rucc_ir/

opcode.rs

//! The instruction set.
//!
//! Design: `spec/08-ir.md` section 8.3.
//!
//! The set is small enough to enumerate and it is closed. Adding an opcode is a spec change,
//! because the verifier, the printer, the parser, the rewrite rules and the lowering all have
//! to learn it, and an opcode that only half of them know about is a silent miscompilation
//! waiting for the right input.
//!
//! Two things are deliberately absent. There is no `getelementptr`: pointer arithmetic is
//! [`Opcode::PtrAdd`] over a byte offset the frontend computed, because C never needs the
//! multi-index form and its absence removes a well known source of complexity. And there is no
//! `phi`: values arriving at a block are the block's parameters, passed by the branch, so
//! there is no operand list positionally tied to a predecessor list kept somewhere else.

use std::fmt;

/// One instruction of the IR.
///
/// The names are the textual form exactly, so [`Opcode::name`] and [`Opcode::from_name`] are
/// what the printer and the parser use, and neither carries a table of its own that could
/// drift from this one.
///
/// The enum is not `non_exhaustive`, deliberately. The set is closed, so a pass that matches
/// on every opcode should stop compiling when one is added rather than fall into a wildcard
/// arm that quietly does the wrong thing.
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub enum Opcode {
    // Constants. A constant is an instruction rather than an operand kind, so that every
    // operand is a value and every value has one definition, which is what makes the
    // dominance check in the verifier a single rule rather than a rule with exceptions.
    /// An integer constant, `iconst.i32 7`.
    IConst,
    /// A floating point constant, `fconst.f64 0x1.8p+1`.
    FConst,
    /// A vector constant with every lane the same, `splat.i8x16 0`.
    Splat,
    /// The address of a global or a function, `global_addr @counter`.
    GlobalAddr,
    /// The address of a block in this function, `block_addr block3`.
    ///
    /// The one instruction that names a block without being a branch, which is what GNU's
    /// `&&label` is. Where it goes is [`Opcode::IndirectBr`], and the two are only useful
    /// together: an address on its own is a number that nothing can do anything with.
    BlockAddr,

    // Arithmetic.
    /// Integer addition.
    Add,
    /// Integer subtraction.
    Sub,
    /// Integer multiplication.
    Mul,
    /// Signed division.
    SDiv,
    /// Unsigned division.
    UDiv,
    /// Signed remainder, with the sign of the dividend.
    SRem,
    /// Unsigned remainder.
    URem,
    /// Bitwise and.
    And,
    /// Bitwise or.
    Or,
    /// Bitwise exclusive or.
    Xor,
    /// Shift left.
    Shl,
    /// Logical shift right, shifting in zeroes.
    LShr,
    /// Arithmetic shift right, shifting in the sign bit.
    AShr,
    /// Floating point addition.
    FAdd,
    /// Floating point subtraction.
    FSub,
    /// Floating point multiplication.
    FMul,
    /// Floating point division.
    FDiv,
    /// Floating point remainder.
    FRem,
    /// Floating point negation, which flips the sign bit and is not `0 - x`.
    FNeg,
    /// Fused multiply-add, rounded once.
    Fma,

    // Comparison.
    /// Integer comparison, producing `i1` or a vector of `i1`.
    ICmp,
    /// Floating point comparison, producing `i1` or a vector of `i1`.
    FCmp,

    // Selection.
    /// One of two values, chosen by a bit. `select c, a, b` is `a` when `c` is one.
    ///
    /// This is what control flow becomes when it stops being control flow.
    /// `spec/optimizer/22-phiopt-and-if-conversion.md` section 22.2 makes it the lowering target
    /// for a diamond whose two arms compute a value, and the reason it is an opcode rather than a
    /// pattern is that it is the form the rule set is written against: `select(c, a, a) -> a` and
    /// `select(c, 1, 0) -> zext(c)` are ordinary rules once the shape has a name.
    ///
    /// Both arms are evaluated, which is the whole point and also the whole danger. Whatever
    /// produces one of these owes the argument that evaluating the arm that is not chosen is
    /// harmless, and section 22.6 is the list of ways that argument goes wrong.
    Select,

    // Conversion.
    /// Narrows an integer, discarding the high bits.
    Trunc,
    /// Widens an integer, copying the sign bit.
    SExt,
    /// Widens an integer, filling with zeroes.
    ZExt,
    /// Narrows a floating point value.
    FPTrunc,
    /// Widens a floating point value.
    FPExt,
    /// Floating point to signed integer.
    FPToSI,
    /// Floating point to unsigned integer.
    FPToUI,
    /// Signed integer to floating point.
    SIToFP,
    /// Unsigned integer to floating point.
    UIToFP,
    /// An address to an integer of the same width.
    PtrToInt,
    /// An integer to an address.
    IntToPtr,
    /// A reinterpretation of the same bits at the same width.
    Bitcast,

    // Memory.
    /// Memory as the function found it, which is where a memory SSA chain starts.
    ///
    /// It produces one `mem` and takes nothing, and it belongs at the top of the entry block.
    /// GCC calls the same thing the default definition of `.MEM` and LLVM calls it
    /// `liveOnEntry`. It exists as an instruction rather than as a parameter of the entry block
    /// because the entry block's parameters are the function's parameters and the verifier
    /// checks them against the signature, and memory is not an argument anybody passed.
    MemEntry,
    /// A stack slot. In the entry block, or marked dynamic for a variable length array.
    Alloca,
    /// A read.
    Load,
    /// A write, producing no value.
    Store,
    /// Address arithmetic: an address and a byte offset.
    PtrAdd,
    /// A copy of a known size between addresses that do not overlap.
    Memcpy,
    /// A copy of a known size between addresses that may overlap.
    Memmove,
    /// A fill of a known size with one byte.
    Memset,
    /// An atomic read.
    AtomicLoad,
    /// An atomic write.
    AtomicStore,
    /// An atomic read-modify-write, carrying which operation in [`RmwOp`](crate::RmwOp).
    AtomicRmw,
    /// An atomic compare and exchange, producing the old value and whether it succeeded.
    Cmpxchg,
    /// A memory barrier.
    Fence,

