(vec_i32_ty))
        }
    }

    fn extract_value(&mut self, agg_val: Value<'ll>, idx: u64) -> Value<'ll> {
        assert_eq!(idx as c_uint as u64, idx);
        unsafe { llvm::LLVMBuildExtractValue(self.llbuilder(), agg_val, idx as c_uint, UNNAMED) }
    }

    fn insert_value(&mut self, agg_val: Value<'ll>, elt: Value<'ll>, idx: u64) -> Value<'ll> {
        assert_eq!(idx as c_uint as u64, idx);
        unsafe {
            llvm::LLVMBuildInsertValue(self.llbuilder(), agg_val, elt, idx as c_uint, UNNAMED)
        }
    }

    fn set_personality_fn(&mut self, personality: Value<'ll>) {
        unsafe {
            llvm::LLVMSetPersonalityFn(self.llfn(), personality);
        }
    }

    fn cleanup_landing_pad(&mut self, pers_fn: Value<'ll>) -> (Value<'ll>, Value<'ll>) {
        let ty = self.type_struct(&[self.type_ptr(), self.type_i32()]);
        let landing_pad = self.landing_pad(ty, pers_fn, 0);
        unsafe {
            llvm::LLVMSetCleanup(landing_pad, llvm::TRUE);
        }
        (self.extract_value(landing_pad, 0), self.extract_value(landing_pad, 1))
    }

    fn filter_landing_pad(&mut self, pers_fn: Value<'ll>) {
        let ty = self.type_struct(&[self.type_ptr(), self.type_i32()]);
        let landing_pad = self.landing_pad(ty, pers_fn, 1);
        self.add_clause(landing_pad, self.const_array(self.type_ptr(), &[]));
    }

    fn resume(&mut self, exn0: Value<'ll>, exn1: Value<'ll>) {
        let ty = self.type_struct(&[self.type_ptr(), self.type_i32()]);
        let mut exn = self.const_poison(ty);
        exn = self.insert_value(exn, exn0, 0);
        exn = self.insert_value(exn, exn1, 1);
        unsafe {
            llvm::LLVMBuildResume(self.llbuilder(), exn);
        }
    }

    fn cleanup_pad(&mut self, parent: Option<Value<'ll>>, args: &[Value<'ll>]) -> Funclet<'ll> {
        let ret = unsafe {
            llvm::LLVMBuildCleanupPad(
                self.llbuilder(),
                parent,
                args.as_ptr(),
                args.len() as c_uint,
                c"cleanuppad".as_ptr(),
            )
        };
        Funclet::new(ret.expect("LLVM does not have support for cleanuppad"))
    }

    fn cleanup_ret(&mut self, funclet: &Funclet<'ll>, unwind: Option<&'ll BasicBlock>) {
        unsafe {
            llvm::LLVMBuildCleanupRet(self.llbuilder(), funclet.cleanuppad(), unwind)
                .expect("LLVM does not have support for cleanupret");
        }
    }

    fn catch_pad(&mut self, parent: Value<'ll>, args: &[Value<'ll>]) -> Funclet<'ll> {
        let ret = unsafe {
            llvm::LLVMBuildCatchPad(
                self.llbuilder(),
                parent,
                args.as_ptr(),
                args.len() as c_uint,
                c"catchpad".as_ptr(),
            )
        };
        Funclet::new(ret.expect("LLVM does not have support for catchpad"))
    }

    fn catch_switch(
        &mut self,
        parent: Option<Value<'ll>>,
        unwind: Option<&'ll BasicBlock>,
        handlers: &[&'ll BasicBlock],
    ) -> Value<'ll> {
        let ret = unsafe {
            llvm::LLVMBuildCatchSwitch(
                self.llbuilder(),
                parent,
                unwind,
                handlers.len() as c_uint,
                c"catchswitch".as_ptr(),
            )
        };
        let ret = ret.expect("LLVM does not have support for catchswitch");
        for handler in handlers {
            unsafe {
                llvm::LLVMAddHandler(ret, handler);
            }
        }
        ret
    }

    fn get_funclet_cleanuppad(&self, funclet: &Funclet<'ll>) -> Value<'ll> {
        funclet.cleanuppad()
    }

