moved crates in lib/ to src/, renamed crates, modified some files' text (#660)

moved crates in lib/ to src/, renamed crates, modified some files' text (#660)

cranelift/codegen/src/isa/arm32/abi.rs

@@ -0,0 +1,35 @@
+//! ARM ABI implementation.
+
+use super::registers::{D, GPR, Q, S};
+use crate::ir;
+use crate::isa::RegClass;
+use crate::regalloc::RegisterSet;
+use crate::settings as shared_settings;
+
+/// Legalize `sig`.
+pub fn legalize_signature(
+    _sig: &mut ir::Signature,
+    _flags: &shared_settings::Flags,
+    _current: bool,
+) {
+    unimplemented!()
+}
+
+/// Get register class for a type appearing in a legalized signature.
+pub fn regclass_for_abi_type(ty: ir::Type) -> RegClass {
+    if ty.is_int() {
+        GPR
+    } else {
+        match ty.bits() {
+            32 => S,
+            64 => D,
+            128 => Q,
+            _ => panic!("Unexpected {} ABI type for arm32", ty),
+        }
+    }
+}
+
+/// Get the set of allocatable registers for `func`.
+pub fn allocatable_registers(_func: &ir::Function) -> RegisterSet {
+    unimplemented!()
+}

cranelift/codegen/src/isa/arm32/binemit.rs

@@ -0,0 +1,7 @@
+//! Emitting binary ARM32 machine code.
+
+use crate::binemit::{bad_encoding, CodeSink};
+use crate::ir::{Function, Inst};
+use crate::regalloc::RegDiversions;
+
+include!(concat!(env!("OUT_DIR"), "/binemit-arm32.rs"));

cranelift/codegen/src/isa/arm32/enc_tables.rs

@@ -0,0 +1,9 @@
+//! Encoding tables for ARM32 ISA.
+
+use crate::isa;
+use crate::isa::constraints::*;
+use crate::isa::enc_tables::*;
+use crate::isa::encoding::RecipeSizing;
+
+include!(concat!(env!("OUT_DIR"), "/encoding-arm32.rs"));
+include!(concat!(env!("OUT_DIR"), "/legalize-arm32.rs"));

cranelift/codegen/src/isa/arm32/mod.rs

@@ -0,0 +1,136 @@
+//! ARM 32-bit Instruction Set Architecture.
+
+mod abi;
+mod binemit;
+mod enc_tables;
+mod registers;
+pub mod settings;
+
+use super::super::settings as shared_settings;
+#[cfg(feature = "testing_hooks")]
+use crate::binemit::CodeSink;
+use crate::binemit::{emit_function, MemoryCodeSink};
+use crate::ir;
+use crate::isa::enc_tables::{self as shared_enc_tables, lookup_enclist, Encodings};
+use crate::isa::Builder as IsaBuilder;
+use crate::isa::{EncInfo, RegClass, RegInfo, TargetIsa};
+use crate::regalloc;
+use core::fmt;
+use std::boxed::Box;
+use target_lexicon::{Architecture, Triple};
+
+#[allow(dead_code)]
+struct Isa {
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    isa_flags: settings::Flags,
+    cpumode: &'static [shared_enc_tables::Level1Entry<u16>],
+}
+
+/// Get an ISA builder for creating ARM32 targets.
+pub fn isa_builder(triple: Triple) -> IsaBuilder {
+    IsaBuilder {
+        triple,
+        setup: settings::builder(),
+        constructor: isa_constructor,
+    }
+}
+
+fn isa_constructor(
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    builder: shared_settings::Builder,
+) -> Box<TargetIsa> {
+    let level1 = match triple.architecture {
+        Architecture::Thumbv6m | Architecture::Thumbv7em | Architecture::Thumbv7m => {
+            &enc_tables::LEVEL1_T32[..]
+        }
+        Architecture::Arm
+        | Architecture::Armv4t
+        | Architecture::Armv5te
+        | Architecture::Armv7
+        | Architecture::Armv7s => &enc_tables::LEVEL1_A32[..],
+        _ => panic!(),
+    };
+    Box::new(Isa {
+        triple,
+        isa_flags: settings::Flags::new(&shared_flags, builder),
+        shared_flags,
+        cpumode: level1,
+    })
+}
+
+impl TargetIsa for Isa {
+    fn name(&self) -> &'static str {
+        "arm32"
+    }
+
+    fn triple(&self) -> &Triple {
+        &self.triple
+    }
+
+    fn flags(&self) -> &shared_settings::Flags {
+        &self.shared_flags
+    }
+
+    fn register_info(&self) -> RegInfo {
+        registers::INFO.clone()
+    }
+
+    fn encoding_info(&self) -> EncInfo {
+        enc_tables::INFO.clone()
+    }
+
+    fn legal_encodings<'a>(
+        &'a self,
+        func: &'a ir::Function,
+        inst: &'a ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Encodings<'a> {
+        lookup_enclist(
+            ctrl_typevar,
+            inst,
+            func,
+            self.cpumode,
+            &enc_tables::LEVEL2[..],
+            &enc_tables::ENCLISTS[..],
+            &enc_tables::LEGALIZE_ACTIONS[..],
+            &enc_tables::RECIPE_PREDICATES[..],
+            &enc_tables::INST_PREDICATES[..],
+            self.isa_flags.predicate_view(),
+        )
+    }
+
+    fn legalize_signature(&self, sig: &mut ir::Signature, current: bool) {
+        abi::legalize_signature(sig, &self.shared_flags, current)
+    }
+
+    fn regclass_for_abi_type(&self, ty: ir::Type) -> RegClass {
+        abi::regclass_for_abi_type(ty)
+    }
+
+    fn allocatable_registers(&self, func: &ir::Function) -> regalloc::RegisterSet {
+        abi::allocatable_registers(func)
+    }
+
+    #[cfg(feature = "testing_hooks")]
+    fn emit_inst(
+        &self,
+        func: &ir::Function,
+        inst: ir::Inst,
+        divert: &mut regalloc::RegDiversions,
+        sink: &mut CodeSink,
+    ) {
+        binemit::emit_inst(func, inst, divert, sink)
+    }
+
+    fn emit_function_to_memory(&self, func: &ir::Function, sink: &mut MemoryCodeSink) {
+        emit_function(func, binemit::emit_inst, sink)
+    }
+}
+
+impl fmt::Display for Isa {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{}\n{}", self.shared_flags, self.isa_flags)
+    }
+}

cranelift/codegen/src/isa/arm32/registers.rs

@@ -0,0 +1,68 @@
+//! ARM32 register descriptions.
+
+use crate::isa::registers::{RegBank, RegClass, RegClassData, RegInfo, RegUnit};
+
+include!(concat!(env!("OUT_DIR"), "/registers-arm32.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::{D, GPR, INFO, S};
+    use crate::isa::RegUnit;
+    use std::string::{String, ToString};
+
+    #[test]
+    fn unit_encodings() {
+        assert_eq!(INFO.parse_regunit("s0"), Some(0));
+        assert_eq!(INFO.parse_regunit("s31"), Some(31));
+        assert_eq!(INFO.parse_regunit("s32"), Some(32));
+        assert_eq!(INFO.parse_regunit("r0"), Some(64));
+        assert_eq!(INFO.parse_regunit("r15"), Some(79));
+    }
+
+    #[test]
+    fn unit_names() {
+        fn uname(ru: RegUnit) -> String {
+            INFO.display_regunit(ru).to_string()
+        }
+
+        assert_eq!(uname(0), "%s0");
+        assert_eq!(uname(1), "%s1");
+        assert_eq!(uname(31), "%s31");
+        assert_eq!(uname(64), "%r0");
+    }
+
+    #[test]
+    fn overlaps() {
+        // arm32 has the most interesting register geometries, so test `regs_overlap()` here.
+        use crate::isa::regs_overlap;
+
+        let r0 = GPR.unit(0);
+        let r1 = GPR.unit(1);
+        let r2 = GPR.unit(2);
+
+        assert!(regs_overlap(GPR, r0, GPR, r0));
+        assert!(regs_overlap(GPR, r2, GPR, r2));
+        assert!(!regs_overlap(GPR, r0, GPR, r1));
+        assert!(!regs_overlap(GPR, r1, GPR, r0));
+        assert!(!regs_overlap(GPR, r2, GPR, r1));
+        assert!(!regs_overlap(GPR, r1, GPR, r2));
+
+        let s0 = S.unit(0);
+        let s1 = S.unit(1);
+        let s2 = S.unit(2);
+        let s3 = S.unit(3);
+        let d0 = D.unit(0);
+        let d1 = D.unit(1);
+
+        assert!(regs_overlap(S, s0, D, d0));
+        assert!(regs_overlap(S, s1, D, d0));
+        assert!(!regs_overlap(S, s0, D, d1));
+        assert!(!regs_overlap(S, s1, D, d1));
+        assert!(regs_overlap(S, s2, D, d1));
+        assert!(regs_overlap(S, s3, D, d1));
+        assert!(!regs_overlap(D, d1, S, s1));
+        assert!(regs_overlap(D, d1, S, s2));
+        assert!(!regs_overlap(D, d0, D, d1));
+        assert!(regs_overlap(D, d1, D, d1));
+    }
+}

cranelift/codegen/src/isa/arm32/settings.rs

@@ -0,0 +1,9 @@
+//! ARM32 Settings.
+
+use crate::settings::{self, detail, Builder};
+use core::fmt;
+
+// Include code generated by `cranelift-codegen/meta-python/gen_settings.py`. This file contains a public
+// `Flags` struct with an impl for all of the settings defined in
+// `cranelift-codegen/meta-python/isa/arm32/settings.py`.
+include!(concat!(env!("OUT_DIR"), "/settings-arm32.rs"));

cranelift/codegen/src/isa/arm64/abi.rs

@@ -0,0 +1,30 @@
+//! ARM 64 ABI implementation.
+
+use super::registers::{FPR, GPR};
+use crate::ir;
+use crate::isa::RegClass;
+use crate::regalloc::RegisterSet;
+use crate::settings as shared_settings;
+
+/// Legalize `sig`.
+pub fn legalize_signature(
+    _sig: &mut ir::Signature,
+    _flags: &shared_settings::Flags,
+    _current: bool,
+) {
+    unimplemented!()
+}
+
+/// Get register class for a type appearing in a legalized signature.
+pub fn regclass_for_abi_type(ty: ir::Type) -> RegClass {
+    if ty.is_int() {
+        GPR
+    } else {
+        FPR
+    }
+}
+
+/// Get the set of allocatable registers for `func`.
+pub fn allocatable_registers(_func: &ir::Function) -> RegisterSet {
+    unimplemented!()
+}

cranelift/codegen/src/isa/arm64/binemit.rs

@@ -0,0 +1,7 @@
+//! Emitting binary ARM64 machine code.
+
+use crate::binemit::{bad_encoding, CodeSink};
+use crate::ir::{Function, Inst};
+use crate::regalloc::RegDiversions;
+
+include!(concat!(env!("OUT_DIR"), "/binemit-arm64.rs"));

cranelift/codegen/src/isa/arm64/enc_tables.rs

@@ -0,0 +1,9 @@
+//! Encoding tables for ARM64 ISA.
+
+use crate::isa;
+use crate::isa::constraints::*;
+use crate::isa::enc_tables::*;
+use crate::isa::encoding::RecipeSizing;
+
+include!(concat!(env!("OUT_DIR"), "/encoding-arm64.rs"));
+include!(concat!(env!("OUT_DIR"), "/legalize-arm64.rs"));