    // Memory safety. Design: `spec/safe-memory/06-instrumentation.md` section 6.2.2. None of
    // these is emitted unless `-fsafety` asked for it, and a function compiled without it
    // contains not one of them.
    /// The capability of a pointer value, taken from the pointer's provenance.
    ///
    /// One operand, the pointer, and one capability out. This is the general question and the other
    /// producers are all the special cases of it that have a cheap answer, which is why it is the
    /// one that always works and the one to reach for last. A pointer whose provenance the compiler
    /// still holds should have got its capability from wherever that provenance came from: from the
    /// allocation site, from the aux slot beside it in memory, from the frame its caller wrote, or
    /// from the derivation that narrowed it. One of these is what is left when none of those did.
    ///
    /// So it has two lowerings and which one it gets is not a property of the instruction. A pointer
    /// the lowering can trace back to an allocation site gets the cheap answer, which is a subtract
    /// and a load off the header the allocator wrote. Anything else gets the plane walk, which is
    /// the expensive answer and the only one that is always available, and comes back marked as
    /// recovered so the summary counts it as the weakening it is.
    ///
    /// An interior pointer is always the second of those, and that is the answer to the question
    /// document 05 section 5.2.3 leaves open rather than a gap in it. The header is behind the
    /// payload so that finding it is a subtract by a constant, which is true of a pointer to the
    /// base of an object and false of a pointer to the middle of one, and the lowering asks for the
    /// base by construction rather than by checking.
    CapOf,
    /// The capability in the auxiliary slot beside a stored pointer, read back.
    ///
    /// A pointer written to memory and read again has to bring its capability with it, and where
    /// the capability lives is document 05's question rather than this one's. What this says is
    /// that a capability comes back from an address.
    ///
    /// Three operands: the capability of the object the word is in, the address of the word, and
    /// the pointer that was loaded from it. The first is what says where the slot is, since the
    /// aux is in front of the object and the arithmetic is over the object's own base and extent,
    /// and the third is what the slot's two numbers are relative to. Both of those are facts about
    /// the representation document 05 section 5.2.2 chose, and an instrumented load has both
    /// values in hand already, since the capability is the one its bounds check used and the
    /// pointer is what the load produced.
    CapLoad,
    /// The other half of [`Opcode::CapLoad`], writing one into the slot beside a pointer.
    ///
    /// Four operands: the capability of the object the word is in, the address of the word, the
    /// pointer being stored there, and that pointer's capability. The first three are the same
    /// three [`Opcode::CapLoad`] takes and for the same reasons, and the fourth is what is being
    /// written down.
    CapStore,
    /// The capability that permits nothing, which is what a null pointer has.
    CapNull,
    /// A capability narrowed to a sub-object of what it covered.
    ///
    /// Only under `-fsafety-subobject`. Narrowing is what catches an overflow from one member of
    /// a struct into the next, and it is separate because C code that walks off the end of a
    /// member on purpose exists and a project has to be able to say so.
    CapNarrow,
    /// The capability for an address that arrived from outside, recovered from the planes.
    CapRecover,
    /// How many bytes from an address on the capability covers, asking for no more than a limit.
    ///
    /// Three operands, the capability, the address and how many bytes the asker wants, and one
    /// integer result that is never more than that limit. It is what
    /// `spec/safe-memory/07-check-elimination.md` section 7.4 needs to split a loop: the checked
    /// part and the unchecked part are divided at `min(n, extent / sizeof(T))`, and the extent is
    /// the half of that a compiler cannot work out on its own.
    ///
    /// The limit is an operand because under milestone S1 answering means walking the lifetime
    /// plane, and a walk that stops at the number of bytes the loop was going to read anyway is
    /// bounded by the work the loop is already doing. An answer smaller than the truth costs
    /// iterations in the checked half and is never wrong, which is what makes stopping early
    /// allowed. Once a capability carries its own bounds, which is milestone S2, this is a
    /// subtraction on the capability and the limit is one `min`.
    CapExtent,
    /// How many bytes below an address on the capability covers, asking for no more than a limit.
    ///
    /// The mirror of [`Opcode::CapExtent`], with the same three operands and the same kind of
    /// answer. What it counts is the bytes ending at the address rather than the bytes starting
    /// there, so an answer of `n` says that `[addr - n, addr)` belongs to one thing. The address
    /// itself is one past what is asked about, which is what a walk from high to low needs: the
    /// question is about where the walk ends up, and where it ends up is below where it began.
    ///
    /// The ownership asked about is the byte below the address rather than the byte at it, since
    /// the address may be one past the end of the object and the object is what the question is
    /// about. Everything else, the limit operand and why an answer short of the truth is allowed,
    /// is [`Opcode::CapExtent`]'s.
    CapExtentBack,
    /// The capabilities of a call's pointer arguments, handed over beside the arguments.
    ///
    /// Document 05 section 5.3 is the whole of why this is an instruction of its own rather than
    /// more operands on the call. An instrumented function's calling convention is unchanged, so a
    /// pointer argument goes in the register it always went in and `sizeof(void *)` is still eight,
    /// which is what lets an object this compiler built link against one nobody instrumented. The
    /// capability therefore travels out of band, in a small per thread frame the caller writes and
    /// the callee reads, indexed by argument position.
    ///
    /// One operand per pointer argument, in the order the call passes them, and no operand for the
    /// arguments that are not pointers. A call handing over more pointers than the frame has room
    /// for describes the ones that fit and the callee recovers the rest, which is a weakening the
    /// summary counts rather than a refusal. How many fit is the runtime's number and
    /// `rucc_safety::frame` is where the compiler keeps it, so nothing here is a bound on the
    /// operand count.
    ///
    /// It goes immediately in front of the call it is about, which is how the two are tied
    /// together, the same way [`Opcode::MetaRelease`] is tied to the atomic it goes in front of.
    CapPublish,
    /// There is no frame for this call, which is what a callee that might not be instrumented gets.
    ///
    /// The other half of [`Opcode::CapPublish`] and not the absence of one. Publishing to a callee
    /// that never reads the frame leaves it in place for whatever that callee calls back into, and
    /// a callback entered from uninstrumented code holding somebody else's capabilities is worse
    /// than one entered holding none. So a call the compiler cannot say reads frames says there is
    /// nothing rather than saying nothing at all, and document 10 section 10.8's callback recovers
    /// its arguments the way an entry from outside always does.
    CapClear,
    /// The capability of a pointer parameter, out of the frame the caller published.
    ///
    /// The reading end of [`Opcode::CapPublish`], in the callee rather than in the caller. Two
    /// operands, the parameter itself and which position it is in the call, and one capability out.
    ///
    /// The parameter is an operand because it is the answer when there is no frame. A function
    /// entered from code this build never compiled finds nothing published, and so does one whose
    /// caller could not vouch for it and said `cap_clear`, and in both cases the capability has to be
    /// worked out from the pointer the way [`Opcode::CapRecover`] does. That is the expensive answer
    /// the frame exists to avoid, it is available whatever the caller did, and it is counted as the
    /// weakening it is.
    ///
    /// The position is an operand rather than a payload for the same reason [`Opcode::CapNarrow`]'s
    /// offset and length are: it is a number the runtime is handed. A position the frame does not
    /// reach is not an error either, since it is the recovery again, so nothing here has to know how
    /// many capabilities a frame holds.
    CapArg,
    /// The capability of the pointer this function is returning, left where its caller reads.
    ///
    /// One operand, the capability, and it goes immediately in front of the `ret` it is about, which
    /// is the tie [`Opcode::CapPublish`] has with its call. A function that returns from several
    /// places has one of these in front of each of them, because what is being said is about the
    /// value leaving by that particular one.
    ///
    /// This is the only thing in the design that writes into a frame somebody else made. It is
    /// allowed to because the frame is the caller's stack and the caller is waiting for this
    /// function to return, and it is needed because a returned pointer is the one value that crosses
    /// a call in the other direction.
    CapYield,
    /// The capability of the pointer a call gave back, out of the frame the call was published with.
    ///
    /// The reading end of [`Opcode::CapYield`], in the caller. One operand, the pointer that came
    /// back, and one capability out. The operand is there for the reason [`Opcode::CapArg`]'s is: a
    /// callee that wrote nothing leaves the bottom capability where the caller put it, and the
    /// bottom one is the signal to work the answer out from the pointer instead.
    ///
    /// It goes immediately after the call it is about, and that call has to be one a `cap_publish`
    /// goes in front of. A call that says there is no frame has no frame for a callee to have
    /// written into, so there would be nothing for this to read.
    CapResult,
    /// An access is within its capability's bounds, aligned, and permitted.
    ///
    /// The size and the alignment are the access's, and they are in the memory payload rather
    /// than in operands because they are what the front end knew and not what the program
    /// computed.
    ///
    /// A third operand overrides how many bytes are asked about, and it exists for the one check
    /// the front end did not write. `spec/safe-memory/07-check-elimination.md` section 7.4 replaces
    /// the checks in a loop that runs `n` times with one check over `n * sizeof(T)` bytes, and that
    /// is a length the program computes rather than one anybody knew when the access was parsed. The
    /// payload still holds the alignment and the type information of the access the check came from,
    /// and its size becomes the size of one of them rather than the size of the question.
    CheckBounds,
    /// The capability's provenance is still live.
    CheckLive,
    /// The access agrees with the type plane, which is the effective type rule of C 6.5.
    CheckType,
    /// The bytes the access reads have been written.
    CheckInit,
    /// A pointer derived from another stays inside the capability the first one had.
    ///
    /// Three operands, because the answer is about the new pointer and the question is about
    /// the old one's capability.
    CheckDeriv,
    /// The metadata this access is about to consult has not been changed under it.
    ///
    /// Judgement J9, which document 09 section 9.5 specifies and which document 04 section 4.5
    /// keeps out of J1 for the reason the `restrict` checks are kept out: it is a statement about
    /// two operations rather than about one. The payload holds the size of the access, because the
    /// question is whether any granule of the range this touches carries a write by another thread.
    CheckRace,
    /// This read did not reach a byte another `restrict` pointer of the same block wrote.
    ///
    /// Judgement J8, which document 09 section 9.6 specifies and which document 04 section 4.6
    /// keeps out of J1 because it is a statement about a pair of accesses rather than about one.
    /// The payload holds the size of the access and the two numbers saying which pointer it went
    /// through, and those are the same two [`crate::Restrict`] carries on the access itself.
    ///
    /// Read and write are two opcodes rather than one with a flag on it, because every bit of
    /// [`crate::Flags`] is spoken for and because the operand would then be a constant the program
    /// does not compute, which is the thing [`Opcode::CheckBounds`] says belongs in the payload.
    CheckRestrictRead,
    /// The same about a write, which is the half of the pair that makes the other half a violation.
    ///
    /// Two accesses that both only read are not a violation of anything, so what the scope records
    /// is which of them wrote and the check refuses a pair only when at least one did.
    CheckRestrictWrite,
    /// A storage instance begins here, over a range, with a class.
    ///
    /// Judgement J4. This is the `alloca` for an automatic instance and the allocator's report
    /// for an allocated one, and the range is a pointer and a length in registers rather than a
    /// payload, because the length of a variable length array is not known when the instruction
    /// is written down.
    MetaBegin,
    /// A storage instance ends here, which is judgement J5.
    ///
    /// Every capability for it fails from this point on and keeps failing after the address is
    /// handed out again, which is what makes the check a use after free check rather than a use
    /// after reallocation one.
    MetaEnd,
    /// The effective type of a range is now this one.
    MetaType,
    /// The effective types of a range are now the ones the range it was copied from had.
    ///
    /// Three operands, because a copy has two ranges and one length: the destination, the source,
    /// and how many bytes moved. A `memcpy` does not store through a type, so there is no type to
    /// name here and naming one would be wrong: C 6.5 says the copied bytes keep the effective type
    /// they had, whatever that was, and the only place that is written down is the plane over the
    /// source.
    MetaTypeCopy,
    /// The bytes of a range are now initialized.
    MetaInit,
    /// The bytes of a range are now initialized wherever the range they were copied from was.
    ///
    /// Three operands, for the reason [`Opcode::MetaTypeCopy`] has three. A copy does not write
    /// values of its own, so whether a destination byte holds anything is whether the byte it came
    /// from did, and the only place that is written down is the plane over the source. That is what
    /// makes a structure filled member by member and then copied whole still have padding nothing
    /// wrote, which is the infoleak the plane is for.
    MetaInitCopy,
    /// This thread wrote a range, at whatever step of its own counting it has reached.
    ///
    /// The epoch plane's only write, from `spec/safe-memory/09-type-init-and-races.md` section 9.5.
    /// Two operands like the other plane writes, and the range is a pointer shaped slot rather than
    /// whatever the access covered: the plane holds one stamp per eight bytes because that is what
    /// a pointer comes in, and a granule two threads share is one holding no pointer.
    ///
    /// It carries no thread and no count. Which thread is running and how far it has counted are
    /// both facts about the moment the program reaches this, so the runtime reads them and nothing
    /// here could name them.
    MetaEpoch,
    /// Everything this thread has done so far is published at the atomic object named here.
    ///
    /// One half of a synchronization edge in the sense of section 9.5, and the half that goes in
    /// front of the atomic that carries it. The operand is the object's address, because that is
    /// the key whoever takes the other end will look the clock up under, and there is no payload:
    /// which thread is publishing and how far it has counted are facts about the moment the program
    /// reaches this, the same way they are for [`Opcode::MetaEpoch`].
    ///
    /// This exists because an atomic is not a call. Every other edge the monitor knows about is a
    /// `pthread` function and is interposed, and a C11 release store is a machine instruction with
    /// nothing to interpose, so the compiler is the only thing that can say the edge was there.
    MetaRelease,
    /// Everything published at the atomic object named here is now ordered before this thread.
    ///
    /// The other half of [`Opcode::MetaRelease`], and it goes after the atomic rather than in front
    /// of it, because the ordering it takes is the ordering the atomic just read.
    ///
    /// A missing edge is the one kind of missing instrumentation in the whole safety pass that
    /// costs a false report rather than a missed one: two threads that really were ordered by an
    /// edge nobody recorded look concurrent, and a race is reported against a program doing nothing
    /// wrong. That is why the edges go in before the race check is ever on by default.
    MetaAcquire,
    /// Everything this thread has done so far is published at no object in particular.
    ///
    /// What a release fence is, and the reason it cannot reuse [`Opcode::MetaRelease`]: a fence
    /// orders against every other thread rather than against one object, so there is no address to
    /// key it on and it takes no operands at all. The relaxed atomic that usually sits beside it in
    /// the source is not the key either, because the fence orders everything, not that one word.
    ///
    /// The runtime pays for that with one cell shared by every fence in the program, which orders
    /// more pairs of threads than the program really ordered. That direction is safe. A thread put
    /// further ahead than it needed to be reports fewer races, never a wrong one.
    MetaFenceRelease,
    /// Everything published at any release fence is now ordered before this thread.
    ///
    /// The other half of [`Opcode::MetaFenceRelease`], after the fence rather than in front of it,
    /// for the same reason [`Opcode::MetaAcquire`] goes after its atomic.
    MetaFenceAcquire,
    /// A range leaves the monitor's authority, or comes back, which is judgement J7.
    MetaTransfer,
    /// A declared exemption starts here, with the reason it was declared.
    ///
    /// Not an optimization hint. Everything between this and its `safe_region_end` is code the
    /// monitor is told not to judge, so the reason it carries is a trust set entry, and
    /// `spec/safe-memory/10-boundaries.md` section 10.2 counts them per build precisely so that
    /// a reviewer can read what a binary's guarantee rests on.
    SafeRegionBegin,
    /// The end of the region the last `safe_region_begin` opened.
    SafeRegionEnd,
    /// A block that declares `restrict` pointers begins here, over a slot to keep its record in.
    ///
    /// The operand is the storage the record lives in, which is the block's own stack slot, and the
    /// payload says how large it is. The clique the block was given is [`crate::Restrict::clique`]
    /// of the payload and how many pointers it declares is [`crate::Restrict::base`], which is the
    /// one place that field counts bases rather than naming one.
    ///
    /// A marker rather than something the front end could fold into the accesses, because the
    /// promise is about the block's dynamic extent: a function called twice has made the promise
    /// twice, and what the second call reached says nothing about the first.
    RestrictEnter,
    /// The block the last `restrict_enter` opened ends here.
    ///
    /// The operand is the same slot, so that the record can be unlinked from whatever encloses it
    /// without the runtime having to keep a list of its own.
    RestrictLeave,