    // Atomic Operations
    fn atomic_cmpxchg(
        &mut self,
        dst: Value<'ll>,
        cmp: Value<'ll>,
        src: Value<'ll>,
        order: rustc_middle::ty::AtomicOrdering,
        failure_order: rustc_middle::ty::AtomicOrdering,
        weak: bool,
    ) -> (Value<'ll>, Value<'ll>) {
        unsafe {
            let value = llvm::LLVMBuildAtomicCmpXchg(
                self.llbuilder(),
                dst,
                cmp,
                src,
                AtomicOrdering::from_generic(order),
                AtomicOrdering::from_generic(failure_order),
                llvm::FALSE, // SingleThreaded
            );
            llvm::LLVMSetWeak(value, weak.to_llvm_bool());
            let val = self.extract_value(value, 0);
            let success = self.extract_value(value, 1);
            (val, success)
        }
    }

    fn atomic_rmw(
        &mut self,
        op: rustc_codegen_ssa::common::AtomicRmwBinOp,
        dst: Value<'ll>,
        src: Value<'ll>,
        order: rustc_middle::ty::AtomicOrdering,
        ret_ptr: bool,
    ) -> Value<'ll> {
        // FIXME: If `ret_ptr` is true and `src` is not a pointer, we *should* tell LLVM that the
        // LHS is a pointer and the operation should be provenance-preserving, but LLVM does not
        // currently support that (https://github.com/llvm/llvm-project/issues/120837).
        let mut res = unsafe {
            llvm::LLVMBuildAtomicRMW(
                self.llbuilder(),
                AtomicRmwBinOp::from_generic(op),
                dst,
                src,
                AtomicOrdering::from_generic(order),
                llvm::FALSE, // SingleThreaded
            )
        };
        if ret_ptr && self.val_ty(res) != self.type_ptr() {
            res = self.inttoptr(res, self.type_ptr());
        }
        res
    }

    fn atomic_fence(
        &mut self,
        order: rustc_middle::ty::AtomicOrdering,
        scope: SynchronizationScope,
    ) {
        let single_threaded = match scope {
            SynchronizationScope::SingleThread => true,
            SynchronizationScope::CrossThread => false,
        };
        unsafe {
            llvm::LLVMBuildFence(
                self.llbuilder(),
                AtomicOrdering::from_generic(order),
                single_threaded.to_llvm_bool(),
                UNNAMED,
            );
        }
    }

    fn set_invariant_load(&mut self, load: Value<'ll>) {
        self.set_metadata_node(load, llvm::MD_invariant_load, &[]);
    }

    fn lifetime_start(&mut self, ptr: Value<'ll>, size: Size) {
        self.call_lifetime_intrinsic("llvm.lifetime.start", ptr, size);
    }

    fn lifetime_end(&mut self, ptr: Value<'ll>, size: Size) {
        self.call_lifetime_intrinsic("llvm.lifetime.end", ptr, size);
    }

    fn call(
        &mut self,
        llty: Type<'ll>,
        caller_attrs: Option<&CodegenFnAttrs>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
        llfn: Value<'ll>,
        return_slot: ReturnSlot<Value<'ll>>,
        args: &[Value<'ll>],
        funclet: Option<&Funclet<'ll>>,
        callee_instance: Option<Instance<'tcx>>,
    ) -> Value<'ll> {
        // If this function returns indirectly (`PassMode::Indirect`),
        // the `return_slot` should be the first argument.
        let args = match return_slot {
            ReturnSlot::Direct => args.to_vec(),
            ReturnSlot::Indirect(sret_ptr) => {
                let mut args = args.to_vec();
                args.insert(0, sret_ptr);
                args
            }
        };
        debug!("call {:?} with args ({:?})", llfn, args);
        let args = self.check_call("call", llty, llfn, &args);
        let funclet_bundle = funclet.map(|funclet| funclet.bundle());
        let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
        if let Some(funclet_bundle) = funclet_bundle {
            bundles.push(funclet_bundle);
        }

        // Emit CFI pointer type membership test
        self.cfi_type_test(caller_attrs, fn_abi, callee_instance, llfn);

        // Emit KCFI operand bundle
        let kcfi_bundle = self.kcfi_operand_bundle(caller_attrs, fn_abi, callee_instance, llfn);
        if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.as_ref()) {
            bundles.push(kcfi_bundle);
        }

        let pauth = self.ptrauth_operand_bundle(llfn, fn_abi);
        if let Some(p) = pauth.as_ref().map(|b| b.as_ref()) {
            bundles.push(p);
        }

        let call = unsafe {
            llvm::LLVMBuildCallWithOperandBundles(
                self.llbuilder(),
                llty,
                llfn,
                args.as_ptr(),
                args.len() as c_uint,
                bundles.as_ptr(),
                bundles.len() as c_uint,
                c"".as_ptr(),
            )
        };