cranelift/codegen/src/isa/arm64/mod.rs

@@ -0,0 +1,123 @@
+//! ARM 64-bit Instruction Set Architecture.
+
+mod abi;
+mod binemit;
+mod enc_tables;
+mod registers;
+pub mod settings;
+
+use super::super::settings as shared_settings;
+#[cfg(feature = "testing_hooks")]
+use crate::binemit::CodeSink;
+use crate::binemit::{emit_function, MemoryCodeSink};
+use crate::ir;
+use crate::isa::enc_tables::{lookup_enclist, Encodings};
+use crate::isa::Builder as IsaBuilder;
+use crate::isa::{EncInfo, RegClass, RegInfo, TargetIsa};
+use crate::regalloc;
+use core::fmt;
+use std::boxed::Box;
+use target_lexicon::Triple;
+
+#[allow(dead_code)]
+struct Isa {
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    isa_flags: settings::Flags,
+}
+
+/// Get an ISA builder for creating ARM64 targets.
+pub fn isa_builder(triple: Triple) -> IsaBuilder {
+    IsaBuilder {
+        triple,
+        setup: settings::builder(),
+        constructor: isa_constructor,
+    }
+}
+
+fn isa_constructor(
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    builder: shared_settings::Builder,
+) -> Box<TargetIsa> {
+    Box::new(Isa {
+        triple,
+        isa_flags: settings::Flags::new(&shared_flags, builder),
+        shared_flags,
+    })
+}
+
+impl TargetIsa for Isa {
+    fn name(&self) -> &'static str {
+        "arm64"
+    }
+
+    fn triple(&self) -> &Triple {
+        &self.triple
+    }
+
+    fn flags(&self) -> &shared_settings::Flags {
+        &self.shared_flags
+    }
+
+    fn register_info(&self) -> RegInfo {
+        registers::INFO.clone()
+    }
+
+    fn encoding_info(&self) -> EncInfo {
+        enc_tables::INFO.clone()
+    }
+
+    fn legal_encodings<'a>(
+        &'a self,
+        func: &'a ir::Function,
+        inst: &'a ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Encodings<'a> {
+        lookup_enclist(
+            ctrl_typevar,
+            inst,
+            func,
+            &enc_tables::LEVEL1_A64[..],
+            &enc_tables::LEVEL2[..],
+            &enc_tables::ENCLISTS[..],
+            &enc_tables::LEGALIZE_ACTIONS[..],
+            &enc_tables::RECIPE_PREDICATES[..],
+            &enc_tables::INST_PREDICATES[..],
+            self.isa_flags.predicate_view(),
+        )
+    }
+
+    fn legalize_signature(&self, sig: &mut ir::Signature, current: bool) {
+        abi::legalize_signature(sig, &self.shared_flags, current)
+    }
+
+    fn regclass_for_abi_type(&self, ty: ir::Type) -> RegClass {
+        abi::regclass_for_abi_type(ty)
+    }
+
+    fn allocatable_registers(&self, func: &ir::Function) -> regalloc::RegisterSet {
+        abi::allocatable_registers(func)
+    }
+
+    #[cfg(feature = "testing_hooks")]
+    fn emit_inst(
+        &self,
+        func: &ir::Function,
+        inst: ir::Inst,
+        divert: &mut regalloc::RegDiversions,
+        sink: &mut CodeSink,
+    ) {
+        binemit::emit_inst(func, inst, divert, sink)
+    }
+
+    fn emit_function_to_memory(&self, func: &ir::Function, sink: &mut MemoryCodeSink) {
+        emit_function(func, binemit::emit_inst, sink)
+    }
+}
+
+impl fmt::Display for Isa {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{}\n{}", self.shared_flags, self.isa_flags)
+    }
+}

cranelift/codegen/src/isa/arm64/registers.rs

@@ -0,0 +1,39 @@
+//! ARM64 register descriptions.
+
+use crate::isa::registers::{RegBank, RegClass, RegClassData, RegInfo, RegUnit};
+
+include!(concat!(env!("OUT_DIR"), "/registers-arm64.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::INFO;
+    use crate::isa::RegUnit;
+    use std::string::{String, ToString};
+
+    #[test]
+    fn unit_encodings() {
+        assert_eq!(INFO.parse_regunit("x0"), Some(0));
+        assert_eq!(INFO.parse_regunit("x31"), Some(31));
+        assert_eq!(INFO.parse_regunit("v0"), Some(32));
+        assert_eq!(INFO.parse_regunit("v31"), Some(63));
+
+        assert_eq!(INFO.parse_regunit("x32"), None);
+        assert_eq!(INFO.parse_regunit("v32"), None);
+    }
+
+    #[test]
+    fn unit_names() {
+        fn uname(ru: RegUnit) -> String {
+            INFO.display_regunit(ru).to_string()
+        }
+
+        assert_eq!(uname(0), "%x0");
+        assert_eq!(uname(1), "%x1");
+        assert_eq!(uname(31), "%x31");
+        assert_eq!(uname(32), "%v0");
+        assert_eq!(uname(33), "%v1");
+        assert_eq!(uname(63), "%v31");
+        assert_eq!(uname(64), "%nzcv");
+        assert_eq!(uname(65), "%INVALID65");
+    }
+}

cranelift/codegen/src/isa/arm64/settings.rs

@@ -0,0 +1,9 @@
+//! ARM64 Settings.
+
+use crate::settings::{self, detail, Builder};
+use core::fmt;
+
+// Include code generated by `cranelift-codegen/meta-python/gen_settings.py`. This file contains a public
+// `Flags` struct with an impl for all of the settings defined in
+// `cranelift-codegen/meta-python/isa/arm64/settings.py`.
+include!(concat!(env!("OUT_DIR"), "/settings-arm64.rs"));

cranelift/codegen/src/isa/call_conv.rs

@@ -0,0 +1,60 @@
+use core::fmt;
+use core::str;
+use target_lexicon::{CallingConvention, Triple};
+
+/// Calling convention identifiers.
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
+pub enum CallConv {
+    /// Best performance, not ABI-stable
+    Fast,
+    /// Smallest caller code size, not ABI-stable
+    Cold,
+    /// System V-style convention used on many platforms
+    SystemV,
+    /// Windows "fastcall" convention, also used for x64 and ARM
+    WindowsFastcall,
+    /// SpiderMonkey WebAssembly convention
+    Baldrdash,
+    /// Specialized convention for the probestack function
+    Probestack,
+}
+
+impl CallConv {
+    /// Return the default calling convention for the given target triple.
+    pub fn triple_default(triple: &Triple) -> Self {
+        match triple.default_calling_convention() {
+            // Default to System V for unknown targets because most everything
+            // uses System V.
+            Ok(CallingConvention::SystemV) | Err(()) => CallConv::SystemV,
+            Ok(CallingConvention::WindowsFastcall) => CallConv::WindowsFastcall,
+        }
+    }
+}
+
+impl fmt::Display for CallConv {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        f.write_str(match *self {
+            CallConv::Fast => "fast",
+            CallConv::Cold => "cold",
+            CallConv::SystemV => "system_v",
+            CallConv::WindowsFastcall => "windows_fastcall",
+            CallConv::Baldrdash => "baldrdash",
+            CallConv::Probestack => "probestack",
+        })
+    }
+}
+
+impl str::FromStr for CallConv {
+    type Err = ();
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "fast" => Ok(CallConv::Fast),
+            "cold" => Ok(CallConv::Cold),
+            "system_v" => Ok(CallConv::SystemV),
+            "windows_fastcall" => Ok(CallConv::WindowsFastcall),
+            "baldrdash" => Ok(CallConv::Baldrdash),
+            "probestack" => Ok(CallConv::Probestack),
+            _ => Err(()),
+        }
+    }
+}

cranelift/codegen/src/isa/constraints.rs

@@ -0,0 +1,207 @@
+//! Register constraints for instruction operands.
+//!
+//! An encoding recipe specifies how an instruction is encoded as binary machine code, but it only
+//! works if the operands and results satisfy certain constraints. Constraints on immediate
+//! operands are checked by instruction predicates when the recipe is chosen.
+//!
+//! It is the register allocator's job to make sure that the register constraints on value operands
+//! are satisfied.
+
+use crate::binemit::CodeOffset;
+use crate::ir::{Function, Inst, ValueLoc};
+use crate::isa::{RegClass, RegUnit};
+use crate::regalloc::RegDiversions;
+
+/// Register constraint for a single value operand or instruction result.
+#[derive(PartialEq, Debug)]
+pub struct OperandConstraint {
+    /// The kind of constraint.
+    pub kind: ConstraintKind,
+
+    /// The register class of the operand.
+    ///
+    /// This applies to all kinds of constraints, but with slightly different meaning.
+    pub regclass: RegClass,
+}
+
+impl OperandConstraint {
+    /// Check if this operand constraint is satisfied by the given value location.
+    /// For tied constraints, this only checks the register class, not that the
+    /// counterpart operand has the same value location.
+    pub fn satisfied(&self, loc: ValueLoc) -> bool {
+        match self.kind {
+            ConstraintKind::Reg | ConstraintKind::Tied(_) => {
+                if let ValueLoc::Reg(reg) = loc {
+                    self.regclass.contains(reg)
+                } else {
+                    false
+                }
+            }
+            ConstraintKind::FixedReg(reg) | ConstraintKind::FixedTied(reg) => {
+                loc == ValueLoc::Reg(reg) && self.regclass.contains(reg)
+            }
+            ConstraintKind::Stack => {
+                if let ValueLoc::Stack(_) = loc {
+                    true
+                } else {
+                    false
+                }
+            }
+        }
+    }
+}
+
+/// The different kinds of operand constraints.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum ConstraintKind {
+    /// This operand or result must be a register from the given register class.
+    Reg,
+
+    /// This operand or result must be a fixed register.
+    ///
+    /// The constraint's `regclass` field is the top-level register class containing the fixed
+    /// register.
+    FixedReg(RegUnit),
+
+    /// This result value must use the same register as an input value operand.
+    ///
+    /// The associated number is the index of the input value operand this result is tied to. The
+    /// constraint's `regclass` field is the same as the tied operand's register class.
+    ///
+    /// When an (in, out) operand pair is tied, this constraint kind appears in both the `ins` and
+    /// the `outs` arrays. The constraint for the in operand is `Tied(out)`, and the constraint for
+    /// the out operand is `Tied(in)`.
+    Tied(u8),
+
+    /// This operand must be a fixed register, and it has a tied counterpart.
+    ///
+    /// This works just like `FixedReg`, but additionally indicates that there are identical
+    /// input/output operands for this fixed register. For an input operand, this means that the
+    /// value will be clobbered by the instruction
+    FixedTied(RegUnit),
+
+    /// This operand must be a value in a stack slot.
+    ///
+    /// The constraint's `regclass` field is the register class that would normally be used to load
+    /// and store values of this type.
+    Stack,
+}
+
+/// Value operand constraints for an encoding recipe.
+#[derive(PartialEq, Clone)]
+pub struct RecipeConstraints {
+    /// Constraints for the instruction's fixed value operands.
+    ///
+    /// If the instruction takes a variable number of operands, the register constraints for those
+    /// operands must be computed dynamically.
+    ///
+    /// - For branches and jumps, EBB arguments must match the expectations of the destination EBB.
+    /// - For calls and returns, the calling convention ABI specifies constraints.
+    pub ins: &'static [OperandConstraint],
+
+    /// Constraints for the instruction's fixed results.
+    ///
+    /// If the instruction produces a variable number of results, it's probably a call and the
+    /// constraints must be derived from the calling convention ABI.
+    pub outs: &'static [OperandConstraint],
+
+    /// Are any of the input constraints `FixedReg`?
+    pub fixed_ins: bool,
+
+    /// Are any of the output constraints `FixedReg`?
+    pub fixed_outs: bool,
+
+    /// Are there any tied operands?
+    pub tied_ops: bool,
+
+    /// Does this instruction clobber the CPU flags?
+    ///
+    /// When true, SSA values of type `iflags` or `fflags` can not be live across the instruction.
+    pub clobbers_flags: bool,
+}
+
+impl RecipeConstraints {
+    /// Check that these constraints are satisfied by the operands on `inst`.
+    pub fn satisfied(&self, inst: Inst, divert: &RegDiversions, func: &Function) -> bool {
+        for (&arg, constraint) in func.dfg.inst_args(inst).iter().zip(self.ins) {
+            let loc = divert.get(arg, &func.locations);
+
+            if let ConstraintKind::Tied(out_index) = constraint.kind {
+                let out_val = func.dfg.inst_results(inst)[out_index as usize];
+                let out_loc = func.locations[out_val];
+                if loc != out_loc {
+                    return false;
+                }
+            }
+
+            if !constraint.satisfied(loc) {
+                return false;
+            }
+        }
+
+        for (&arg, constraint) in func.dfg.inst_results(inst).iter().zip(self.outs) {
+            let loc = divert.get(arg, &func.locations);
+            if !constraint.satisfied(loc) {
+                return false;
+            }
+        }
+
+        true
+    }
+}
+
+/// Constraints on the range of a branch instruction.
+///
+/// A branch instruction usually encodes its destination as a signed n-bit offset from an origin.
+/// The origin depends on the ISA and the specific instruction:
+///
+/// - RISC-V and ARM Aarch64 use the address of the branch instruction, `origin = 0`.
+/// - x86 uses the address of the instruction following the branch, `origin = 2` for a 2-byte
+///   branch instruction.
+/// - ARM's A32 encoding uses the address of the branch instruction + 8 bytes, `origin = 8`.
+#[derive(Clone, Copy, Debug)]
+pub struct BranchRange {
+    /// Offset in bytes from the address of the branch instruction to the origin used for computing
+    /// the branch displacement. This is the destination of a branch that encodes a 0 displacement.
+    pub origin: u8,
+
+    /// Number of bits in the signed byte displacement encoded in the instruction. This does not
+    /// account for branches that can only target aligned addresses.
+    pub bits: u8,
+}
+
+impl BranchRange {
+    /// Determine if this branch range can represent the range from `branch` to `dest`, where
+    /// `branch` is the code offset of the branch instruction itself and `dest` is the code offset
+    /// of the destination EBB header.
+    ///
+    /// This method does not detect if the range is larger than 2 GB.
+    pub fn contains(self, branch: CodeOffset, dest: CodeOffset) -> bool {
+        let d = dest.wrapping_sub(branch + CodeOffset::from(self.origin)) as i32;
+        let s = 32 - self.bits;
+        d == d << s >> s
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn branch_range() {
+        // ARM T1 branch.
+        let t1 = BranchRange { origin: 4, bits: 9 };
+        assert!(t1.contains(0, 0));
+        assert!(t1.contains(0, 2));
+        assert!(t1.contains(2, 0));
+        assert!(t1.contains(1000, 1000));
+
+        // Forward limit.
+        assert!(t1.contains(1000, 1258));
+        assert!(!t1.contains(1000, 1260));
+
+        // Backward limit
+        assert!(t1.contains(1000, 748));
+        assert!(!t1.contains(1000, 746));
+    }
+}