    // Control. Every one of these is a terminator.
    /// An unconditional branch, `jump block1(%a, %b)`.
    Jump,
    /// A two-way branch on an `i1`.
    BrIf,
    /// A multi-way branch on an integer, with a default.
    Switch,
    /// A branch to an address, `indirect_br %0, block1, block2`.
    ///
    /// The targets are every block control can arrive at, which is what makes the edges of a
    /// computed `goto` ordinary edges: nothing else in the compiler has to know that the
    /// address decides which one it is. A target that is not listed is a branch that does not
    /// happen, so a frontend that leaves one out has made a promise on the program's behalf.
    IndirectBr,
    /// A return, with the values the signature says.
    Return,
    /// A place control cannot reach, which the frontend emits after a `noreturn` call.
    Unreachable,

    // Calls.
    /// A call to a named function.
    Call,
    /// A call through an address, carrying the signature it is called with.
    CallIndirect,
    /// A call in tail position that reuses the frame, which is a terminator.
    TailCall,

    // Intrinsics, which is the closed part. The open part is `TargetIntrinsic`.
    /// Count leading zeroes.
    Ctlz,
    /// Count trailing zeroes.
    Cttz,
    /// Count set bits.
    Ctpop,
    /// Reverse the bytes.
    Bswap,
    /// Reverse the bits.
    Bitreverse,
    /// Signed addition, producing the result and whether it overflowed.
    SAddOverflow,
    /// Unsigned addition, producing the result and whether it overflowed.
    UAddOverflow,
    /// Signed subtraction, producing the result and whether it overflowed.
    SSubOverflow,
    /// Unsigned subtraction, producing the result and whether it overflowed.
    USubOverflow,
    /// Signed multiplication, producing the result and whether it overflowed.
    SMulOverflow,
    /// Unsigned multiplication, producing the result and whether it overflowed.
    UMulOverflow,
    /// `__builtin_expect`, which is the value with a hint attached.
    Expect,
    /// `__builtin_unreachable` as a hint on a path, distinct from the terminator.
    UnreachableHint,
    /// `__builtin_trap`, which stops the program where it stands.
    ///
    /// Not a terminator, for the reason `unreachable_hint` is not one: what ends a block here is
    /// control going somewhere, and this goes nowhere at all. The block it is in goes on being
    /// lowered and whatever follows it is written and never run, which costs a few bytes nothing
    /// reaches and keeps every pass that walks a block from needing a second shape for it.
    Trap,
    /// `__builtin_prefetch`.
    Prefetch,
    /// `__builtin_frame_address`.
    FrameAddress,
    /// `__builtin_return_address`.
    ReturnAddress,
    /// `__builtin_thread_pointer`, the address of the storage the running thread has.
    ///
    /// It takes nothing and answers a pointer. Unlike the two above it there is no walk to do and
    /// no frame to have kept: the machine holds the address in a place of its own, so this is one
    /// instruction on every target that has the builtin at all.
    ThreadPointer,
    /// The start of a variable argument list.
    VaStart,
    /// One argument off a variable argument list, which moves the list on as it reads it. Two
    /// of these on one list are two arguments and never one argument read twice, so whatever
    /// decides which instructions may be folded together has to leave these alone.
    VaArg,
    /// One argument off a variable argument list, when that argument is an object rather than a
    /// value, which is what a `struct` or a `union` read out of one is.
    ///
    /// It answers the address of the object rather than the object, because an aggregate is not
    /// a value and there is nothing for one result to be. Where the object arrives in registers
    /// there is no address until something makes one, so what this asks of a target is a place
    /// to put the registers and the address of that place, which is the copy every psABI's own
    /// description of the algorithm makes. It moves the list on for the reason [`Opcode::VaArg`]
    /// does.
    VaObject,
    /// The end of a variable argument list.
    VaEnd,
    /// A copy of a variable argument list.
    VaCopy,
    /// The stack pointer, saved before a variable length array.
    StackSave,
    /// The stack pointer, restored after one.
    StackRestore,
    /// The marker a `setjmp` leaves, which pins everything live across it.
    SetjmpMarker,
    /// The marker a `longjmp` leaves.
    LongjmpMarker,
    /// A target-specific intrinsic, named rather than enumerated, for the vector builtins.
    TargetIntrinsic,