        if let Some(fn_abi) = fn_abi {
            fn_abi.apply_attrs_callsite(self, call);
        }
        call
    }

    fn tail_call(
        &mut self,
        llty: Self::Type,
        caller_attrs: Option<&CodegenFnAttrs>,
        fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
        llfn: Self::Value,
        return_slot: ReturnSlot<Self::Value>,
        args: &[Self::Value],
        funclet: Option<&Self::Funclet>,
        callee_instance: Option<Instance<'tcx>>,
    ) {
        let call = self.call(
            llty,
            caller_attrs,
            Some(fn_abi),
            llfn,
            return_slot,
            args,
            funclet,
            callee_instance,
        );
        llvm::LLVMSetTailCallKind(call, llvm::TailCallKind::MustTail);

        match &fn_abi.ret.mode {
            PassMode::Ignore | PassMode::Indirect { .. } => self.ret_void(),
            PassMode::Direct(_) | PassMode::Pair { .. } | PassMode::Cast { .. } => self.ret(call),
            PassMode::IndirectUnsized { .. } => bug!("unsized returns are not supported"),
        }
    }

    fn zext(&mut self, val: Value<'ll>, dest_ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildZExt(self.llbuilder(), val, dest_ty, UNNAMED) }
    }

    fn apply_attrs_to_cleanup_callsite(&mut self, llret: Value<'ll>) {
        // Cleanup is always the cold path.
        let cold_inline = llvm::AttributeKind::Cold.create_attr(self.llcx);
        attributes::apply_to_callsite(llret, llvm::AttributePlace::Function, &[cold_inline]);
    }
}

impl<'ll> StaticBuilderMethods for Builder<'_, 'll, '_> {
    fn get_static(&mut self, def_id: DefId) -> Value<'ll> {
        // Forward to the `get_static` method of `CodegenCx`
        let global = self.cx().get_static(def_id);
        if self.cx().tcx.is_thread_local_static(def_id) {
            let pointer =
                self.call_intrinsic("llvm.threadlocal.address", &[self.val_ty(global)], &[global]);
            // Cast to default address space if globals are in a different addrspace
            self.pointercast(pointer, self.type_ptr())
        } else {
            // Cast to default address space if globals are in a different addrspace
            self.cx().const_pointercast(global, self.type_ptr())
        }
    }
}

impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
    pub(crate) fn llfn(&self) -> Value<'ll> {
        unsafe { llvm::LLVMGetBasicBlockParent(self.llbb()) }
    }

    fn generate_ubsan_cfi_diag_data(
        &mut self,
        span: rustc_span::Span,
        expected_ty: String,
        check_kind: u8,
    ) -> Value<'ll> {
        let cx = self.cx();
        let tcx = cx.tcx;

        let loc = tcx.sess.source_map().lookup_char_pos(span.lo());

        let filename_str = format!("{}\0", loc.file.name.prefer_local_unconditionally());
        let filename_val = cx.const_bytes(filename_str.as_bytes());
        let filename_ptr = cx.static_addr_of_impl(filename_val, Align::ONE, None);

        // SourceLocation UBSan struct: { const char *filename, uint32_t line, uint32_t column }
        let source_location = cx.const_struct(
            &[
                filename_ptr,
                cx.const_u32(loc.line as u32),
                // UBSan columns are 1-based
                cx.const_u32(loc.col.0 as u32 + 1),
            ],
            false, // packed = false
        );

        let ty_name = format!("{}\0", expected_ty);
        let ty_name_val = cx.const_bytes(ty_name.as_bytes());

        // TypeDescriptor UBSan struct: { uint16_t TypeKind, uint16_t TypeInfo, const char *TypeName }
        let type_descriptor =
            cx.const_struct(&[cx.const_i16(0xffffu16 as i16), cx.const_i16(0), ty_name_val], false);

        let type_descriptor_ptr =
            cx.static_addr_of_impl(type_descriptor, Align::from_bytes(2).unwrap(), None);

        // CFICheckFailData UBSan struct: { uint8_t CheckKind, SourceLocation Loc, TypeDescriptor *Type }
        let cfi_check_fail_data = cx
            .const_struct(&[cx.const_u8(check_kind), source_location, type_descriptor_ptr], false);
        let align = tcx.data_layout.aggregate_align;