cranelift/codegen/src/isa/enc_tables.rs

@@ -0,0 +1,292 @@
+//! Support types for generated encoding tables.
+//!
+//! This module contains types and functions for working with the encoding tables generated by
+//! `cranelift-codegen/meta-python/gen_encoding.py`.
+
+use crate::constant_hash::{probe, Table};
+use crate::ir::{Function, InstructionData, Opcode, Type};
+use crate::isa::{Encoding, Legalize};
+use crate::settings::PredicateView;
+use core::ops::Range;
+
+/// A recipe predicate.
+///
+/// This is a predicate function capable of testing ISA and instruction predicates simultaneously.
+///
+/// A None predicate is always satisfied.
+pub type RecipePredicate = Option<fn(PredicateView, &InstructionData) -> bool>;
+
+/// An instruction predicate.
+///
+/// This is a predicate function that needs to be tested in addition to the recipe predicate. It
+/// can't depend on ISA settings.
+pub type InstPredicate = fn(&Function, &InstructionData) -> bool;
+
+/// Legalization action to perform when no encoding can be found for an instruction.
+///
+/// This is an index into an ISA-specific table of legalization actions.
+pub type LegalizeCode = u8;
+
+/// Level 1 hash table entry.
+///
+/// One level 1 hash table is generated per CPU mode. This table is keyed by the controlling type
+/// variable, using `INVALID` for non-polymorphic instructions.
+///
+/// The hash table values are references to level 2 hash tables, encoded as an offset in `LEVEL2`
+/// where the table begins, and the binary logarithm of its length. All the level 2 hash tables
+/// have a power-of-two size.
+///
+/// Entries are generic over the offset type. It will typically be `u32` or `u16`, depending on the
+/// size of the `LEVEL2` table.
+///
+/// Empty entries are encoded with a `!0` value for `log2len` which will always be out of range.
+/// Entries that have a `legalize` value but no level 2 table have an `offset` field that is out of
+/// bounds.
+pub struct Level1Entry<OffT: Into<u32> + Copy> {
+    pub ty: Type,
+    pub log2len: u8,
+    pub legalize: LegalizeCode,
+    pub offset: OffT,
+}
+
+impl<OffT: Into<u32> + Copy> Level1Entry<OffT> {
+    /// Get the level 2 table range indicated by this entry.
+    fn range(&self) -> Range<usize> {
+        let b = self.offset.into() as usize;
+        b..b + (1 << self.log2len)
+    }
+}
+
+impl<OffT: Into<u32> + Copy> Table<Type> for [Level1Entry<OffT>] {
+    fn len(&self) -> usize {
+        self.len()
+    }
+
+    fn key(&self, idx: usize) -> Option<Type> {
+        if self[idx].log2len != !0 {
+            Some(self[idx].ty)
+        } else {
+            None
+        }
+    }
+}
+
+/// Level 2 hash table entry.
+///
+/// The second level hash tables are keyed by `Opcode`, and contain an offset into the `ENCLISTS`
+/// table where the encoding recipes for the instruction are stored.
+///
+/// Entries are generic over the offset type which depends on the size of `ENCLISTS`. A `u16`
+/// offset allows the entries to be only 32 bits each. There is no benefit to dropping down to `u8`
+/// for tiny ISAs. The entries won't shrink below 32 bits since the opcode is expected to be 16
+/// bits.
+///
+/// Empty entries are encoded with a `NotAnOpcode` `opcode` field.
+pub struct Level2Entry<OffT: Into<u32> + Copy> {
+    pub opcode: Option<Opcode>,
+    pub offset: OffT,
+}
+
+impl<OffT: Into<u32> + Copy> Table<Opcode> for [Level2Entry<OffT>] {
+    fn len(&self) -> usize {
+        self.len()
+    }
+
+    fn key(&self, idx: usize) -> Option<Opcode> {
+        self[idx].opcode
+    }
+}
+
+/// Two-level hash table lookup and iterator construction.
+///
+/// Given the controlling type variable and instruction opcode, find the corresponding encoding
+/// list.
+///
+/// Returns an iterator that produces legal encodings for `inst`.
+pub fn lookup_enclist<'a, OffT1, OffT2>(
+    ctrl_typevar: Type,
+    inst: &'a InstructionData,
+    func: &'a Function,
+    level1_table: &'static [Level1Entry<OffT1>],
+    level2_table: &'static [Level2Entry<OffT2>],
+    enclist: &'static [EncListEntry],
+    legalize_actions: &'static [Legalize],
+    recipe_preds: &'static [RecipePredicate],
+    inst_preds: &'static [InstPredicate],
+    isa_preds: PredicateView<'a>,
+) -> Encodings<'a>
+where
+    OffT1: Into<u32> + Copy,
+    OffT2: Into<u32> + Copy,
+{
+    let (offset, legalize) = match probe(level1_table, ctrl_typevar, ctrl_typevar.index()) {
+        Err(l1idx) => {
+            // No level 1 entry found for the type.
+            // We have a sentinel entry with the default legalization code.
+            (!0, level1_table[l1idx].legalize)
+        }
+        Ok(l1idx) => {
+            // We have a valid level 1 entry for this type.
+            let l1ent = &level1_table[l1idx];
+            let offset = match level2_table.get(l1ent.range()) {
+                Some(l2tab) => {
+                    let opcode = inst.opcode();
+                    match probe(l2tab, opcode, opcode as usize) {
+                        Ok(l2idx) => l2tab[l2idx].offset.into() as usize,
+                        Err(_) => !0,
+                    }
+                }
+                // The l1ent range is invalid. This means that we just have a customized
+                // legalization code for this type. The level 2 table is empty.
+                None => !0,
+            };
+            (offset, l1ent.legalize)
+        }
+    };
+
+    // Now we have an offset into `enclist` that is `!0` when no encoding list could be found.
+    // The default legalization code is always valid.
+    Encodings::new(
+        offset,
+        legalize,
+        inst,
+        func,
+        enclist,
+        legalize_actions,
+        recipe_preds,
+        inst_preds,
+        isa_preds,
+    )
+}
+
+/// Encoding list entry.
+///
+/// Encoding lists are represented as sequences of u16 words.
+pub type EncListEntry = u16;
+
+/// Number of bits used to represent a predicate. c.f. `meta-python/gen_encoding.py`.
+const PRED_BITS: u8 = 12;
+const PRED_MASK: usize = (1 << PRED_BITS) - 1;
+/// First code word representing a predicate check. c.f. `meta-python/gen_encoding.py`.
+const PRED_START: usize = 0x1000;
+
+/// An iterator over legal encodings for the instruction.
+pub struct Encodings<'a> {
+    // Current offset into `enclist`, or out of bounds after we've reached the end.
+    offset: usize,
+    // Legalization code to use of no encoding is found.
+    legalize: LegalizeCode,
+    inst: &'a InstructionData,
+    func: &'a Function,
+    enclist: &'static [EncListEntry],
+    legalize_actions: &'static [Legalize],
+    recipe_preds: &'static [RecipePredicate],
+    inst_preds: &'static [InstPredicate],
+    isa_preds: PredicateView<'a>,
+}
+
+impl<'a> Encodings<'a> {
+    /// Creates a new instance of `Encodings`.
+    ///
+    /// This iterator provides search for encodings that applies to the given instruction. The
+    /// encoding lists are laid out such that first call to `next` returns valid entry in the list
+    /// or `None`.
+    pub fn new(
+        offset: usize,
+        legalize: LegalizeCode,
+        inst: &'a InstructionData,
+        func: &'a Function,
+        enclist: &'static [EncListEntry],
+        legalize_actions: &'static [Legalize],
+        recipe_preds: &'static [RecipePredicate],
+        inst_preds: &'static [InstPredicate],
+        isa_preds: PredicateView<'a>,
+    ) -> Self {
+        Encodings {
+            offset,
+            inst,
+            func,
+            legalize,
+            isa_preds,
+            recipe_preds,
+            inst_preds,
+            enclist,
+            legalize_actions,
+        }
+    }
+
+    /// Get the legalization action that caused the enumeration of encodings to stop.
+    /// This can be the default legalization action for the type or a custom code for the
+    /// instruction.
+    ///
+    /// This method must only be called after the iterator returns `None`.
+    pub fn legalize(&self) -> Legalize {
+        debug_assert_eq!(self.offset, !0, "Premature Encodings::legalize()");
+        self.legalize_actions[self.legalize as usize]
+    }
+
+    /// Check if the `rpred` recipe predicate is satisfied.
+    fn check_recipe(&self, rpred: RecipePredicate) -> bool {
+        match rpred {
+            Some(p) => p(self.isa_preds, self.inst),
+            None => true,
+        }
+    }
+
+    /// Check an instruction or isa predicate.
+    fn check_pred(&self, pred: usize) -> bool {
+        if let Some(&p) = self.inst_preds.get(pred) {
+            p(self.func, self.inst)
+        } else {
+            let pred = pred - self.inst_preds.len();
+            self.isa_preds.test(pred)
+        }
+    }
+}
+
+impl<'a> Iterator for Encodings<'a> {
+    type Item = Encoding;
+
+    fn next(&mut self) -> Option<Encoding> {
+        while let Some(entryref) = self.enclist.get(self.offset) {
+            let entry = *entryref as usize;
+
+            // Check for "recipe+bits".
+            let recipe = entry >> 1;
+            if let Some(&rpred) = self.recipe_preds.get(recipe) {
+                let bits = self.offset + 1;
+                if entry & 1 == 0 {
+                    self.offset += 2; // Next entry.
+                } else {
+                    self.offset = !0; // Stop.
+                }
+                if self.check_recipe(rpred) {
+                    return Some(Encoding::new(recipe as u16, self.enclist[bits]));
+                }
+                continue;
+            }
+
+            // Check for "stop with legalize".
+            if entry < PRED_START {
+                self.legalize = (entry - 2 * self.recipe_preds.len()) as LegalizeCode;
+                self.offset = !0; // Stop.
+                return None;
+            }
+
+            // Finally, this must be a predicate entry.
+            let pred_entry = entry - PRED_START;
+            let skip = pred_entry >> PRED_BITS;
+            let pred = pred_entry & PRED_MASK;
+
+            if self.check_pred(pred) {
+                self.offset += 1;
+            } else if skip == 0 {
+                self.offset = !0; // Stop.
+                return None;
+            } else {
+                self.offset += 1 + skip;
+            }
+        }
+        None
+    }
+}