    /// Inline assembly. A terminator when it has labels, which is `asm goto`.
    InlineAsm,
}

impl Opcode {
    /// The textual form, which is also what the parser reads.
    #[must_use]
    pub const fn name(self) -> &'static str {
        match self {
            Self::IConst => "iconst",
            Self::FConst => "fconst",
            Self::Splat => "splat",
            Self::GlobalAddr => "global_addr",
            Self::BlockAddr => "block_addr",
            Self::Add => "add",
            Self::Sub => "sub",
            Self::Mul => "mul",
            Self::SDiv => "sdiv",
            Self::UDiv => "udiv",
            Self::SRem => "srem",
            Self::URem => "urem",
            Self::And => "and",
            Self::Or => "or",
            Self::Xor => "xor",
            Self::Shl => "shl",
            Self::LShr => "lshr",
            Self::AShr => "ashr",
            Self::FAdd => "fadd",
            Self::FSub => "fsub",
            Self::FMul => "fmul",
            Self::FDiv => "fdiv",
            Self::FRem => "frem",
            Self::FNeg => "fneg",
            Self::Fma => "fma",
            Self::ICmp => "icmp",
            Self::FCmp => "fcmp",
            Self::Select => "select",
            Self::Trunc => "trunc",
            Self::SExt => "sext",
            Self::ZExt => "zext",
            Self::FPTrunc => "fptrunc",
            Self::FPExt => "fpext",
            Self::FPToSI => "fptosi",
            Self::FPToUI => "fptoui",
            Self::SIToFP => "sitofp",
            Self::UIToFP => "uitofp",
            Self::PtrToInt => "ptrtoint",
            Self::IntToPtr => "inttoptr",
            Self::Bitcast => "bitcast",
            Self::MemEntry => "mem_entry",
            Self::Alloca => "alloca",
            Self::Load => "load",
            Self::Store => "store",
            Self::PtrAdd => "ptr_add",
            Self::Memcpy => "memcpy",
            Self::Memmove => "memmove",
            Self::Memset => "memset",
            Self::AtomicLoad => "atomic_load",
            Self::AtomicStore => "atomic_store",
            Self::AtomicRmw => "atomic_rmw",
            Self::Cmpxchg => "cmpxchg",
            Self::Fence => "fence",
            Self::CapOf => "cap_of",
            Self::CapLoad => "cap_load",
            Self::CapStore => "cap_store",
            Self::CapNull => "cap_null",
            Self::CapNarrow => "cap_narrow",
            Self::CapRecover => "cap_recover",
            Self::CapExtent => "cap_extent",
            Self::CapExtentBack => "cap_extent_back",
            Self::CapPublish => "cap_publish",
            Self::CapClear => "cap_clear",
            Self::CapArg => "cap_arg",
            Self::CapYield => "cap_yield",
            Self::CapResult => "cap_result",
            Self::CheckBounds => "check_bounds",
            Self::CheckLive => "check_live",
            Self::CheckType => "check_type",
            Self::CheckInit => "check_init",
            Self::CheckDeriv => "check_deriv",
            Self::CheckRace => "check_race",
            Self::CheckRestrictRead => "check_restrict_read",
            Self::CheckRestrictWrite => "check_restrict_write",
            Self::MetaBegin => "meta_begin",
            Self::MetaEnd => "meta_end",
            Self::MetaType => "meta_type",
            Self::MetaTypeCopy => "meta_type_copy",
            Self::MetaInit => "meta_init",
            Self::MetaInitCopy => "meta_init_copy",
            Self::MetaEpoch => "meta_epoch",
            Self::MetaRelease => "meta_release",
            Self::MetaAcquire => "meta_acquire",
            Self::MetaFenceRelease => "meta_fence_release",
            Self::MetaFenceAcquire => "meta_fence_acquire",
            Self::MetaTransfer => "meta_transfer",
            Self::SafeRegionBegin => "safe_region_begin",
            Self::SafeRegionEnd => "safe_region_end",
            Self::RestrictEnter => "restrict_enter",
            Self::RestrictLeave => "restrict_leave",
            Self::Jump => "jump",
            Self::BrIf => "br_if",
            Self::Switch => "switch",
            Self::IndirectBr => "indirect_br",
            Self::Return => "return",
            Self::Unreachable => "unreachable",
            Self::Call => "call",
            Self::CallIndirect => "call_indirect",
            Self::TailCall => "tail_call",
            Self::Ctlz => "ctlz",
            Self::Cttz => "cttz",
            Self::Ctpop => "ctpop",
            Self::Bswap => "bswap",
            Self::Bitreverse => "bitreverse",
            Self::SAddOverflow => "sadd_overflow",
            Self::UAddOverflow => "uadd_overflow",
            Self::SSubOverflow => "ssub_overflow",
            Self::USubOverflow => "usub_overflow",
            Self::SMulOverflow => "smul_overflow",
            Self::UMulOverflow => "umul_overflow",
            Self::Expect => "expect",
            Self::UnreachableHint => "unreachable_hint",
            Self::Trap => "trap",
            Self::Prefetch => "prefetch",
            Self::FrameAddress => "frame_address",
            Self::ReturnAddress => "return_address",
            Self::ThreadPointer => "thread_pointer",
            Self::VaStart => "va_start",
            Self::VaArg => "va_arg",
            Self::VaObject => "va_object",
            Self::VaEnd => "va_end",
            Self::VaCopy => "va_copy",
            Self::StackSave => "stacksave",
            Self::StackRestore => "stackrestore",
            Self::SetjmpMarker => "setjmp_marker",
            Self::LongjmpMarker => "longjmp_marker",
            Self::TargetIntrinsic => "target_intrinsic",
            Self::InlineAsm => "inline_asm",
        }
    }

    /// Every opcode, in the order they are declared.
    ///
    /// The parser walks this rather than holding a second table, because a second table is a
    /// table that can disagree with the first one.
    pub fn all() -> impl Iterator<Item = Self> {
        ALL.iter().copied()
    }

    /// The opcode with that name, if there is one.
    #[must_use]
    pub fn from_name(name: &str) -> Option<Self> {
        ALL.iter().copied().find(|op| op.name() == name)
    }

    /// Whether this ends a block.
    ///
    /// [`Opcode::InlineAsm`] is not here and is the one instruction whose answer depends on
    /// the instruction rather than on the opcode: `asm goto` has successors and everything
    /// else does not. Ask the instruction, not the opcode.
    #[must_use]
    pub const fn is_terminator(self) -> bool {
        matches!(
            self,
            Self::Jump
                | Self::BrIf
                | Self::Switch
                | Self::IndirectBr
                | Self::Return
                | Self::Unreachable
                | Self::TailCall
        )
    }

    /// Whether the operands can be swapped without changing the result.
    ///
    /// The floating point cases are commutative even under the strictest rounding, because
    /// swapping the operands of an addition does not change which of them is a NaN, and the
    /// sign of a NaN result is not something we promise anything about either way.
    #[must_use]
    pub const fn is_commutative(self) -> bool {
        matches!(
            self,
            Self::Add
                | Self::Mul
                | Self::And
                | Self::Or
                | Self::Xor
                | Self::FAdd
                | Self::FMul
                | Self::SAddOverflow
                | Self::UAddOverflow
                | Self::SMulOverflow
                | Self::UMulOverflow
        )
    }

    /// Whether this reads or writes memory, or has an effect the optimizer has to preserve.
    ///
    /// An instruction that answers no can be deleted when nothing uses its result, moved
    /// across a call, and merged with another one computing the same thing. Everything else
    /// has to be argued about individually, so the conservative answer is the true one here
    /// and the list of exceptions is the part that is checked.
    #[must_use]
    pub const fn has_effects(self) -> bool {
        !matches!(
            self,
            Self::IConst
                | Self::FConst
                | Self::Splat
                | Self::GlobalAddr
                | Self::BlockAddr
                | Self::Add
                | Self::Sub
                | Self::Mul
                | Self::SDiv
                | Self::UDiv
                | Self::SRem
                | Self::URem
                | Self::And
                | Self::Or
                | Self::Xor
                | Self::Shl
                | Self::LShr
                | Self::AShr
                | Self::FAdd
                | Self::FSub
                | Self::FMul
                | Self::FDiv
                | Self::FRem
                | Self::FNeg
                | Self::Fma
                | Self::ICmp
                | Self::FCmp
                | Self::Select
                | Self::Trunc
                | Self::SExt
                | Self::ZExt
                | Self::FPTrunc
                | Self::FPExt
                | Self::FPToSI
                | Self::FPToUI
                | Self::SIToFP
                | Self::UIToFP
                | Self::PtrToInt
                | Self::IntToPtr
                | Self::Bitcast
                | Self::PtrAdd
                | Self::Ctlz
                | Self::Cttz
                | Self::Ctpop
                | Self::Bswap
                | Self::Bitreverse
                | Self::SAddOverflow
                | Self::UAddOverflow
                | Self::SSubOverflow
                | Self::USubOverflow
                | Self::SMulOverflow
                | Self::UMulOverflow
                | Self::Expect
                | Self::FrameAddress
                | Self::ReturnAddress
                // The same address for as long as the thread runs, and a thread cannot change
                // which one it is part way through a function, so two of these in one function
                // are the same value and either may be moved to where the other is.
                | Self::ThreadPointer
                | Self::MemEntry
                // Three of the capability instructions are arithmetic on a pointer's
                // provenance and touch nothing. The other three do: `cap_load` and
                // `cap_store` are an access, and `cap_recover` reads the planes.
                | Self::CapOf
                | Self::CapNull
                | Self::CapNarrow
        )
    }