        // Returns the final opaque pointer to the struct to be passed to __ubsan_handle_cfi_check_fail
        cx.static_addr_of_mut(cfi_check_fail_data, align, Some("__ubsan_cfi_check_fail_data"))
    }
}

impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
    fn position_at_start(&mut self, llbb: &'ll BasicBlock) {
        unsafe {
            llvm::LLVMRustPositionBuilderAtStart(self.llbuilder(), llbb);
        }
    }
}
impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
    fn align_metadata(&mut self, load: Value<'ll>, align: Align) {
        let md = [llvm::LLVMValueAsMetadata(self.cx.const_u64(align.bytes()))];
        self.set_metadata_node(load, llvm::MD_align, &md);
    }

    fn noundef_metadata(&mut self, load: Value<'ll>) {
        self.set_metadata_node(load, llvm::MD_noundef, &[]);
    }

    pub(crate) fn set_unpredictable(&mut self, inst: Value<'ll>) {
        self.set_metadata_node(inst, llvm::MD_unpredictable, &[]);
    }

    fn write_operand_repeatedly_optimized(
        &mut self,
        cg_elem: OperandRef<'tcx, Value<'ll>>,
        count: u64,
        dest: PlaceRef<'tcx, Value<'ll>>,
    ) {
        let zero = self.const_usize(0);
        let count = self.const_usize(count);

        let header_bb = self.append_sibling_block("repeat_loop_header");
        let body_bb = self.append_sibling_block("repeat_loop_body");
        let next_bb = self.append_sibling_block("repeat_loop_next");

        self.br(header_bb);

        let mut header_bx = Self::build(self.cx, header_bb);
        let i = header_bx.phi(self.val_ty(zero), &[zero], &[self.llbb()]);

        let keep_going = header_bx.icmp(IntPredicate::IntULT, i, count);
        header_bx.cond_br(keep_going, body_bb, next_bb);

        let mut body_bx = Self::build(self.cx, body_bb);
        let dest_elem = dest.project_index(&mut body_bx, i);
        cg_elem.val.store(&mut body_bx, dest_elem);

        let next = body_bx.unchecked_uadd(i, self.const_usize(1));
        body_bx.br(header_bb);
        header_bx.add_incoming_to_phi(i, next, body_bb);

        *self = Self::build(self.cx, next_bb);
    }

    fn write_operand_repeatedly_unoptimized(
        &mut self,
        cg_elem: OperandRef<'tcx, Value<'ll>>,
        count: u64,
        dest: PlaceRef<'tcx, Value<'ll>>,
    ) {
        let zero = self.const_usize(0);
        let count = self.const_usize(count);
        let start = dest.project_index(self, zero).val.llval;
        let end = dest.project_index(self, count).val.llval;

        let header_bb = self.append_sibling_block("repeat_loop_header");
        let body_bb = self.append_sibling_block("repeat_loop_body");
        let next_bb = self.append_sibling_block("repeat_loop_next");

        self.br(header_bb);

        let mut header_bx = Self::build(self.cx, header_bb);
        let current = header_bx.phi(self.val_ty(start), &[start], &[self.llbb()]);

        let keep_going = header_bx.icmp(IntPredicate::IntNE, current, end);
        header_bx.cond_br(keep_going, body_bb, next_bb);

        let mut body_bx = Self::build(self.cx, body_bb);
        let align = dest.val.align.restrict_for_offset(dest.layout.field(self.cx(), 0).size);
        cg_elem
            .val
            .store(&mut body_bx, PlaceRef::new_sized_aligned(current, cg_elem.layout, align));

        let next = body_bx.inbounds_ptradd(current, self.const_usize(cg_elem.layout.size.bytes()));
        body_bx.br(header_bb);
        header_bx.add_incoming_to_phi(current, next, body_bb);

        *self = Self::build(self.cx, next_bb);
    }

    pub(crate) fn minimum_number_nsz(&mut self, lhs: Value<'ll>, rhs: Value<'ll>) -> Value<'ll> {
        let call = self.call_intrinsic("llvm.minimumnum", &[self.val_ty(lhs)], &[lhs, rhs]);
        unsafe { llvm::LLVMRustSetNoSignedZeros(call) };
        call
    }

    pub(crate) fn maximum_number_nsz(&mut self, lhs: Value<'ll>, rhs: Value<'ll>) -> Value<'ll> {
        let call = self.call_intrinsic("llvm.maximumnum", &[self.val_ty(lhs)], &[lhs, rhs]);
        unsafe { llvm::LLVMRustSetNoSignedZeros(call) };
        call
    }

    pub(crate) fn insert_element(
        &mut self,
        vec: Value<'ll>,
        elt: Value<'ll>,
        idx: Value<'ll>,
    ) -> Value<'ll> {
        unsafe { llvm::LLVMBuildInsertElement(self.llbuilder(), vec, elt, idx, UNNAMED) }
    }