cranelift/codegen/src/isa/encoding.rs

@@ -0,0 +1,157 @@
+//! The `Encoding` struct.
+
+use crate::binemit::CodeOffset;
+use crate::ir::{Function, Inst};
+use crate::isa::constraints::{BranchRange, RecipeConstraints};
+use crate::regalloc::RegDiversions;
+use core::fmt;
+
+/// Bits needed to encode an instruction as binary machine code.
+///
+/// The encoding consists of two parts, both specific to the target ISA: An encoding *recipe*, and
+/// encoding *bits*. The recipe determines the native instruction format and the mapping of
+/// operands to encoded bits. The encoding bits provide additional information to the recipe,
+/// typically parts of the opcode.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct Encoding {
+    recipe: u16,
+    bits: u16,
+}
+
+impl Encoding {
+    /// Create a new `Encoding` containing `(recipe, bits)`.
+    pub fn new(recipe: u16, bits: u16) -> Self {
+        Self { recipe, bits }
+    }
+
+    /// Get the recipe number in this encoding.
+    pub fn recipe(self) -> usize {
+        self.recipe as usize
+    }
+
+    /// Get the recipe-specific encoding bits.
+    pub fn bits(self) -> u16 {
+        self.bits
+    }
+
+    /// Is this a legal encoding, or the default placeholder?
+    pub fn is_legal(self) -> bool {
+        self != Self::default()
+    }
+}
+
+/// The default encoding is the illegal one.
+impl Default for Encoding {
+    fn default() -> Self {
+        Self::new(0xffff, 0xffff)
+    }
+}
+
+/// ISA-independent display of an encoding.
+impl fmt::Display for Encoding {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        if self.is_legal() {
+            write!(f, "{}#{:02x}", self.recipe, self.bits)
+        } else {
+            write!(f, "-")
+        }
+    }
+}
+
+/// Temporary object that holds enough context to properly display an encoding.
+/// This is meant to be created by `EncInfo::display()`.
+pub struct DisplayEncoding {
+    pub encoding: Encoding,
+    pub recipe_names: &'static [&'static str],
+}
+
+impl fmt::Display for DisplayEncoding {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        if self.encoding.is_legal() {
+            write!(
+                f,
+                "{}#{:02x}",
+                self.recipe_names[self.encoding.recipe()],
+                self.encoding.bits
+            )
+        } else {
+            write!(f, "-")
+        }
+    }
+}
+
+type SizeCalculatorFn = fn(&RecipeSizing, Inst, &RegDiversions, &Function) -> u8;
+
+/// Returns the base size of the Recipe, assuming it's fixed. This is the default for most
+/// encodings; others can be variable and longer than this base size, depending on the registers
+/// they're using and use a different function, specific per platform.
+pub fn base_size(sizing: &RecipeSizing, _: Inst, _: &RegDiversions, _: &Function) -> u8 {
+    sizing.base_size
+}
+
+/// Code size information for an encoding recipe.
+///
+/// All encoding recipes correspond to an exact instruction size.
+pub struct RecipeSizing {
+    /// Size in bytes of instructions encoded with this recipe.
+    pub base_size: u8,
+
+    /// Method computing the real instruction's size, given inputs and outputs.
+    pub compute_size: SizeCalculatorFn,
+
+    /// Allowed branch range in this recipe, if any.
+    ///
+    /// All encoding recipes for branches have exact branch range information.
+    pub branch_range: Option<BranchRange>,
+}
+
+/// Information about all the encodings in this ISA.
+#[derive(Clone)]
+pub struct EncInfo {
+    /// Constraints on value operands per recipe.
+    pub constraints: &'static [RecipeConstraints],
+
+    /// Code size information per recipe.
+    pub sizing: &'static [RecipeSizing],
+
+    /// Names of encoding recipes.
+    pub names: &'static [&'static str],
+}
+
+impl EncInfo {
+    /// Get the value operand constraints for `enc` if it is a legal encoding.
+    pub fn operand_constraints(&self, enc: Encoding) -> Option<&'static RecipeConstraints> {
+        self.constraints.get(enc.recipe())
+    }
+
+    /// Create an object that can display an ISA-dependent encoding properly.
+    pub fn display(&self, enc: Encoding) -> DisplayEncoding {
+        DisplayEncoding {
+            encoding: enc,
+            recipe_names: self.names,
+        }
+    }
+
+    /// Get the precise size in bytes of instructions encoded with `enc`.
+    ///
+    /// Returns 0 for illegal encodings.
+    pub fn byte_size(
+        &self,
+        enc: Encoding,
+        inst: Inst,
+        divert: &RegDiversions,
+        func: &Function,
+    ) -> CodeOffset {
+        self.sizing.get(enc.recipe()).map_or(0, |s| {
+            let compute_size = s.compute_size;
+            CodeOffset::from(compute_size(&s, inst, divert, func))
+        })
+    }
+
+    /// Get the branch range that is supported by `enc`, if any.
+    ///
+    /// This will never return `None` for a legal branch encoding.
+    pub fn branch_range(&self, enc: Encoding) -> Option<BranchRange> {
+        self.sizing.get(enc.recipe()).and_then(|s| s.branch_range)
+    }
+}