    /// Whether an instruction with this opcode touches memory.
    ///
    /// This is what decides whether it takes a memory operand once memory SSA is built, per
    /// document 09 of `spec/optimizer`. It is written as the exceptions to touching memory
    /// rather than as a list of what does, for the reason document 08.6 gives about the escape
    /// analysis: an opcode added later has to end up on the conservative side by default, and a
    /// list of what touches memory would silently leave a new one out.
    ///
    /// `mem_entry` answers no. It produces memory rather than touching it, which is the whole
    /// of what it is for.
    #[must_use]
    pub const fn touches_memory(self) -> bool {
        if !self.has_effects() {
            return false;
        }
        !matches!(
            self,
            // Fresh storage nothing could have been reading, and the pointer that names it.
            Self::Alloca
                // The stack pointer, which is a register and not memory. Putting it back is a
                // different matter and is below, because it takes storage away.
                | Self::StackSave
                // Control, which goes somewhere rather than touching anything. A tail call is
                // not here, because it is a call.
                | Self::Jump
                | Self::BrIf
                | Self::Switch
                | Self::IndirectBr
                | Self::Return
                | Self::Unreachable
                | Self::UnreachableHint
        )
    }

    /// Whether an instruction with this opcode writes memory, and so produces a new version of
    /// it rather than only reading the version it was given.
    ///
    /// Everything that touches memory writes it except the ones that plainly do not. A `fence`
    /// writes nothing and is still a write here, because document 09.5 says an atomic or a
    /// barrier is a definition nothing walks past, and giving it one is how that is expressed
    /// in a representation whose only ordering is the memory chain.
    ///
    /// The checks read the planes and change nothing, which
    /// `spec/safe-memory/06-instrumentation.md` section 6.2.4 states as the word `readonly`. A
    /// check that trapped is a program that stopped and there is no version of memory after it
    /// for anything to observe, so the trap costs nothing here. What it does cost is that a
    /// check may not be moved across a plane write, and that is the memory chain saying so
    /// rather than this.
    #[must_use]
    pub const fn writes_memory(self) -> bool {
        self.touches_memory()
            && !matches!(
                self,
                Self::Load
                    | Self::AtomicLoad
                    | Self::Prefetch
                    | Self::CapLoad
                    | Self::CapRecover
                    | Self::CapArg
                    | Self::CapResult
                    | Self::CapExtent
                    | Self::CapExtentBack
                    | Self::CheckBounds
                    | Self::CheckLive
                    | Self::CheckType
                    | Self::CheckInit
                    | Self::CheckDeriv
                    | Self::CheckRace
            )
    }

    /// How many values this produces, for the opcodes where the count is fixed.
    ///
    /// `None` means the count comes from somewhere else: a call takes it from its signature,
    /// and inline assembly takes it from its output constraints. A tail call is not one of
    /// them, because whatever it returns goes straight out of the function and there is no
    /// instruction after it to use anything.
    #[must_use]
    pub const fn results(self) -> Option<u8> {
        match self {
            Self::Call | Self::CallIndirect | Self::InlineAsm => None,
            Self::Cmpxchg
            | Self::SAddOverflow
            | Self::UAddOverflow
            | Self::SSubOverflow
            | Self::USubOverflow
            | Self::SMulOverflow
            | Self::UMulOverflow => Some(2),
            Self::Store
            | Self::Memcpy
            | Self::Memmove
            | Self::Memset
            | Self::AtomicStore
            | Self::Fence
            | Self::Prefetch
            | Self::VaStart
            | Self::VaEnd
            | Self::VaCopy
            | Self::StackRestore
            | Self::UnreachableHint
            | Self::Trap
            | Self::SetjmpMarker
            | Self::LongjmpMarker
            | Self::CapStore
            | Self::CapPublish
            | Self::CapClear
            | Self::CapYield
            | Self::CheckBounds
            | Self::CheckLive
            | Self::CheckType
            | Self::CheckInit
            | Self::CheckDeriv
            | Self::CheckRace
            | Self::CheckRestrictRead
            | Self::CheckRestrictWrite
            | Self::MetaBegin
            | Self::MetaEnd
            | Self::MetaType
            | Self::MetaTypeCopy
            | Self::MetaInit
            | Self::MetaInitCopy
            | Self::MetaEpoch
            | Self::MetaRelease
            | Self::MetaAcquire
            | Self::MetaFenceRelease
            | Self::MetaFenceAcquire
            | Self::MetaTransfer
            | Self::SafeRegionBegin
            | Self::SafeRegionEnd
            | Self::RestrictEnter
            | Self::RestrictLeave => Some(0),
            _ if self.is_terminator() => Some(0),
            _ => Some(1),
        }
    }

    /// Whether an instruction with this opcode produces a capability.
    ///
    /// Seven of the thirteen `cap` instructions. The other six consume one instead, or none at all:
    /// `cap_store` writes one beside a pointer, `cap_extent` and `cap_extent_back` ask one a
    /// question about itself and answer with a number, `cap_publish` hands a call's worth of them
    /// to a callee, `cap_yield` leaves one where the caller of this function will look for it, and
    /// `cap_clear` takes no operands because saying there is no frame is not a statement about any
    /// capability. The reason this is a question about the opcode rather than
    /// about the result type is that the verifier asks it the other way round: it walks the results
    /// looking for a `cap` and needs to know whether the instruction under it was entitled to make
    /// one.
    #[must_use]
    pub const fn makes_capability(self) -> bool {
        matches!(
            self,
            Self::CapOf
                | Self::CapLoad
                | Self::CapNull
                | Self::CapNarrow
                | Self::CapRecover
                | Self::CapArg
                | Self::CapResult
        )
    }

    /// Which operand of a capability producer names the pointer the capability is about.
    ///
    /// Four of the seven [`Opcode::makes_capability`] lists, and they all mean the same thing by it:
    /// the capability describes the object that pointer is in. Where they differ is only in how the
    /// answer was arrived at, which is a walk of the lifetime plane for `cap_recover`, a read of the
    /// slot beside the word for `cap_load`, a read of the caller's frame for `cap_arg`, and whatever
    /// the back end has at hand for `cap_of`.
    ///
    /// That is worth stating as one question because the optimizer asks it. A rule that discharges a
    /// check by knowing which pointer the check's capability is about has no business caring which
    /// producer supplied it, and while `cap_of` was the only one anything emitted, asking for the
    /// opcode by name and taking operand zero was the same question. It stopped being the same
    /// question when tamnd/rucc#1241 started emitting the cheap producers, and a rule that still
    /// asked by name would quietly discharge less the better the code got.
    ///
    /// The three that answer nothing are the three that are not about a pointer at all. A
    /// `cap_narrow` is about another capability, a `cap_null` is about nothing by construction, and
    /// a `cap_result` is about a pointer the callee returned, which is a value this function has
    /// only as the call's own result and not as an operand.
    #[must_use]
    pub const fn capability_names(self) -> Option<usize> {
        match self {
            Self::CapOf | Self::CapRecover | Self::CapArg => Some(0),
            // The third, because the first two are the container's capability and the address of
            // the word, and the pointer this one is about is the value that came out of the word.
            Self::CapLoad => Some(2),
            _ => None,
        }
    }