    pub(crate) fn shuffle_vector(
        &mut self,
        v1: Value<'ll>,
        v2: Value<'ll>,
        mask: Value<'ll>,
    ) -> Value<'ll> {
        unsafe { llvm::LLVMBuildShuffleVector(self.llbuilder(), v1, v2, mask, UNNAMED) }
    }

    pub(crate) fn vector_reduce_fadd(&mut self, acc: Value<'ll>, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.fadd", &[self.val_ty(src)], &[acc, src])
    }
    pub(crate) fn vector_reduce_fmul(&mut self, acc: Value<'ll>, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.fmul", &[self.val_ty(src)], &[acc, src])
    }
    pub(crate) fn vector_reduce_fadd_reassoc(
        &mut self,
        acc: Value<'ll>,
        src: Value<'ll>,
    ) -> Value<'ll> {
        unsafe {
            let instr =
                self.call_intrinsic("llvm.vector.reduce.fadd", &[self.val_ty(src)], &[acc, src]);
            llvm::LLVMRustSetAllowReassoc(instr);
            instr
        }
    }
    pub(crate) fn vector_reduce_fmul_reassoc(
        &mut self,
        acc: Value<'ll>,
        src: Value<'ll>,
    ) -> Value<'ll> {
        unsafe {
            let instr =
                self.call_intrinsic("llvm.vector.reduce.fmul", &[self.val_ty(src)], &[acc, src]);
            llvm::LLVMRustSetAllowReassoc(instr);
            instr
        }
    }
    pub(crate) fn vector_reduce_add(&mut self, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.add", &[self.val_ty(src)], &[src])
    }
    pub(crate) fn vector_reduce_mul(&mut self, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.mul", &[self.val_ty(src)], &[src])
    }
    pub(crate) fn vector_reduce_and(&mut self, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.and", &[self.val_ty(src)], &[src])
    }
    pub(crate) fn vector_reduce_or(&mut self, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.or", &[self.val_ty(src)], &[src])
    }
    pub(crate) fn vector_reduce_xor(&mut self, src: Value<'ll>) -> Value<'ll> {
        self.call_intrinsic("llvm.vector.reduce.xor", &[self.val_ty(src)], &[src])
    }
    pub(crate) fn vector_reduce_min(&mut self, src: Value<'ll>, is_signed: bool) -> Value<'ll> {
        self.call_intrinsic(
            if is_signed { "llvm.vector.reduce.smin" } else { "llvm.vector.reduce.umin" },
            &[self.val_ty(src)],
            &[src],
        )
    }
    pub(crate) fn vector_reduce_max(&mut self, src: Value<'ll>, is_signed: bool) -> Value<'ll> {
        self.call_intrinsic(
            if is_signed { "llvm.vector.reduce.smax" } else { "llvm.vector.reduce.umax" },
            &[self.val_ty(src)],
            &[src],
        )
    }
}
impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
    pub(crate) fn add_clause(&mut self, landing_pad: Value<'ll>, clause: Value<'ll>) {
        unsafe {
            llvm::LLVMAddClause(landing_pad, clause);
        }
    }

    pub(crate) fn catch_ret(
        &mut self,
        funclet: &Funclet<'ll>,
        unwind: &'ll BasicBlock,
    ) -> Value<'ll> {
        let ret =
            unsafe { llvm::LLVMBuildCatchRet(self.llbuilder(), funclet.cleanuppad(), unwind) };
        ret.expect("LLVM does not have support for catchret")
    }

    pub(crate) fn check_call<'b>(
        &mut self,
        typ: &str,
        fn_ty: Type<'ll>,
        llfn: Value<'ll>,
        args: &'b [Value<'ll>],
    ) -> Cow<'b, [Value<'ll>]> {
        assert!(
            self.cx.type_kind(fn_ty) == TypeKind::Function,
            "builder::{typ} not passed a function, but {fn_ty:?}"
        );

        let param_tys = self.cx.func_params_types(fn_ty);

        let all_args_match = iter::zip(&param_tys, args.iter().map(|&v| self.cx.val_ty(v)))
            .all(|(expected_ty, actual_ty)| *expected_ty == actual_ty);

        if all_args_match {
            return Cow::Borrowed(args);
        }

        let casted_args: Vec<_> = iter::zip(param_tys, args)
            .enumerate()
            .map(|(i, (expected_ty, &actual_val))| {
                let actual_ty = self.cx.val_ty(actual_val);
                if expected_ty != actual_ty {
                    debug!(
                        "type mismatch in function call of {:?}. \
                            Expected {:?} for param {}, got {:?}; injecting bitcast",
                        llfn, expected_ty, i, actual_ty
                    );
                    self.bitcast(actual_val, expected_ty)
                } else {
                    actual_val
                }
            })
            .collect();