cranelift/codegen/src/isa/mod.rs

@@ -0,0 +1,375 @@
+//! Instruction Set Architectures.
+//!
+//! The `isa` module provides a `TargetIsa` trait which provides the behavior specialization needed
+//! by the ISA-independent code generator. The sub-modules of this module provide definitions for
+//! the instruction sets that Cranelift can target. Each sub-module has it's own implementation of
+//! `TargetIsa`.
+//!
+//! # Constructing a `TargetIsa` instance
+//!
+//! The target ISA is built from the following information:
+//!
+//! - The name of the target ISA as a string. Cranelift is a cross-compiler, so the ISA to target
+//!   can be selected dynamically. Individual ISAs can be left out when Cranelift is compiled, so a
+//!   string is used to identify the proper sub-module.
+//! - Values for settings that apply to all ISAs. This is represented by a `settings::Flags`
+//!   instance.
+//! - Values for ISA-specific settings.
+//!
+//! The `isa::lookup()` function is the main entry point which returns an `isa::Builder`
+//! appropriate for the requested ISA:
+//!
+//! ```
+//! # extern crate cranelift_codegen;
+//! # #[macro_use] extern crate target_lexicon;
+//! # fn main() {
+//! use cranelift_codegen::isa;
+//! use cranelift_codegen::settings::{self, Configurable};
+//! use std::str::FromStr;
+//! use target_lexicon::Triple;
+//!
+//! let shared_builder = settings::builder();
+//! let shared_flags = settings::Flags::new(shared_builder);
+//!
+//! match isa::lookup(triple!("riscv32")) {
+//!     Err(_) => {
+//!         // The RISC-V target ISA is not available.
+//!     }
+//!     Ok(mut isa_builder) => {
+//!         isa_builder.set("supports_m", "on");
+//!         let isa = isa_builder.finish(shared_flags);
+//!     }
+//! }
+//! # }
+//! ```
+//!
+//! The configured target ISA trait object is a `Box<TargetIsa>` which can be used for multiple
+//! concurrent function compilations.
+
+pub use crate::isa::call_conv::CallConv;
+pub use crate::isa::constraints::{
+    BranchRange, ConstraintKind, OperandConstraint, RecipeConstraints,
+};
+pub use crate::isa::encoding::{base_size, EncInfo, Encoding};
+pub use crate::isa::registers::{regs_overlap, RegClass, RegClassIndex, RegInfo, RegUnit};
+pub use crate::isa::stack::{StackBase, StackBaseMask, StackRef};
+
+use crate::binemit;
+use crate::flowgraph;
+use crate::ir;
+use crate::isa::enc_tables::Encodings;
+use crate::regalloc;
+use crate::result::CodegenResult;
+use crate::settings;
+use crate::settings::SetResult;
+use crate::timing;
+use core::fmt;
+use failure_derive::Fail;
+use std::boxed::Box;
+use target_lexicon::{Architecture, PointerWidth, Triple};
+
+#[cfg(build_riscv)]
+mod riscv;
+
+#[cfg(build_x86)]
+mod x86;
+
+#[cfg(build_arm32)]
+mod arm32;
+
+#[cfg(build_arm64)]
+mod arm64;
+
+mod call_conv;
+mod constraints;
+mod enc_tables;
+mod encoding;
+pub mod registers;
+mod stack;
+
+/// Returns a builder that can create a corresponding `TargetIsa`
+/// or `Err(LookupError::Unsupported)` if not enabled.
+macro_rules! isa_builder {
+    ($module:ident, $name:ident) => {{
+        #[cfg($name)]
+        fn $name(triple: Triple) -> Result<Builder, LookupError> {
+            Ok($module::isa_builder(triple))
+        };
+        #[cfg(not($name))]
+        fn $name(_triple: Triple) -> Result<Builder, LookupError> {
+            Err(LookupError::Unsupported)
+        }
+        $name
+    }};
+}
+
+/// Look for a supported ISA with the given `name`.
+/// Return a builder that can create a corresponding `TargetIsa`.
+pub fn lookup(triple: Triple) -> Result<Builder, LookupError> {
+    match triple.architecture {
+        Architecture::Riscv32 | Architecture::Riscv64 => isa_builder!(riscv, build_riscv)(triple),
+        Architecture::I386 | Architecture::I586 | Architecture::I686 | Architecture::X86_64 => {
+            isa_builder!(x86, build_x86)(triple)
+        }
+        Architecture::Thumbv6m
+        | Architecture::Thumbv7em
+        | Architecture::Thumbv7m
+        | Architecture::Arm
+        | Architecture::Armv4t
+        | Architecture::Armv5te
+        | Architecture::Armv7
+        | Architecture::Armv7s => isa_builder!(arm32, build_arm32)(triple),
+        Architecture::Aarch64 => isa_builder!(arm64, build_arm64)(triple),
+        _ => Err(LookupError::Unsupported),
+    }
+}
+
+/// Describes reason for target lookup failure
+#[derive(Fail, PartialEq, Eq, Copy, Clone, Debug)]
+pub enum LookupError {
+    /// Support for this target was disabled in the current build.
+    #[fail(display = "Support for this target is disabled")]
+    SupportDisabled,
+
+    /// Support for this target has not yet been implemented.
+    #[fail(display = "Support for this target has not been implemented yet")]
+    Unsupported,
+}
+
+/// Builder for a `TargetIsa`.
+/// Modify the ISA-specific settings before creating the `TargetIsa` trait object with `finish`.
+pub struct Builder {
+    triple: Triple,
+    setup: settings::Builder,
+    constructor: fn(Triple, settings::Flags, settings::Builder) -> Box<TargetIsa>,
+}
+
+impl Builder {
+    /// Combine the ISA-specific settings with the provided ISA-independent settings and allocate a
+    /// fully configured `TargetIsa` trait object.
+    pub fn finish(self, shared_flags: settings::Flags) -> Box<TargetIsa> {
+        (self.constructor)(self.triple, shared_flags, self.setup)
+    }
+}
+
+impl settings::Configurable for Builder {
+    fn set(&mut self, name: &str, value: &str) -> SetResult<()> {
+        self.setup.set(name, value)
+    }
+
+    fn enable(&mut self, name: &str) -> SetResult<()> {
+        self.setup.enable(name)
+    }
+}
+
+/// After determining that an instruction doesn't have an encoding, how should we proceed to
+/// legalize it?
+///
+/// The `Encodings` iterator returns a legalization function to call.
+pub type Legalize =
+    fn(ir::Inst, &mut ir::Function, &mut flowgraph::ControlFlowGraph, &TargetIsa) -> bool;
+
+/// This struct provides information that a frontend may need to know about a target to
+/// produce Cranelift IR for the target.
+#[derive(Clone, Copy)]
+pub struct TargetFrontendConfig {
+    /// The default calling convention of the target.
+    pub default_call_conv: CallConv,
+
+    /// The pointer width of the target.
+    pub pointer_width: PointerWidth,
+}
+
+impl TargetFrontendConfig {
+    /// Get the pointer type of this target.
+    pub fn pointer_type(self) -> ir::Type {
+        ir::Type::int(u16::from(self.pointer_bits())).unwrap()
+    }
+
+    /// Get the width of pointers on this target, in units of bits.
+    pub fn pointer_bits(self) -> u8 {
+        self.pointer_width.bits()
+    }
+
+    /// Get the width of pointers on this target, in units of bytes.
+    pub fn pointer_bytes(self) -> u8 {
+        self.pointer_width.bytes()
+    }
+}
+
+/// Methods that are specialized to a target ISA. Implies a Display trait that shows the
+/// shared flags, as well as any isa-specific flags.
+pub trait TargetIsa: fmt::Display + Sync {
+    /// Get the name of this ISA.
+    fn name(&self) -> &'static str;
+
+    /// Get the target triple that was used to make this trait object.
+    fn triple(&self) -> &Triple;
+
+    /// Get the ISA-independent flags that were used to make this trait object.
+    fn flags(&self) -> &settings::Flags;
+
+    /// Get the default calling convention of this target.
+    fn default_call_conv(&self) -> CallConv {
+        CallConv::triple_default(self.triple())
+    }
+
+    /// Get the pointer type of this ISA.
+    fn pointer_type(&self) -> ir::Type {
+        ir::Type::int(u16::from(self.pointer_bits())).unwrap()
+    }
+
+    /// Get the width of pointers on this ISA.
+    fn pointer_width(&self) -> PointerWidth {
+        self.triple().pointer_width().unwrap()
+    }
+
+    /// Get the width of pointers on this ISA, in units of bits.
+    fn pointer_bits(&self) -> u8 {
+        self.pointer_width().bits()
+    }
+
+    /// Get the width of pointers on this ISA, in units of bytes.
+    fn pointer_bytes(&self) -> u8 {
+        self.pointer_width().bytes()
+    }
+
+    /// Get the information needed by frontends producing Cranelift IR.
+    fn frontend_config(&self) -> TargetFrontendConfig {
+        TargetFrontendConfig {
+            default_call_conv: self.default_call_conv(),
+            pointer_width: self.pointer_width(),
+        }
+    }
+
+    /// Does the CPU implement scalar comparisons using a CPU flags register?
+    fn uses_cpu_flags(&self) -> bool {
+        false
+    }
+
+    /// Does the CPU implement multi-register addressing?
+    fn uses_complex_addresses(&self) -> bool {
+        false
+    }
+
+    /// Get a data structure describing the registers in this ISA.
+    fn register_info(&self) -> RegInfo;
+
+    /// Returns an iterator over legal encodings for the instruction.
+    fn legal_encodings<'a>(
+        &'a self,
+        func: &'a ir::Function,
+        inst: &'a ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Encodings<'a>;
+
+    /// Encode an instruction after determining it is legal.
+    ///
+    /// If `inst` can legally be encoded in this ISA, produce the corresponding `Encoding` object.
+    /// Otherwise, return `Legalize` action.
+    ///
+    /// This is also the main entry point for determining if an instruction is legal.
+    fn encode(
+        &self,
+        func: &ir::Function,
+        inst: &ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Result<Encoding, Legalize> {
+        let mut iter = self.legal_encodings(func, inst, ctrl_typevar);
+        iter.next().ok_or_else(|| iter.legalize())
+    }
+
+    /// Get a data structure describing the instruction encodings in this ISA.
+    fn encoding_info(&self) -> EncInfo;
+
+    /// Legalize a function signature.
+    ///
+    /// This is used to legalize both the signature of the function being compiled and any called
+    /// functions. The signature should be modified by adding `ArgumentLoc` annotations to all
+    /// arguments and return values.
+    ///
+    /// Arguments with types that are not supported by the ABI can be expanded into multiple
+    /// arguments:
+    ///
+    /// - Integer types that are too large to fit in a register can be broken into multiple
+    ///   arguments of a smaller integer type.
+    /// - Floating point types can be bit-cast to an integer type of the same size, and possible
+    ///   broken into smaller integer types.
+    /// - Vector types can be bit-cast and broken down into smaller vectors or scalars.
+    ///
+    /// The legalizer will adapt argument and return values as necessary at all ABI boundaries.
+    ///
+    /// When this function is called to legalize the signature of the function currently being
+    /// compiled, `current` is true. The legalized signature can then also contain special purpose
+    /// arguments and return values such as:
+    ///
+    /// - A `link` argument representing the link registers on RISC architectures that don't push
+    ///   the return address on the stack.
+    /// - A `link` return value which will receive the value that was passed to the `link`
+    ///   argument.
+    /// - An `sret` argument can be added if one wasn't present already. This is necessary if the
+    ///   signature returns more values than registers are available for returning values.
+    /// - An `sret` return value can be added if the ABI requires a function to return its `sret`
+    ///   argument in a register.
+    ///
+    /// Arguments and return values for the caller's frame pointer and other callee-saved registers
+    /// should not be added by this function. These arguments are not added until after register
+    /// allocation.
+    fn legalize_signature(&self, sig: &mut ir::Signature, current: bool);
+
+    /// Get the register class that should be used to represent an ABI argument or return value of
+    /// type `ty`. This should be the top-level register class that contains the argument
+    /// registers.
+    ///
+    /// This function can assume that it will only be asked to provide register classes for types
+    /// that `legalize_signature()` produces in `ArgumentLoc::Reg` entries.
+    fn regclass_for_abi_type(&self, ty: ir::Type) -> RegClass;
+
+    /// Get the set of allocatable registers that can be used when compiling `func`.
+    ///
+    /// This set excludes reserved registers like the stack pointer and other special-purpose
+    /// registers.
+    fn allocatable_registers(&self, func: &ir::Function) -> regalloc::RegisterSet;
+
+    /// Compute the stack layout and insert prologue and epilogue code into `func`.
+    ///
+    /// Return an error if the stack frame is too large.
+    fn prologue_epilogue(&self, func: &mut ir::Function) -> CodegenResult<()> {
+        let _tt = timing::prologue_epilogue();
+        // This default implementation is unlikely to be good enough.
+        use crate::ir::stackslot::{StackOffset, StackSize};
+        use crate::stack_layout::layout_stack;
+
+        let word_size = StackSize::from(self.pointer_bytes());
+
+        // Account for the SpiderMonkey standard prologue pushes.
+        if func.signature.call_conv == CallConv::Baldrdash {
+            let bytes = StackSize::from(self.flags().baldrdash_prologue_words()) * word_size;
+            let mut ss = ir::StackSlotData::new(ir::StackSlotKind::IncomingArg, bytes);
+            ss.offset = Some(-(bytes as StackOffset));
+            func.stack_slots.push(ss);
+        }
+
+        layout_stack(&mut func.stack_slots, word_size)?;
+        Ok(())
+    }
+
+    /// Emit binary machine code for a single instruction into the `sink` trait object.
+    ///
+    /// Note that this will call `put*` methods on the `sink` trait object via its vtable which
+    /// is not the fastest way of emitting code.
+    ///
+    /// This function is under the "testing_hooks" feature, and is only suitable for use by
+    /// test harnesses. It increases code size, and is inefficient.
+    #[cfg(feature = "testing_hooks")]
+    fn emit_inst(
+        &self,
+        func: &ir::Function,
+        inst: ir::Inst,
+        divert: &mut regalloc::RegDiversions,
+        sink: &mut binemit::CodeSink,
+    );
+
+    /// Emit a whole function into memory.
+    fn emit_function_to_memory(&self, func: &ir::Function, sink: &mut binemit::MemoryCodeSink);
+}