    /// Which payload an instruction with this opcode carries.
    ///
    /// The printer reads the payload it finds and does not need this. The parser has only the
    /// opcode when it reaches the operands, so this is where the two of them agree on what
    /// comes after them. An instruction carrying a payload of some other kind prints as text
    /// the parser cannot read back, which is why the verifier checks it against
    /// [`Extra::kind`](crate::Extra::kind) rather than leaving it to be found later.
    #[must_use]
    pub const fn extra_kind(self) -> ExtraKind {
        match self {
            Self::IConst | Self::FConst | Self::Splat => ExtraKind::Imm,
            Self::GlobalAddr | Self::TargetIntrinsic => ExtraKind::Symbol,
            Self::ICmp => ExtraKind::IntPred,
            Self::FCmp => ExtraKind::FloatPred,
            Self::Alloca
            | Self::Load
            | Self::Store
            | Self::Memcpy
            | Self::Memmove
            | Self::Memset
            | Self::AtomicLoad
            | Self::AtomicStore
            | Self::Cmpxchg
            // Four of the checks are about a run of bytes and the payload is where the size
            // of that run is, along with the alignment `check_bounds` wants and the aliasing
            // node `check_type` compares against. The other two ask a question about a
            // pointer and not about a range, so they carry nothing.
            | Self::CheckBounds
            | Self::CheckType
            | Self::CheckInit
            | Self::CheckRace
            // The two `restrict` checks and the marker that opens their scope. The first two carry
            // the size of the access and the two numbers saying which pointer it went through, and
            // the third carries the size of the slot and the numbers describing the scope itself.
            | Self::CheckRestrictRead
            | Self::CheckRestrictWrite
            | Self::RestrictEnter => ExtraKind::Mem,
            // The plane writes. What each one needs beyond the range is different, and the range
            // itself is operands, since the length of a variable length array is a value.
            Self::MetaBegin => ExtraKind::Class,
            Self::MetaTransfer => ExtraKind::Owner,
            Self::MetaType => ExtraKind::Node,
            Self::SafeRegionBegin => ExtraKind::Reason,
            Self::VaObject => ExtraKind::VaObject,
            Self::AtomicRmw => ExtraKind::Rmw,
            Self::Fence => ExtraKind::Order,
            Self::Prefetch => ExtraKind::Prefetch,
            // How far up the chain of frames to walk, which is a number written in the instruction
            // and never a value. The builtins these came from take a constant and nothing else, for
            // the reason `prefetch` takes one: the instructions this becomes are a walk of that
            // length, and a length not known until the program runs has nothing to walk.
            Self::FrameAddress | Self::ReturnAddress => ExtraKind::Depth,
            Self::Jump | Self::BrIf | Self::BlockAddr | Self::IndirectBr => ExtraKind::Targets,
            Self::Switch => ExtraKind::Switch,
            Self::Call | Self::CallIndirect | Self::TailCall => ExtraKind::Call,
            Self::InlineAsm => ExtraKind::Asm,
            _ => ExtraKind::None,
        }
    }
}

/// Which of [`Extra`](crate::Extra)'s shapes an instruction carries.
///
/// The same list of names, without any of the payloads, so that a question about an opcode can
/// be answered without an instruction to look at.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub enum ExtraKind {
    /// Nothing.
    None,
    /// A constant.
    Imm,
    /// A name.
    Symbol,
    /// An integer comparison predicate.
    IntPred,
    /// A floating point comparison predicate.
    FloatPred,
    /// An access.
    Mem,
    /// An atomic read-modify-write.
    Rmw,
    /// A barrier's ordering.
    Order,
    /// What a prefetch is a hint about.
    Prefetch,
    /// How many frames up to walk.
    Depth,
    /// Branch targets.
    Targets,
    /// A call.
    Call,
    /// A `switch`.
    Switch,
    /// Inline assembly.
    Asm,
    /// An object read off a variable argument list.
    VaObject,
    /// What kind of storage an instance is.
    Class,
    /// Who a range of memory went to.
    Owner,
    /// A metadata node.
    Node,
    /// Why a declared exemption is there.
    Reason,
}

impl ExtraKind {
    /// What it is, in words, for a message that names two of them and has to read as English.
    #[must_use]
    pub const fn name(self) -> &'static str {
        match self {
            Self::None => "nothing",
            Self::Imm => "a constant",
            Self::Symbol => "a name",
            Self::IntPred => "an integer comparison",
            Self::FloatPred => "a floating point comparison",
            Self::Mem => "an access",
            Self::Rmw => "a read-modify-write",
            Self::Order => "an ordering",
            Self::Prefetch => "a prefetch hint",
            Self::Depth => "a depth",
            Self::Targets => "branch targets",
            Self::Call => "a call",
            Self::Switch => "a switch",
            Self::Asm => "inline assembly",
            Self::VaObject => "an object off a variable argument list",
            Self::Class => "a storage class",
            Self::Owner => "an owner",
            Self::Node => "a metadata node",
            Self::Reason => "a reason",
        }
    }
}

impl fmt::Display for Opcode {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(self.name())
    }
}

/// Every opcode, which is what [`Opcode::all`] hands out.
///
/// This is written out rather than derived, and the test below is what keeps it complete: it
/// checks the count against [`Opcode::InlineAsm`], the last variant, so a new opcode that is
/// not added here fails the build rather than going quietly missing from the parser.
static ALL: &[Opcode] = &[
    Opcode::IConst,
    Opcode::FConst,
    Opcode::Splat,
    Opcode::GlobalAddr,
    Opcode::BlockAddr,
    Opcode::Add,
    Opcode::Sub,
    Opcode::Mul,
    Opcode::SDiv,
    Opcode::UDiv,
    Opcode::SRem,
    Opcode::URem,
    Opcode::And,
    Opcode::Or,
    Opcode::Xor,
    Opcode::Shl,
    Opcode::LShr,
    Opcode::AShr,
    Opcode::FAdd,
    Opcode::FSub,
    Opcode::FMul,
    Opcode::FDiv,
    Opcode::FRem,
    Opcode::FNeg,
    Opcode::Fma,
    Opcode::ICmp,
    Opcode::FCmp,
    Opcode::Select,
    Opcode::Trunc,
    Opcode::SExt,
    Opcode::ZExt,
    Opcode::FPTrunc,
    Opcode::FPExt,
    Opcode::FPToSI,
    Opcode::FPToUI,
    Opcode::SIToFP,
    Opcode::UIToFP,
    Opcode::PtrToInt,
    Opcode::IntToPtr,
    Opcode::Bitcast,
    Opcode::MemEntry,
    Opcode::Alloca,
    Opcode::Load,
    Opcode::Store,
    Opcode::PtrAdd,
    Opcode::Memcpy,
    Opcode::Memmove,
    Opcode::Memset,
    Opcode::AtomicLoad,
    Opcode::AtomicStore,
    Opcode::AtomicRmw,
    Opcode::Cmpxchg,
    Opcode::Fence,
    Opcode::CapOf,
    Opcode::CapLoad,
    Opcode::CapStore,
    Opcode::CapNull,
    Opcode::CapNarrow,
    Opcode::CapRecover,
    Opcode::CapExtent,
    Opcode::CapExtentBack,
    Opcode::CapPublish,
    Opcode::CapClear,
    Opcode::CapArg,
    Opcode::CapYield,
    Opcode::CapResult,
    Opcode::CheckBounds,
    Opcode::CheckLive,
    Opcode::CheckType,
    Opcode::CheckInit,
    Opcode::CheckDeriv,
    Opcode::CheckRace,
    Opcode::CheckRestrictRead,
    Opcode::CheckRestrictWrite,
    Opcode::MetaBegin,
    Opcode::MetaEnd,
    Opcode::MetaType,
    Opcode::MetaTypeCopy,
    Opcode::MetaInit,
    Opcode::MetaInitCopy,
    Opcode::MetaEpoch,
    Opcode::MetaRelease,
    Opcode::MetaAcquire,
    Opcode::MetaFenceRelease,
    Opcode::MetaFenceAcquire,
    Opcode::MetaTransfer,
    Opcode::SafeRegionBegin,
    Opcode::SafeRegionEnd,
    Opcode::RestrictEnter,
    Opcode::RestrictLeave,
    Opcode::Jump,
    Opcode::BrIf,
    Opcode::Switch,
    Opcode::IndirectBr,
    Opcode::Return,
    Opcode::Unreachable,
    Opcode::Call,
    Opcode::CallIndirect,
    Opcode::TailCall,
    Opcode::Ctlz,
    Opcode::Cttz,
    Opcode::Ctpop,
    Opcode::Bswap,
    Opcode::Bitreverse,
    Opcode::SAddOverflow,
    Opcode::UAddOverflow,
    Opcode::SSubOverflow,
    Opcode::USubOverflow,
    Opcode::SMulOverflow,
    Opcode::UMulOverflow,
    Opcode::Expect,
    Opcode::UnreachableHint,
    Opcode::Trap,
    Opcode::Prefetch,
    Opcode::FrameAddress,
    Opcode::ReturnAddress,
    Opcode::ThreadPointer,
    Opcode::VaStart,
    Opcode::VaArg,
    Opcode::VaObject,
    Opcode::VaEnd,
    Opcode::VaCopy,
    Opcode::StackSave,
    Opcode::StackRestore,
    Opcode::SetjmpMarker,
    Opcode::LongjmpMarker,
    Opcode::TargetIntrinsic,
    Opcode::InlineAsm,
];