        Cow::Owned(casted_args)
    }

    pub(crate) fn va_arg(&mut self, list: Value<'ll>, ty: Type<'ll>) -> Value<'ll> {
        unsafe { llvm::LLVMBuildVAArg(self.llbuilder(), list, ty, UNNAMED) }
    }
}

impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
    pub(crate) fn call_intrinsic(
        &mut self,
        base_name: impl Into<Cow<'static, str>>,
        type_params: &[Type<'ll>],
        args: &[Value<'ll>],
    ) -> Value<'ll> {
        let (ty, f) = self.cx.get_intrinsic(base_name.into(), type_params);
        // No LLVM intrinsic returns its data indirectly (via `sret`).
        self.call(ty, None, None, f, ReturnSlot::Direct, args, None, None)
    }

    fn call_lifetime_intrinsic(&mut self, intrinsic: &'static str, ptr: Value<'ll>, size: Size) {
        let size = size.bytes();
        if size == 0 {
            return;
        }

        if !self.cx().sess().emit_lifetime_markers() {
            return;
        }

        // LLVM 22 requires the lifetime intrinsic to act directly on the alloca,
        // there can't be an addrspacecast in between.
        let ptr = unsafe { llvm::LLVMRustStripPointerCasts(ptr) };
        self.call_intrinsic(intrinsic, &[self.val_ty(ptr)], &[ptr]);
    }
}
impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> {
    pub(crate) fn phi(
        &mut self,
        ty: Type<'ll>,
        vals: &[Value<'ll>],
        bbs: &[&'ll BasicBlock],
    ) -> Value<'ll> {
        assert_eq!(vals.len(), bbs.len());
        let phi = unsafe { llvm::LLVMBuildPhi(self.llbuilder(), ty, UNNAMED) };
        unsafe {
            llvm::LLVMAddIncoming(phi, vals.as_ptr(), bbs.as_ptr(), vals.len() as c_uint);
            phi
        }
    }

    fn add_incoming_to_phi(&mut self, phi: Value<'ll>, val: Value<'ll>, bb: &'ll BasicBlock) {
        unsafe {
            llvm::LLVMAddIncoming(phi, &val, &bb, 1 as c_uint);
        }
    }
}
impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
    pub(crate) fn landing_pad(
        &mut self,
        ty: Type<'ll>,
        pers_fn: Value<'ll>,
        num_clauses: usize,
    ) -> Value<'ll> {
        // Use LLVMSetPersonalityFn to set the personality. It supports arbitrary Consts while,
        // LLVMBuildLandingPad requires the argument to be a Function (as of LLVM 12). The
        // personality lives on the parent function anyway.
        self.set_personality_fn(pers_fn);
        unsafe {
            llvm::LLVMBuildLandingPad(self.llbuilder(), ty, None, num_clauses as c_uint, UNNAMED)
        }
    }

    pub(crate) fn callbr(
        &mut self,
        llty: Type<'ll>,
        fn_attrs: Option<&CodegenFnAttrs>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
        llfn: Value<'ll>,
        args: &[Value<'ll>],
        default_dest: &'ll BasicBlock,
        indirect_dest: &[&'ll BasicBlock],
        funclet: Option<&Funclet<'ll>>,
        instance: Option<Instance<'tcx>>,
    ) -> Value<'ll> {
        debug!("invoke {:?} with args ({:?})", llfn, args);

        let args = self.check_call("callbr", llty, llfn, args);
        let funclet_bundle = funclet.map(|funclet| funclet.bundle());
        let mut bundles: SmallVec<[_; 2]> = SmallVec::new();
        if let Some(funclet_bundle) = funclet_bundle {
            bundles.push(funclet_bundle);
        }

        // Emit CFI pointer type membership test
        self.cfi_type_test(fn_attrs, fn_abi, instance, llfn);

        // Emit KCFI operand bundle
        let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn);
        if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|bundle| bundle.as_ref()) {
            bundles.push(kcfi_bundle);
        }

        let pauth = self.ptrauth_operand_bundle(llfn, fn_abi);
        if let Some(p) = pauth.as_ref().map(|b| b.as_ref()) {
            bundles.push(p);
        }

        let callbr = unsafe {
            llvm::LLVMBuildCallBr(
                self.llbuilder(),
                llty,
                llfn,
                default_dest,
                indirect_dest.as_ptr(),
                indirect_dest.len() as c_uint,
                args.as_ptr(),
                args.len() as c_uint,
                bundles.as_ptr(),
                bundles.len() as c_uint,
                UNNAMED,
            )
        };
        if let Some(fn_abi) = fn_abi {
            fn_abi.apply_attrs_callsite(self, callbr);
        }
        callbr
    }