cranelift/codegen/src/isa/registers.rs

@@ -0,0 +1,325 @@
+//! Data structures describing the registers in an ISA.
+
+use crate::entity::EntityRef;
+use core::fmt;
+
+/// Register units are the smallest units of register allocation.
+///
+/// Normally there is a 1-1 correspondence between registers and register units, but when an ISA
+/// has aliasing registers, the aliasing can be modeled with registers that cover multiple
+/// register units.
+///
+/// The register allocator will enforce that each register unit only gets used for one thing.
+pub type RegUnit = u16;
+
+/// A bit mask indexed by register units.
+///
+/// The size of this type is determined by the target ISA that has the most register units defined.
+/// Currently that is arm32 which has 64+16 units.
+///
+/// This type should be coordinated with meta-python/cdsl/registers.py.
+pub type RegUnitMask = [u32; 3];
+
+/// A bit mask indexed by register classes.
+///
+/// The size of this type is determined by the ISA with the most register classes.
+///
+/// This type should be coordinated with meta-python/cdsl/isa.py.
+pub type RegClassMask = u32;
+
+/// Guaranteed maximum number of top-level register classes with pressure tracking in any ISA.
+///
+/// This can be increased, but should be coordinated with meta-python/cdsl/isa.py.
+pub const MAX_TRACKED_TOPRCS: usize = 4;
+
+/// The register units in a target ISA are divided into disjoint register banks. Each bank covers a
+/// contiguous range of register units.
+///
+/// The `RegBank` struct provides a static description of a register bank.
+pub struct RegBank {
+    /// The name of this register bank as defined in the ISA's `registers.py` file.
+    pub name: &'static str,
+
+    /// The first register unit in this bank.
+    pub first_unit: RegUnit,
+
+    /// The total number of register units in this bank.
+    pub units: RegUnit,
+
+    /// Array of specially named register units. This array can be shorter than the number of units
+    /// in the bank.
+    pub names: &'static [&'static str],
+
+    /// Name prefix to use for those register units in the bank not covered by the `names` array.
+    /// The remaining register units will be named this prefix followed by their decimal offset in
+    /// the bank. So with a prefix `r`, registers will be named `r8`, `r9`, ...
+    pub prefix: &'static str,
+
+    /// Index of the first top-level register class in this bank.
+    pub first_toprc: usize,
+
+    /// Number of top-level register classes in this bank.
+    ///
+    /// The top-level register classes in a bank are guaranteed to be numbered sequentially from
+    /// `first_toprc`, and all top-level register classes across banks come before any sub-classes.
+    pub num_toprcs: usize,
+
+    /// Is register pressure tracking enabled for this bank?
+    pub pressure_tracking: bool,
+}
+
+impl RegBank {
+    /// Does this bank contain `regunit`?
+    fn contains(&self, regunit: RegUnit) -> bool {
+        regunit >= self.first_unit && regunit - self.first_unit < self.units
+    }
+
+    /// Try to parse a regunit name. The name is not expected to begin with `%`.
+    fn parse_regunit(&self, name: &str) -> Option<RegUnit> {
+        match self.names.iter().position(|&x| x == name) {
+            Some(offset) => {
+                // This is one of the special-cased names.
+                Some(offset as RegUnit)
+            }
+            None => {
+                // Try a regular prefixed name.
+                if name.starts_with(self.prefix) {
+                    name[self.prefix.len()..].parse().ok()
+                } else {
+                    None
+                }
+            }
+        }
+        .and_then(|offset| {
+            if offset < self.units {
+                Some(offset + self.first_unit)
+            } else {
+                None
+            }
+        })
+    }
+
+    /// Write `regunit` to `w`, assuming that it belongs to this bank.
+    /// All regunits are written with a `%` prefix.
+    fn write_regunit(&self, f: &mut fmt::Formatter, regunit: RegUnit) -> fmt::Result {
+        let offset = regunit - self.first_unit;
+        assert!(offset < self.units);
+        if (offset as usize) < self.names.len() {
+            write!(f, "%{}", self.names[offset as usize])
+        } else {
+            write!(f, "%{}{}", self.prefix, offset)
+        }
+    }
+}
+
+/// A register class reference.
+///
+/// All register classes are statically defined in tables generated from the meta descriptions.
+pub type RegClass = &'static RegClassData;
+
+/// Data about a register class.
+///
+/// A register class represents a subset of the registers in a bank. It describes the set of
+/// permitted registers for a register operand in a given encoding of an instruction.
+///
+/// A register class can be a subset of another register class. The top-level register classes are
+/// disjoint.
+pub struct RegClassData {
+    /// The name of the register class.
+    pub name: &'static str,
+
+    /// The index of this class in the ISA's RegInfo description.
+    pub index: u8,
+
+    /// How many register units to allocate per register.
+    pub width: u8,
+
+    /// Index of the register bank this class belongs to.
+    pub bank: u8,
+
+    /// Index of the top-level register class contains this one.
+    pub toprc: u8,
+
+    /// The first register unit in this class.
+    pub first: RegUnit,
+
+    /// Bit-mask of sub-classes of this register class, including itself.
+    ///
+    /// Bits correspond to RC indexes.
+    pub subclasses: RegClassMask,
+
+    /// Mask of register units in the class. If `width > 1`, the mask only has a bit set for the
+    /// first register unit in each allocatable register.
+    pub mask: RegUnitMask,
+
+    /// The global `RegInfo` instance containing this register class.
+    pub info: &'static RegInfo,
+}
+
+impl RegClassData {
+    /// Get the register class index corresponding to the intersection of `self` and `other`.
+    ///
+    /// This register class is guaranteed to exist if the register classes overlap. If the register
+    /// classes don't overlap, returns `None`.
+    pub fn intersect_index(&self, other: RegClass) -> Option<RegClassIndex> {
+        // Compute the set of common subclasses.
+        let mask = self.subclasses & other.subclasses;
+
+        if mask == 0 {
+            // No overlap.
+            None
+        } else {
+            // Register class indexes are topologically ordered, so the largest common subclass has
+            // the smallest index.
+            Some(RegClassIndex(mask.trailing_zeros() as u8))
+        }
+    }
+
+    /// Get the intersection of `self` and `other`.
+    pub fn intersect(&self, other: RegClass) -> Option<RegClass> {
+        self.intersect_index(other).map(|rci| self.info.rc(rci))
+    }
+
+    /// Returns true if `other` is a subclass of this register class.
+    /// A register class is considered to be a subclass of itself.
+    pub fn has_subclass<RCI: Into<RegClassIndex>>(&self, other: RCI) -> bool {
+        self.subclasses & (1 << other.into().0) != 0
+    }
+
+    /// Get the top-level register class containing this class.
+    pub fn toprc(&self) -> RegClass {
+        self.info.rc(RegClassIndex(self.toprc))
+    }
+
+    /// Get a specific register unit in this class.
+    pub fn unit(&self, offset: usize) -> RegUnit {
+        let uoffset = offset * usize::from(self.width);
+        self.first + uoffset as RegUnit
+    }
+
+    /// Does this register class contain `regunit`?
+    pub fn contains(&self, regunit: RegUnit) -> bool {
+        self.mask[(regunit / 32) as usize] & (1u32 << (regunit % 32)) != 0
+    }
+}
+
+impl fmt::Display for RegClassData {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        f.write_str(self.name)
+    }
+}
+
+impl fmt::Debug for RegClassData {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        f.write_str(self.name)
+    }
+}
+
+/// Within an ISA, register classes are uniquely identified by their index.
+impl PartialEq for RegClassData {
+    fn eq(&self, other: &Self) -> bool {
+        self.index == other.index
+    }
+}
+
+/// A small reference to a register class.
+///
+/// Use this when storing register classes in compact data structures. The `RegInfo::rc()` method
+/// can be used to get the real register class reference back.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct RegClassIndex(u8);
+
+impl EntityRef for RegClassIndex {
+    fn new(idx: usize) -> Self {
+        RegClassIndex(idx as u8)
+    }
+
+    fn index(self) -> usize {
+        usize::from(self.0)
+    }
+}
+
+impl From<RegClass> for RegClassIndex {
+    fn from(rc: RegClass) -> Self {
+        RegClassIndex(rc.index)
+    }
+}
+
+impl fmt::Display for RegClassIndex {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "rci{}", self.0)
+    }
+}
+
+/// Test of two registers overlap.
+///
+/// A register is identified as a `(RegClass, RegUnit)` pair. The register class is needed to
+/// determine the width (in regunits) of the register.
+pub fn regs_overlap(rc1: RegClass, reg1: RegUnit, rc2: RegClass, reg2: RegUnit) -> bool {
+    let end1 = reg1 + RegUnit::from(rc1.width);
+    let end2 = reg2 + RegUnit::from(rc2.width);
+    !(end1 <= reg2 || end2 <= reg1)
+}
+
+/// Information about the registers in an ISA.
+///
+/// The `RegUnit` data structure collects all relevant static information about the registers in an
+/// ISA.
+#[derive(Clone)]
+pub struct RegInfo {
+    /// All register banks, ordered by their `first_unit`. The register banks are disjoint, but
+    /// there may be holes of unused register unit numbers between banks due to alignment.
+    pub banks: &'static [RegBank],
+
+    /// All register classes ordered topologically so a sub-class always follows its parent.
+    pub classes: &'static [RegClass],
+}
+
+impl RegInfo {
+    /// Get the register bank holding `regunit`.
+    pub fn bank_containing_regunit(&self, regunit: RegUnit) -> Option<&RegBank> {
+        // We could do a binary search, but most ISAs have only two register banks...
+        self.banks.iter().find(|b| b.contains(regunit))
+    }
+
+    /// Try to parse a regunit name. The name is not expected to begin with `%`.
+    pub fn parse_regunit(&self, name: &str) -> Option<RegUnit> {
+        self.banks
+            .iter()
+            .filter_map(|b| b.parse_regunit(name))
+            .next()
+    }
+
+    /// Make a temporary object that can display a register unit.
+    pub fn display_regunit(&self, regunit: RegUnit) -> DisplayRegUnit {
+        DisplayRegUnit {
+            regunit,
+            reginfo: self,
+        }
+    }
+
+    /// Get the register class corresponding to `idx`.
+    pub fn rc(&self, idx: RegClassIndex) -> RegClass {
+        self.classes[idx.index()]
+    }
+
+    /// Get the top-level register class containing the `idx` class.
+    pub fn toprc(&self, idx: RegClassIndex) -> RegClass {
+        self.classes[self.rc(idx).toprc as usize]
+    }
+}
+
+/// Temporary object that holds enough information to print a register unit.
+pub struct DisplayRegUnit<'a> {
+    regunit: RegUnit,
+    reginfo: &'a RegInfo,
+}
+
+impl<'a> fmt::Display for DisplayRegUnit<'a> {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        match self.reginfo.bank_containing_regunit(self.regunit) {
+            Some(b) => b.write_regunit(f, self.regunit),
+            None => write!(f, "%INVALID{}", self.regunit),
+        }
+    }
+}