/// The ten integer comparisons.
///
/// Signedness is on the predicate rather than on the type, for the same reason it is on
/// `sdiv` and `udiv`: the type space is halved and the operation says what it means.
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub enum IntPred {
    /// Equal.
    Eq,
    /// Not equal.
    Ne,
    /// Signed less than.
    Slt,
    /// Signed less than or equal.
    Sle,
    /// Signed greater than.
    Sgt,
    /// Signed greater than or equal.
    Sge,
    /// Unsigned less than.
    Ult,
    /// Unsigned less than or equal.
    Ule,
    /// Unsigned greater than.
    Ugt,
    /// Unsigned greater than or equal.
    Uge,
}

impl IntPred {
    /// The textual form.
    #[must_use]
    pub const fn name(self) -> &'static str {
        match self {
            Self::Eq => "eq",
            Self::Ne => "ne",
            Self::Slt => "slt",
            Self::Sle => "sle",
            Self::Sgt => "sgt",
            Self::Sge => "sge",
            Self::Ult => "ult",
            Self::Ule => "ule",
            Self::Ugt => "ugt",
            Self::Uge => "uge",
        }
    }

    /// The predicate with that name, if there is one.
    #[must_use]
    pub fn from_name(name: &str) -> Option<Self> {
        Self::all().find(|pred| pred.name() == name)
    }

    /// Every predicate.
    pub fn all() -> impl Iterator<Item = Self> {
        [
            Self::Eq,
            Self::Ne,
            Self::Slt,
            Self::Sle,
            Self::Sgt,
            Self::Sge,
            Self::Ult,
            Self::Ule,
            Self::Ugt,
            Self::Uge,
        ]
        .into_iter()
    }

    /// The predicate that holds exactly when this one does not.
    #[must_use]
    pub const fn inverse(self) -> Self {
        match self {
            Self::Eq => Self::Ne,
            Self::Ne => Self::Eq,
            Self::Slt => Self::Sge,
            Self::Sge => Self::Slt,
            Self::Sle => Self::Sgt,
            Self::Sgt => Self::Sle,
            Self::Ult => Self::Uge,
            Self::Uge => Self::Ult,
            Self::Ule => Self::Ugt,
            Self::Ugt => Self::Ule,
        }
    }

    /// The predicate that holds when the operands are given the other way round.
    #[must_use]
    pub const fn swapped(self) -> Self {
        match self {
            Self::Eq => Self::Eq,
            Self::Ne => Self::Ne,
            Self::Slt => Self::Sgt,
            Self::Sgt => Self::Slt,
            Self::Sle => Self::Sge,
            Self::Sge => Self::Sle,
            Self::Ult => Self::Ugt,
            Self::Ugt => Self::Ult,
            Self::Ule => Self::Uge,
            Self::Uge => Self::Ule,
        }
    }

    /// Whether this reads its operands as signed. Equality reads them as neither.
    #[must_use]
    pub const fn is_signed(self) -> bool {
        matches!(self, Self::Slt | Self::Sle | Self::Sgt | Self::Sge)
    }
}

impl fmt::Display for IntPred {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(self.name())
    }
}

/// The floating point comparisons, ordered and unordered.
///
/// An ordered predicate is false if either operand is a NaN, and an unordered one is true. C's
/// `<` is `olt` and C's `!=` is `une`, which is the whole of why both families are here.
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub enum FloatPred {
    /// Always false.
    False,
    /// Ordered and equal.
    Oeq,
    /// Ordered and greater than.
    Ogt,
    /// Ordered and greater than or equal.
    Oge,
    /// Ordered and less than.
    Olt,
    /// Ordered and less than or equal.
    Ole,
    /// Ordered and not equal.
    One,
    /// Ordered, which is to say neither operand is a NaN.
    Ord,
    /// Unordered, which is to say one of them is.
    Uno,
    /// Unordered or equal.
    Ueq,
    /// Unordered or greater than.
    Ugt,
    /// Unordered or greater than or equal.
    Uge,
    /// Unordered or less than.
    Ult,
    /// Unordered or less than or equal.
    Ule,
    /// Unordered or not equal.
    Une,
    /// Always true.
    True,
}

impl FloatPred {
    /// The textual form.
    #[must_use]
    pub const fn name(self) -> &'static str {
        match self {
            Self::False => "false",
            Self::Oeq => "oeq",
            Self::Ogt => "ogt",
            Self::Oge => "oge",
            Self::Olt => "olt",
            Self::Ole => "ole",
            Self::One => "one",
            Self::Ord => "ord",
            Self::Uno => "uno",
            Self::Ueq => "ueq",
            Self::Ugt => "ugt",
            Self::Uge => "uge",
            Self::Ult => "ult",
            Self::Ule => "ule",
            Self::Une => "une",
            Self::True => "true",
        }
    }

    /// The predicate with that name, if there is one.
    #[must_use]
    pub fn from_name(name: &str) -> Option<Self> {
        Self::all().find(|pred| pred.name() == name)
    }

    /// Every predicate.
    pub fn all() -> impl Iterator<Item = Self> {
        [
            Self::False,
            Self::Oeq,
            Self::Ogt,
            Self::Oge,
            Self::Olt,
            Self::Ole,
            Self::One,
            Self::Ord,
            Self::Uno,
            Self::Ueq,
            Self::Ugt,
            Self::Uge,
            Self::Ult,
            Self::Ule,
            Self::Une,
            Self::True,
        ]
        .into_iter()
    }

    /// The predicate that holds exactly when this one does not.
    #[must_use]
    pub const fn inverse(self) -> Self {
        match self {
            Self::False => Self::True,
            Self::Oeq => Self::Une,
            Self::Ogt => Self::Ule,
            Self::Oge => Self::Ult,
            Self::Olt => Self::Uge,
            Self::Ole => Self::Ugt,
            Self::One => Self::Ueq,
            Self::Ord => Self::Uno,
            Self::Uno => Self::Ord,
            Self::Ueq => Self::One,
            Self::Ugt => Self::Ole,
            Self::Uge => Self::Olt,
            Self::Ult => Self::Oge,
            Self::Ule => Self::Ogt,
            Self::Une => Self::Oeq,
            Self::True => Self::False,
        }
    }

    /// The predicate that holds when the operands are given the other way round.
    #[must_use]
    pub const fn swapped(self) -> Self {
        match self {
            Self::Ogt => Self::Olt,
            Self::Olt => Self::Ogt,
            Self::Oge => Self::Ole,
            Self::Ole => Self::Oge,
            Self::Ugt => Self::Ult,
            Self::Ult => Self::Ugt,
            Self::Uge => Self::Ule,
            Self::Ule => Self::Uge,
            same => same,
        }
    }

    /// Whether this is false when either operand is a NaN.
    ///
    /// [`FloatPred::False`] and [`FloatPred::True`] are neither ordered nor unordered, since
    /// they do not look at their operands at all, and both answer no here.
    #[must_use]
    pub const fn is_ordered(self) -> bool {
        matches!(
            self,
            Self::Oeq | Self::Ogt | Self::Oge | Self::Olt | Self::Ole | Self::One | Self::Ord
        )
    }
}

impl fmt::Display for FloatPred {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(self.name())
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn every_opcode_is_in_the_table() {
        // `InlineAsm` is the last variant, so its discriminant plus one is how many there are.
        // A new opcode declared after it moves this number, and a new opcode declared before
        // it and not added to `ALL` moves the length, so either mistake fails here.
        assert_eq!(ALL.len(), Opcode::InlineAsm as usize + 1);
        for (position, &op) in ALL.iter().enumerate() {
            assert_eq!(op as usize, position, "{op} is out of order in ALL");
        }
    }

    #[test]
    fn every_opcode_name_is_one_word_the_reader_can_take() {
        // The textual form keeps the dot for the type suffix and the flags, so an opcode with a
        // dot in it reads back as a shorter opcode with a suffix that is not a type. The safety
        // instructions are spelled `cap_of` and not `cap.of` for this reason, and the
        // specification says so at `spec/safe-memory/06-instrumentation.md` section 6.2.2.
        for opcode in Opcode::all() {
            let name = opcode.name();
            assert!(!name.is_empty(), "an opcode with no name");
            assert!(
                name.bytes().all(|b| b.is_ascii_lowercase() || b.is_ascii_digit() || b == b'_'),
                "{name} is not one word"
            );
        }
    }

    #[test]
    fn every_opcode_has_its_own_name_and_finds_it_again() {
        let mut names: Vec<&str> = Opcode::all().map(Opcode::name).collect();
        let total = names.len();
        names.sort_unstable();
        names.dedup();
        assert_eq!(names.len(), total, "two opcodes share a name");
        for op in Opcode::all() {
            assert_eq!(Opcode::from_name(op.name()), Some(op));
        }
        assert_eq!(Opcode::from_name("phi"), None);
        assert_eq!(Opcode::from_name("getelementptr"), None);
        assert_eq!(Opcode::from_name(""), None);
    }