    // Emits CFI pointer type membership tests.
    fn cfi_type_test(
        &mut self,
        fn_attrs: Option<&CodegenFnAttrs>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
        instance: Option<Instance<'tcx>>,
        llfn: Value<'ll>,
    ) {
        let is_indirect_call = unsafe { llvm::LLVMRustIsNonGVFunctionPointerTy(llfn) };
        if self.tcx.sess.is_sanitizer_cfi_enabled()
            && let Some(fn_abi) = fn_abi
            && is_indirect_call
        {
            if let Some(fn_attrs) = fn_attrs
                && fn_attrs.sanitizers.disabled.contains(SanitizerSet::CFI)
            {
                return;
            }
            if crate::llvm::HasStringAttribute(self.llfn(), "no-sanitize-cfi") {
                return;
            }

            let mut options = cfi::TypeIdOptions::empty();
            if self.tcx.sess.is_sanitizer_cfi_generalize_pointers_enabled() {
                options.insert(cfi::TypeIdOptions::GENERALIZE_POINTERS);
            }
            if self.tcx.sess.is_sanitizer_cfi_normalize_integers_enabled() {
                options.insert(cfi::TypeIdOptions::NORMALIZE_INTEGERS);
            }

            if self.cx.is_sanitizer_type_ignored(c"cfi", fn_abi) {
                return;
            }

            let typeid = if let Some(instance) = instance {
                cfi::typeid_for_instance(self.tcx, instance, options)
            } else {
                cfi::typeid_for_fnabi(self.tcx, fn_abi, options)
            };
            let typeid_metadata = self.cx.create_metadata(typeid.as_bytes());
            let dbg_loc = self.get_dbg_loc();

            // Test whether the function pointer is associated with the type identifier using the
            // llvm.type.test intrinsic. The LowerTypeTests link-time optimization pass replaces
            // calls to this intrinsic with code to test type membership.
            let typeid = self.get_metadata_value(typeid_metadata);
            let cond = self.call_intrinsic("llvm.type.test", &[], &[llfn, typeid]);
            let bb_pass = self.append_sibling_block("type_test.pass");
            let bb_fail = self.append_sibling_block("type_test.fail");
            self.cond_br(cond, bb_pass, bb_fail);

            self.switch_to_block(bb_fail);
            if let Some(dbg_loc) = dbg_loc {
                self.set_dbg_loc(dbg_loc);
            }

            let is_diag = self.tcx.sess.opts.unstable_opts.sanitizer_cfi_diag.unwrap_or(false);
            let is_recover =
                self.tcx.sess.opts.unstable_opts.sanitizer_cfi_recover.unwrap_or(false);
            let is_minimal =
                self.tcx.sess.opts.unstable_opts.sanitizer_cfi_minimal_runtime.unwrap_or(false);

            if is_diag || is_recover {
                let fty = if is_minimal {
                    self.cx.type_func(&[], self.cx.type_void())
                } else {
                    self.cx.type_func(
                        &[self.cx.type_ptr(), self.cx.type_isize(), self.cx.type_isize()],
                        self.cx.type_void(),
                    )
                };
                let ubsan_handler = self.declare_cfn(
                    match (is_minimal, is_recover) {
                        (true, true) => "__ubsan_handle_cfi_check_fail_minimal",
                        (true, false) => "__ubsan_handle_cfi_check_fail_minimal_abort",
                        (false, true) => "__ubsan_handle_cfi_check_fail",
                        (false, false) => "__ubsan_handle_cfi_check_fail_abort",
                    },
                    llvm::UnnamedAddr::Global,
                    fty,
                );

                let mut expected_ty = String::from("fn(");
                ty::print::with_no_trimmed_paths!({
                    for (i, arg) in fn_abi.args.iter().enumerate() {
                        if i > 0 {
                            expected_ty.push_str(", ");
                        }
                        use std::fmt::Write;
                        write!(&mut expected_ty, "{}", arg.layout.ty).unwrap();
                    }
                    expected_ty.push(')');
                    if !fn_abi.ret.layout.ty.is_unit() {
                        use std::fmt::Write;
                        write!(&mut expected_ty, " -> {}", fn_abi.ret.layout.ty).unwrap();
                    }
                });