cranelift/codegen/src/isa/riscv/abi.rs

@@ -0,0 +1,144 @@
+//! RISC-V ABI implementation.
+//!
+//! This module implements the RISC-V calling convention through the primary `legalize_signature()`
+//! entry point.
+//!
+//! This doesn't support the soft-float ABI at the moment.
+
+use super::registers::{FPR, GPR};
+use super::settings;
+use crate::abi::{legalize_args, ArgAction, ArgAssigner, ValueConversion};
+use crate::ir::{self, AbiParam, ArgumentExtension, ArgumentLoc, ArgumentPurpose, Type};
+use crate::isa::RegClass;
+use crate::regalloc::RegisterSet;
+use core::i32;
+use target_lexicon::Triple;
+
+struct Args {
+    pointer_bits: u8,
+    pointer_bytes: u8,
+    pointer_type: Type,
+    regs: u32,
+    reg_limit: u32,
+    offset: u32,
+}
+
+impl Args {
+    fn new(bits: u8, enable_e: bool) -> Self {
+        Self {
+            pointer_bits: bits,
+            pointer_bytes: bits / 8,
+            pointer_type: Type::int(u16::from(bits)).unwrap(),
+            regs: 0,
+            reg_limit: if enable_e { 6 } else { 8 },
+            offset: 0,
+        }
+    }
+}
+
+impl ArgAssigner for Args {
+    fn assign(&mut self, arg: &AbiParam) -> ArgAction {
+        fn align(value: u32, to: u32) -> u32 {
+            (value + to - 1) & !(to - 1)
+        }
+
+        let ty = arg.value_type;
+
+        // Check for a legal type.
+        // RISC-V doesn't have SIMD at all, so break all vectors down.
+        if ty.is_vector() {
+            return ValueConversion::VectorSplit.into();
+        }
+
+        // Large integers and booleans are broken down to fit in a register.
+        if !ty.is_float() && ty.bits() > u16::from(self.pointer_bits) {
+            // Align registers and stack to a multiple of two pointers.
+            self.regs = align(self.regs, 2);
+            self.offset = align(self.offset, 2 * u32::from(self.pointer_bytes));
+            return ValueConversion::IntSplit.into();
+        }
+
+        // Small integers are extended to the size of a pointer register.
+        if ty.is_int() && ty.bits() < u16::from(self.pointer_bits) {
+            match arg.extension {
+                ArgumentExtension::None => {}
+                ArgumentExtension::Uext => return ValueConversion::Uext(self.pointer_type).into(),
+                ArgumentExtension::Sext => return ValueConversion::Sext(self.pointer_type).into(),
+            }
+        }
+
+        if self.regs < self.reg_limit {
+            // Assign to a register.
+            let reg = if ty.is_float() {
+                FPR.unit(10 + self.regs as usize)
+            } else {
+                GPR.unit(10 + self.regs as usize)
+            };
+            self.regs += 1;
+            ArgumentLoc::Reg(reg).into()
+        } else {
+            // Assign a stack location.
+            let loc = ArgumentLoc::Stack(self.offset as i32);
+            self.offset += u32::from(self.pointer_bytes);
+            debug_assert!(self.offset <= i32::MAX as u32);
+            loc.into()
+        }
+    }
+}
+
+/// Legalize `sig` for RISC-V.
+pub fn legalize_signature(
+    sig: &mut ir::Signature,
+    triple: &Triple,
+    isa_flags: &settings::Flags,
+    current: bool,
+) {
+    let bits = triple.pointer_width().unwrap().bits();
+
+    let mut args = Args::new(bits, isa_flags.enable_e());
+    legalize_args(&mut sig.params, &mut args);
+
+    let mut rets = Args::new(bits, isa_flags.enable_e());
+    legalize_args(&mut sig.returns, &mut rets);
+
+    if current {
+        let ptr = Type::int(u16::from(bits)).unwrap();
+
+        // Add the link register as an argument and return value.
+        //
+        // The `jalr` instruction implementing a return can technically accept the return address
+        // in any register, but a micro-architecture with a return address predictor will only
+        // recognize it as a return if the address is in `x1`.
+        let link = AbiParam::special_reg(ptr, ArgumentPurpose::Link, GPR.unit(1));
+        sig.params.push(link);
+        sig.returns.push(link);
+    }
+}
+
+/// Get register class for a type appearing in a legalized signature.
+pub fn regclass_for_abi_type(ty: Type) -> RegClass {
+    if ty.is_float() {
+        FPR
+    } else {
+        GPR
+    }
+}
+
+pub fn allocatable_registers(_func: &ir::Function, isa_flags: &settings::Flags) -> RegisterSet {
+    let mut regs = RegisterSet::new();
+    regs.take(GPR, GPR.unit(0)); // Hard-wired 0.
+                                 // %x1 is the link register which is available for allocation.
+    regs.take(GPR, GPR.unit(2)); // Stack pointer.
+    regs.take(GPR, GPR.unit(3)); // Global pointer.
+    regs.take(GPR, GPR.unit(4)); // Thread pointer.
+                                 // TODO: %x8 is the frame pointer. Reserve it?
+
+    // Remove %x16 and up for RV32E.
+    if isa_flags.enable_e() {
+        for u in 16..32 {
+            regs.take(GPR, GPR.unit(u));
+        }
+    }
+
+    regs
+}

cranelift/codegen/src/isa/riscv/binemit.rs

@@ -0,0 +1,182 @@
+//! Emitting binary RISC-V machine code.
+
+use crate::binemit::{bad_encoding, CodeSink, Reloc};
+use crate::ir::{Function, Inst, InstructionData};
+use crate::isa::{RegUnit, StackBaseMask, StackRef};
+use crate::predicates::is_signed_int;
+use crate::regalloc::RegDiversions;
+use core::u32;
+
+include!(concat!(env!("OUT_DIR"), "/binemit-riscv.rs"));
+
+/// R-type instructions.
+///
+///   31     24  19  14     11 6
+///   funct7 rs2 rs1 funct3 rd opcode
+///       25  20  15     12  7      0
+///
+/// Encoding bits: `opcode[6:2] | (funct3 << 5) | (funct7 << 8)`.
+fn put_r<CS: CodeSink + ?Sized>(bits: u16, rs1: RegUnit, rs2: RegUnit, rd: RegUnit, sink: &mut CS) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let funct3 = (bits >> 5) & 0x7;
+    let funct7 = (bits >> 8) & 0x7f;
+    let rs1 = u32::from(rs1) & 0x1f;
+    let rs2 = u32::from(rs2) & 0x1f;
+    let rd = u32::from(rd) & 0x1f;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= rd << 7;
+    i |= funct3 << 12;
+    i |= rs1 << 15;
+    i |= rs2 << 20;
+    i |= funct7 << 25;
+
+    sink.put4(i);
+}
+
+/// R-type instructions with a shift amount instead of rs2.
+///
+///   31     25    19  14     11 6
+///   funct7 shamt rs1 funct3 rd opcode
+///       25    20  15     12  7      0
+///
+/// Both funct7 and shamt contribute to bit 25. In RV64, shamt uses it for shifts > 31.
+///
+/// Encoding bits: `opcode[6:2] | (funct3 << 5) | (funct7 << 8)`.
+fn put_rshamt<CS: CodeSink + ?Sized>(
+    bits: u16,
+    rs1: RegUnit,
+    shamt: i64,
+    rd: RegUnit,
+    sink: &mut CS,
+) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let funct3 = (bits >> 5) & 0x7;
+    let funct7 = (bits >> 8) & 0x7f;
+    let rs1 = u32::from(rs1) & 0x1f;
+    let shamt = shamt as u32 & 0x3f;
+    let rd = u32::from(rd) & 0x1f;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= rd << 7;
+    i |= funct3 << 12;
+    i |= rs1 << 15;
+    i |= shamt << 20;
+    i |= funct7 << 25;
+
+    sink.put4(i);
+}
+
+/// I-type instructions.
+///
+///   31  19  14     11 6
+///   imm rs1 funct3 rd opcode
+///    20  15     12  7      0
+///
+/// Encoding bits: `opcode[6:2] | (funct3 << 5)`
+fn put_i<CS: CodeSink + ?Sized>(bits: u16, rs1: RegUnit, imm: i64, rd: RegUnit, sink: &mut CS) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let funct3 = (bits >> 5) & 0x7;
+    let rs1 = u32::from(rs1) & 0x1f;
+    let rd = u32::from(rd) & 0x1f;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= rd << 7;
+    i |= funct3 << 12;
+    i |= rs1 << 15;
+    i |= (imm << 20) as u32;
+
+    sink.put4(i);
+}
+
+/// U-type instructions.
+///
+///   31  11 6
+///   imm rd opcode
+///    12  7      0
+///
+/// Encoding bits: `opcode[6:2] | (funct3 << 5)`
+fn put_u<CS: CodeSink + ?Sized>(bits: u16, imm: i64, rd: RegUnit, sink: &mut CS) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let rd = u32::from(rd) & 0x1f;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= rd << 7;
+    i |= imm as u32 & 0xfffff000;
+
+    sink.put4(i);
+}
+
+/// SB-type branch instructions.
+///
+///   31  24  19  14     11  6
+///   imm rs2 rs1 funct3 imm opcode
+///    25  20  15     12   7      0
+///
+/// Encoding bits: `opcode[6:2] | (funct3 << 5)`
+fn put_sb<CS: CodeSink + ?Sized>(bits: u16, imm: i64, rs1: RegUnit, rs2: RegUnit, sink: &mut CS) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let funct3 = (bits >> 5) & 0x7;
+    let rs1 = u32::from(rs1) & 0x1f;
+    let rs2 = u32::from(rs2) & 0x1f;
+
+    debug_assert!(is_signed_int(imm, 13, 1), "SB out of range {:#x}", imm);
+    let imm = imm as u32;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= funct3 << 12;
+    i |= rs1 << 15;
+    i |= rs2 << 20;
+
+    // The displacement is completely hashed up.
+    i |= ((imm >> 11) & 0x1) << 7;
+    i |= ((imm >> 1) & 0xf) << 8;
+    i |= ((imm >> 5) & 0x3f) << 25;
+    i |= ((imm >> 12) & 0x1) << 31;
+
+    sink.put4(i);
+}
+
+/// UJ-type jump instructions.
+///
+///   31  11 6
+///   imm rd opcode
+///    12  7      0
+///
+/// Encoding bits: `opcode[6:2]`
+fn put_uj<CS: CodeSink + ?Sized>(bits: u16, imm: i64, rd: RegUnit, sink: &mut CS) {
+    let bits = u32::from(bits);
+    let opcode5 = bits & 0x1f;
+    let rd = u32::from(rd) & 0x1f;
+
+    debug_assert!(is_signed_int(imm, 21, 1), "UJ out of range {:#x}", imm);
+    let imm = imm as u32;
+
+    // 0-6: opcode
+    let mut i = 0x3;
+    i |= opcode5 << 2;
+    i |= rd << 7;
+
+    // The displacement is completely hashed up.
+    i |= imm & 0xff000;
+    i |= ((imm >> 11) & 0x1) << 20;
+    i |= ((imm >> 1) & 0x3ff) << 21;
+    i |= ((imm >> 20) & 0x1) << 31;
+
+    sink.put4(i);
+}

cranelift/codegen/src/isa/riscv/enc_tables.rs

@@ -0,0 +1,17 @@
+//! Encoding tables for RISC-V.
+
+use super::registers::*;
+use crate::ir;
+use crate::isa;
+use crate::isa::constraints::*;
+use crate::isa::enc_tables::*;
+use crate::isa::encoding::{base_size, RecipeSizing};
+
+// Include the generated encoding tables:
+// - `LEVEL1_RV32`
+// - `LEVEL1_RV64`
+// - `LEVEL2`
+// - `ENCLIST`
+// - `INFO`
+include!(concat!(env!("OUT_DIR"), "/encoding-riscv.rs"));
+include!(concat!(env!("OUT_DIR"), "/legalize-riscv.rs"));