    #[test]
    fn the_terminators_are_the_ones_control_leaves_by() {
        let terminators: Vec<&str> =
            Opcode::all().filter(|op| op.is_terminator()).map(Opcode::name).collect();
        assert_eq!(
            terminators,
            ["jump", "br_if", "switch", "indirect_br", "return", "unreachable", "tail_call"]
        );
    }

    #[test]
    fn a_terminator_produces_nothing() {
        for op in Opcode::all().filter(|op| op.is_terminator()) {
            assert_eq!(op.results(), Some(0), "{op}");
        }
    }

    #[test]
    fn the_pair_producing_opcodes_are_the_ones_with_a_flag_beside_the_value() {
        let pairs: Vec<&str> =
            Opcode::all().filter(|op| op.results() == Some(2)).map(Opcode::name).collect();
        assert_eq!(
            pairs,
            [
                "cmpxchg",
                "sadd_overflow",
                "uadd_overflow",
                "ssub_overflow",
                "usub_overflow",
                "smul_overflow",
                "umul_overflow"
            ]
        );
    }

    #[test]
    fn the_capability_instructions_are_the_ones_that_make_a_capability() {
        let makers: Vec<Opcode> = Opcode::all().filter(|op| op.makes_capability()).collect();
        assert_eq!(
            makers,
            vec![
                Opcode::CapOf,
                Opcode::CapLoad,
                Opcode::CapNull,
                Opcode::CapNarrow,
                Opcode::CapRecover,
                Opcode::CapArg,
                Opcode::CapResult
            ]
        );
        // The other three read a capability rather than making one. `cap_store` writes it out and
        // produces nothing at all, and the two extent queries answer with a number.
        assert!(!Opcode::CapStore.makes_capability());
        assert_eq!(Opcode::CapStore.results(), Some(0));
        assert!(!Opcode::CapExtent.makes_capability());
        assert_eq!(Opcode::CapExtent.results(), Some(1));
        assert!(!Opcode::CapExtentBack.makes_capability());
        assert_eq!(Opcode::CapExtentBack.results(), Some(1));
        for opcode in makers {
            assert_eq!(opcode.results(), Some(1), "{}", opcode.name());
        }
    }

    #[test]
    fn a_producer_says_which_of_its_operands_is_the_pointer_it_is_about() {
        // Four of the seven, and the one that is not operand zero is the one whose first two
        // operands are the container and the word rather than the value that came out of it.
        assert_eq!(Opcode::CapOf.capability_names(), Some(0));
        assert_eq!(Opcode::CapRecover.capability_names(), Some(0));
        assert_eq!(Opcode::CapArg.capability_names(), Some(0));
        assert_eq!(Opcode::CapLoad.capability_names(), Some(2));

        // The three that are about something other than a pointer this function has an operand for.
        assert_eq!(Opcode::CapNarrow.capability_names(), None);
        assert_eq!(Opcode::CapNull.capability_names(), None);
        assert_eq!(Opcode::CapResult.capability_names(), None);

        // Nothing that is not a producer answers, since the question is what a capability describes
        // and those have no capability to describe anything with.
        assert_eq!(Opcode::CapStore.capability_names(), None);
        assert_eq!(Opcode::Load.capability_names(), None);
        assert_eq!(Opcode::CheckLive.capability_names(), None);
    }

    #[test]
    fn a_check_reads_the_planes_and_writes_nothing() {
        let checks = [
            Opcode::CheckBounds,
            Opcode::CheckLive,
            Opcode::CheckType,
            Opcode::CheckInit,
            Opcode::CheckDeriv,
            Opcode::CheckRace,
        ];
        for opcode in checks {
            let name = opcode.name();
            // It traps, so it stays where it was put and nothing deletes it for having no
            // result. It reads a plane, so it takes a memory operand. It writes nothing, so
            // the access after it reads the version the check was given.
            assert!(opcode.has_effects(), "{name}");
            assert!(opcode.touches_memory(), "{name}");
            assert!(!opcode.writes_memory(), "{name}");
            assert_eq!(opcode.results(), Some(0), "{name}");
        }
    }

    #[test]
    fn a_restrict_check_writes_memory_because_it_records_what_it_saw() {
        // The one place the sentence above does not hold. Every other check reads a plane and
        // leaves it alone, so the optimizer may hoist one out of a loop or keep the later of two
        // identical ones. These record the range they were asked about into the block's own slot,
        // so a check that ran twice saw two accesses and a check that was hoisted saw one, and
        // either rewrite changes what the next one answers. Saying they write memory is how the
        // memory chain refuses both.
        let recording = [
            Opcode::CheckRestrictRead,
            Opcode::CheckRestrictWrite,
            Opcode::RestrictEnter,
            Opcode::RestrictLeave,
        ];
        for opcode in recording {
            let name = opcode.name();
            assert!(opcode.has_effects(), "{name}");
            assert!(opcode.touches_memory(), "{name}");
            assert!(opcode.writes_memory(), "{name}");
            assert_eq!(opcode.results(), Some(0), "{name}");
        }
    }

    #[test]
    fn the_capability_instructions_that_touch_memory_are_the_five_that_have_to() {
        // `cap_load` and `cap_store` are an access to the slot beside a pointer, and `cap_recover`
        // and the two extent queries read the planes. The other three are arithmetic on a
        // provenance the program already had, so the optimizer may treat them as it treats
        // `ptr_add`.
        assert!(!Opcode::CapOf.has_effects());
        assert!(!Opcode::CapNull.has_effects());
        assert!(!Opcode::CapNarrow.has_effects());
        assert!(Opcode::CapLoad.touches_memory() && !Opcode::CapLoad.writes_memory());
        assert!(Opcode::CapRecover.touches_memory() && !Opcode::CapRecover.writes_memory());
        assert!(Opcode::CapExtent.touches_memory() && !Opcode::CapExtent.writes_memory());
        assert!(Opcode::CapExtentBack.touches_memory() && !Opcode::CapExtentBack.writes_memory());
        assert!(Opcode::CapStore.writes_memory());
    }

    #[test]
    fn memory_has_effects_and_arithmetic_does_not() {
        for op in [Opcode::Load, Opcode::Store, Opcode::Call, Opcode::Alloca, Opcode::Fence] {
            assert!(op.has_effects(), "{op}");
        }
        for op in [Opcode::Add, Opcode::FDiv, Opcode::ICmp, Opcode::PtrAdd, Opcode::IConst] {
            assert!(!op.has_effects(), "{op}");
        }
    }

    #[test]
    fn commuting_is_only_claimed_where_it_holds() {
        assert!(Opcode::Add.is_commutative());
        assert!(Opcode::FAdd.is_commutative());
        assert!(!Opcode::Sub.is_commutative());
        assert!(!Opcode::FDiv.is_commutative());
        assert!(!Opcode::Shl.is_commutative());
    }

    #[test]
    fn an_integer_predicate_inverts_and_swaps_back_to_itself() {
        for pred in IntPred::all() {
            assert_eq!(pred.inverse().inverse(), pred);
            assert_eq!(pred.swapped().swapped(), pred);
            assert_eq!(IntPred::from_name(pred.name()), Some(pred));
        }
        assert_eq!(IntPred::Slt.inverse(), IntPred::Sge);
        assert_eq!(IntPred::Slt.swapped(), IntPred::Sgt);
        assert_eq!(IntPred::from_name("lt"), None);
    }

    #[test]
    fn a_floating_predicate_inverts_across_the_ordered_line() {
        for pred in FloatPred::all() {
            assert_eq!(pred.inverse().inverse(), pred);
            assert_eq!(pred.swapped().swapped(), pred);
            assert_eq!(FloatPred::from_name(pred.name()), Some(pred));
        }
        // Inverting has to cross the line, because the negation of an ordered comparison is
        // true when an operand is a NaN. This is where `!(a < b)` stops being `a >= b`. The
        // two constants are outside it: neither of them looks at its operands.
        for pred in FloatPred::all().filter(|p| !matches!(p, FloatPred::False | FloatPred::True)) {
            assert_ne!(pred.is_ordered(), pred.inverse().is_ordered(), "{pred}");
        }
        assert_eq!(FloatPred::Olt.inverse(), FloatPred::Uge);
        assert_eq!(FloatPred::Olt.swapped(), FloatPred::Ogt);
    }

    #[test]
    fn swapping_a_predicate_keeps_it_ordered_or_unordered() {
        for pred in FloatPred::all() {
            assert_eq!(pred.is_ordered(), pred.swapped().is_ordered(), "{pred}");
        }
        for pred in IntPred::all() {
            assert_eq!(pred.is_signed(), pred.swapped().is_signed(), "{pred}");
        }
    }

    #[test]
    fn no_two_predicates_share_a_name_within_their_family() {
        for names in [
            IntPred::all().map(IntPred::name).collect::<Vec<_>>(),
            FloatPred::all().map(FloatPred::name).collect::<Vec<_>>(),
        ] {
            let total = names.len();
            let mut names = names;
            names.sort_unstable();
            names.dedup();
            assert_eq!(names.len(), total);
        }
    }
}