                // 4 for cfi-icall (indirect call)
                let check_kind = 4;
                let diag_data =
                    self.generate_ubsan_cfi_diag_data(self.span, expected_ty, check_kind);

                let function_address = self.ptrtoint(llfn, self.cx.type_isize());
                let arguments: &[_] = if is_minimal {
                    &[]
                } else {
                    &[diag_data, function_address, self.const_usize(0)]
                };
                self.call(
                    fty,
                    None,
                    None,
                    ubsan_handler,
                    ReturnSlot::Direct,
                    arguments,
                    None,
                    None,
                );
                if is_recover {
                    self.br(bb_pass);
                } else {
                    self.unreachable();
                }
            } else {
                self.abort_immediate();
                self.unreachable();
            }

            self.switch_to_block(bb_pass);
            if let Some(dbg_loc) = dbg_loc {
                self.set_dbg_loc(dbg_loc);
            }
        }
    }

    // Emits KCFI operand bundles.
    fn kcfi_operand_bundle(
        &mut self,
        fn_attrs: Option<&CodegenFnAttrs>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
        instance: Option<Instance<'tcx>>,
        llfn: Value<'ll>,
    ) -> Option<llvm::OperandBundleBox<'ll>> {
        let is_indirect_call = unsafe { llvm::LLVMRustIsNonGVFunctionPointerTy(llfn) };
        let kcfi_bundle = if self.tcx.sess.is_sanitizer_kcfi_enabled()
            && let Some(fn_abi) = fn_abi
            && is_indirect_call
        {
            if let Some(fn_attrs) = fn_attrs
                && fn_attrs.sanitizers.disabled.contains(SanitizerSet::KCFI)
            {
                return None;
            }
            if crate::llvm::HasStringAttribute(self.llfn(), "no-sanitize-kcfi") {
                return None;
            }

            let mut options = kcfi::TypeIdOptions::empty();
            if self.tcx.sess.is_sanitizer_cfi_generalize_pointers_enabled() {
                options.insert(kcfi::TypeIdOptions::GENERALIZE_POINTERS);
            }
            if self.tcx.sess.is_sanitizer_cfi_normalize_integers_enabled() {
                options.insert(kcfi::TypeIdOptions::NORMALIZE_INTEGERS);
            }

            if self.cx.is_sanitizer_type_ignored(c"kcfi", fn_abi) {
                return None;
            }

            let kcfi_typeid = if let Some(instance) = instance {
                kcfi::typeid_for_instance(self.tcx, instance, options)
            } else {
                kcfi::typeid_for_fnabi(self.tcx, fn_abi, options)
            };

            Some(llvm::OperandBundleBox::new("kcfi", &[self.const_u32(kcfi_typeid)]))
        } else {
            None
        };
        kcfi_bundle
    }

    // Emits pauth operand bundle.
    fn ptrauth_operand_bundle(
        &mut self,
        llfn: Value<'ll>,
        fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>,
    ) -> Option<llvm::OperandBundleBox<'ll>> {
        let schema = self.sess().pointer_authentication_functions()?;

        // Pointer authentication support is currently limited to extern "C" calls; filter out other
        // ABIs.
        if fn_abi?.conv != CanonAbi::C {
            return None;
        }
        // Filter out LLVM intrinsics.
        if llvm::get_value_name(llfn).starts_with(b"llvm.") {
            return None;
        }

        // FIXME(jchlanda) Operand bundles should only be attached to indirect function calls.
        // However, function pointer signing is currently performed in `get_fn_addr`, which causes
        // the logic to be applied too broadly, including to function values (not just pointers).
        // As a result, direct calls using signed function values must also receive operand
        // bundles.
        // Once this is resolved, we should analyze each call and skip direct calls. See the
        // discussion in the rust-lang issue: <https://github.com/rust-lang/rust/issues/152532>
        Some(llvm::OperandBundleBox::new(
            "ptrauth",
            &[
                self.const_u32(schema.key as u32),
                self.const_u64(schema.constant_discriminator as u64),
            ],
        ))
    }

    /// Emits a call to `llvm.instrprof.increment`. Used by coverage instrumentation.
    #[instrument(level = "debug", skip(self))]
    pub(crate) fn instrprof_increment(
        &mut self,
        fn_name: Value<'ll>,
        hash: Value<'ll>,
        num_counters: Value<'ll>,
        index: Value<'ll>,
    ) {
        self.call_intrinsic("llvm.instrprof.increment", &[], &[fn_name, hash, num_counters, index]);
    }
}