cranelift/codegen/src/isa/riscv/mod.rs

@@ -0,0 +1,281 @@
+//! RISC-V Instruction Set Architecture.
+
+mod abi;
+mod binemit;
+mod enc_tables;
+mod registers;
+pub mod settings;
+
+use super::super::settings as shared_settings;
+#[cfg(feature = "testing_hooks")]
+use crate::binemit::CodeSink;
+use crate::binemit::{emit_function, MemoryCodeSink};
+use crate::ir;
+use crate::isa::enc_tables::{self as shared_enc_tables, lookup_enclist, Encodings};
+use crate::isa::Builder as IsaBuilder;
+use crate::isa::{EncInfo, RegClass, RegInfo, TargetIsa};
+use crate::regalloc;
+use core::fmt;
+use std::boxed::Box;
+use target_lexicon::{PointerWidth, Triple};
+
+#[allow(dead_code)]
+struct Isa {
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    isa_flags: settings::Flags,
+    cpumode: &'static [shared_enc_tables::Level1Entry<u16>],
+}
+
+/// Get an ISA builder for creating RISC-V targets.
+pub fn isa_builder(triple: Triple) -> IsaBuilder {
+    IsaBuilder {
+        triple,
+        setup: settings::builder(),
+        constructor: isa_constructor,
+    }
+}
+
+fn isa_constructor(
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    builder: shared_settings::Builder,
+) -> Box<TargetIsa> {
+    let level1 = match triple.pointer_width().unwrap() {
+        PointerWidth::U16 => panic!("16-bit RISC-V unrecognized"),
+        PointerWidth::U32 => &enc_tables::LEVEL1_RV32[..],
+        PointerWidth::U64 => &enc_tables::LEVEL1_RV64[..],
+    };
+    Box::new(Isa {
+        triple,
+        isa_flags: settings::Flags::new(&shared_flags, builder),
+        shared_flags,
+        cpumode: level1,
+    })
+}
+
+impl TargetIsa for Isa {
+    fn name(&self) -> &'static str {
+        "riscv"
+    }
+
+    fn triple(&self) -> &Triple {
+        &self.triple
+    }
+
+    fn flags(&self) -> &shared_settings::Flags {
+        &self.shared_flags
+    }
+
+    fn register_info(&self) -> RegInfo {
+        registers::INFO.clone()
+    }
+
+    fn encoding_info(&self) -> EncInfo {
+        enc_tables::INFO.clone()
+    }
+
+    fn legal_encodings<'a>(
+        &'a self,
+        func: &'a ir::Function,
+        inst: &'a ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Encodings<'a> {
+        lookup_enclist(
+            ctrl_typevar,
+            inst,
+            func,
+            self.cpumode,
+            &enc_tables::LEVEL2[..],
+            &enc_tables::ENCLISTS[..],
+            &enc_tables::LEGALIZE_ACTIONS[..],
+            &enc_tables::RECIPE_PREDICATES[..],
+            &enc_tables::INST_PREDICATES[..],
+            self.isa_flags.predicate_view(),
+        )
+    }
+
+    fn legalize_signature(&self, sig: &mut ir::Signature, current: bool) {
+        abi::legalize_signature(sig, &self.triple, &self.isa_flags, current)
+    }
+
+    fn regclass_for_abi_type(&self, ty: ir::Type) -> RegClass {
+        abi::regclass_for_abi_type(ty)
+    }
+
+    fn allocatable_registers(&self, func: &ir::Function) -> regalloc::RegisterSet {
+        abi::allocatable_registers(func, &self.isa_flags)
+    }
+
+    #[cfg(feature = "testing_hooks")]
+    fn emit_inst(
+        &self,
+        func: &ir::Function,
+        inst: ir::Inst,
+        divert: &mut regalloc::RegDiversions,
+        sink: &mut CodeSink,
+    ) {
+        binemit::emit_inst(func, inst, divert, sink)
+    }
+
+    fn emit_function_to_memory(&self, func: &ir::Function, sink: &mut MemoryCodeSink) {
+        emit_function(func, binemit::emit_inst, sink)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::ir::{immediates, types};
+    use crate::ir::{Function, InstructionData, Opcode};
+    use crate::isa;
+    use crate::settings::{self, Configurable};
+    use core::str::FromStr;
+    use std::string::{String, ToString};
+    use target_lexicon::triple;
+
+    fn encstr(isa: &isa::TargetIsa, enc: Result<isa::Encoding, isa::Legalize>) -> String {
+        match enc {
+            Ok(e) => isa.encoding_info().display(e).to_string(),
+            Err(_) => "no encoding".to_string(),
+        }
+    }
+
+    #[test]
+    fn test_64bitenc() {
+        let shared_builder = settings::builder();
+        let shared_flags = settings::Flags::new(shared_builder);
+        let isa = isa::lookup(triple!("riscv64"))
+            .unwrap()
+            .finish(shared_flags);
+
+        let mut func = Function::new();
+        let ebb = func.dfg.make_ebb();
+        let arg64 = func.dfg.append_ebb_param(ebb, types::I64);
+        let arg32 = func.dfg.append_ebb_param(ebb, types::I32);
+
+        // Try to encode iadd_imm.i64 v1, -10.
+        let inst64 = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg64,
+            imm: immediates::Imm64::new(-10),
+        };
+
+        // ADDI is I/0b00100
+        assert_eq!(
+            encstr(&*isa, isa.encode(&func, &inst64, types::I64)),
+            "Ii#04"
+        );
+
+        // Try to encode iadd_imm.i64 v1, -10000.
+        let inst64_large = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg64,
+            imm: immediates::Imm64::new(-10000),
+        };
+
+        // Immediate is out of range for ADDI.
+        assert!(isa.encode(&func, &inst64_large, types::I64).is_err());
+
+        // Create an iadd_imm.i32 which is encodable in RV64.
+        let inst32 = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg32,
+            imm: immediates::Imm64::new(10),
+        };
+
+        // ADDIW is I/0b00110
+        assert_eq!(
+            encstr(&*isa, isa.encode(&func, &inst32, types::I32)),
+            "Ii#06"
+        );
+    }
+
+    // Same as above, but for RV32.
+    #[test]
+    fn test_32bitenc() {
+        let shared_builder = settings::builder();
+        let shared_flags = settings::Flags::new(shared_builder);
+        let isa = isa::lookup(triple!("riscv32"))
+            .unwrap()
+            .finish(shared_flags);
+
+        let mut func = Function::new();
+        let ebb = func.dfg.make_ebb();
+        let arg64 = func.dfg.append_ebb_param(ebb, types::I64);
+        let arg32 = func.dfg.append_ebb_param(ebb, types::I32);
+
+        // Try to encode iadd_imm.i64 v1, -10.
+        let inst64 = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg64,
+            imm: immediates::Imm64::new(-10),
+        };
+
+        // In 32-bit mode, an i64 bit add should be narrowed.
+        assert!(isa.encode(&func, &inst64, types::I64).is_err());
+
+        // Try to encode iadd_imm.i64 v1, -10000.
+        let inst64_large = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg64,
+            imm: immediates::Imm64::new(-10000),
+        };
+
+        // In 32-bit mode, an i64 bit add should be narrowed.
+        assert!(isa.encode(&func, &inst64_large, types::I64).is_err());
+
+        // Create an iadd_imm.i32 which is encodable in RV32.
+        let inst32 = InstructionData::BinaryImm {
+            opcode: Opcode::IaddImm,
+            arg: arg32,
+            imm: immediates::Imm64::new(10),
+        };
+
+        // ADDI is I/0b00100
+        assert_eq!(
+            encstr(&*isa, isa.encode(&func, &inst32, types::I32)),
+            "Ii#04"
+        );
+
+        // Create an imul.i32 which is encodable in RV32, but only when use_m is true.
+        let mul32 = InstructionData::Binary {
+            opcode: Opcode::Imul,
+            args: [arg32, arg32],
+        };
+
+        assert!(isa.encode(&func, &mul32, types::I32).is_err());
+    }
+
+    #[test]
+    fn test_rv32m() {
+        let shared_builder = settings::builder();
+        let shared_flags = settings::Flags::new(shared_builder);
+
+        // Set the supports_m stting which in turn enables the use_m predicate that unlocks
+        // encodings for imul.
+        let mut isa_builder = isa::lookup(triple!("riscv32")).unwrap();
+        isa_builder.enable("supports_m").unwrap();
+
+        let isa = isa_builder.finish(shared_flags);
+
+        let mut func = Function::new();
+        let ebb = func.dfg.make_ebb();
+        let arg32 = func.dfg.append_ebb_param(ebb, types::I32);
+
+        // Create an imul.i32 which is encodable in RV32M.
+        let mul32 = InstructionData::Binary {
+            opcode: Opcode::Imul,
+            args: [arg32, arg32],
+        };
+        assert_eq!(
+            encstr(&*isa, isa.encode(&func, &mul32, types::I32)),
+            "R#10c"
+        );
+    }
+}
+
+impl fmt::Display for Isa {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{}\n{}", self.shared_flags, self.isa_flags)
+    }
+}

cranelift/codegen/src/isa/riscv/registers.rs

@@ -0,0 +1,50 @@
+//! RISC-V register descriptions.
+
+use crate::isa::registers::{RegBank, RegClass, RegClassData, RegInfo, RegUnit};
+
+include!(concat!(env!("OUT_DIR"), "/registers-riscv.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::{FPR, GPR, INFO};
+    use crate::isa::RegUnit;
+    use std::string::{String, ToString};
+
+    #[test]
+    fn unit_encodings() {
+        assert_eq!(INFO.parse_regunit("x0"), Some(0));
+        assert_eq!(INFO.parse_regunit("x31"), Some(31));
+        assert_eq!(INFO.parse_regunit("f0"), Some(32));
+        assert_eq!(INFO.parse_regunit("f31"), Some(63));
+
+        assert_eq!(INFO.parse_regunit("x32"), None);
+        assert_eq!(INFO.parse_regunit("f32"), None);
+    }
+
+    #[test]
+    fn unit_names() {
+        fn uname(ru: RegUnit) -> String {
+            INFO.display_regunit(ru).to_string()
+        }
+
+        assert_eq!(uname(0), "%x0");
+        assert_eq!(uname(1), "%x1");
+        assert_eq!(uname(31), "%x31");
+        assert_eq!(uname(32), "%f0");
+        assert_eq!(uname(33), "%f1");
+        assert_eq!(uname(63), "%f31");
+        assert_eq!(uname(64), "%INVALID64");
+    }
+
+    #[test]
+    fn classes() {
+        assert!(GPR.contains(GPR.unit(0)));
+        assert!(GPR.contains(GPR.unit(31)));
+        assert!(!FPR.contains(GPR.unit(0)));
+        assert!(!FPR.contains(GPR.unit(31)));
+        assert!(!GPR.contains(FPR.unit(0)));
+        assert!(!GPR.contains(FPR.unit(31)));
+        assert!(FPR.contains(FPR.unit(0)));
+        assert!(FPR.contains(FPR.unit(31)));
+    }
+}

cranelift/codegen/src/isa/riscv/settings.rs

@@ -0,0 +1,54 @@
+//! RISC-V Settings.
+
+use crate::settings::{self, detail, Builder};
+use core::fmt;
+
+// Include code generated by `cranelift-codegen/meta-python/gen_settings.py`. This file contains a public
+// `Flags` struct with an impl for all of the settings defined in
+// `cranelift-codegen/meta-python/isa/riscv/settings.py`.
+include!(concat!(env!("OUT_DIR"), "/settings-riscv.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::{builder, Flags};
+    use crate::settings::{self, Configurable};
+    use std::string::ToString;
+
+    #[test]
+    fn display_default() {
+        let shared = settings::Flags::new(settings::builder());
+        let b = builder();
+        let f = Flags::new(&shared, b);
+        assert_eq!(
+            f.to_string(),
+            "[riscv]\n\
+             supports_m = false\n\
+             supports_a = false\n\
+             supports_f = false\n\
+             supports_d = false\n\
+             enable_m = true\n\
+             enable_e = false\n"
+        );
+        // Predicates are not part of the Display output.
+        assert_eq!(f.full_float(), false);
+    }
+
+    #[test]
+    fn predicates() {
+        let shared = settings::Flags::new(settings::builder());
+        let mut b = builder();
+        b.enable("supports_f").unwrap();
+        b.enable("supports_d").unwrap();
+        let f = Flags::new(&shared, b);
+        assert_eq!(f.full_float(), true);
+
+        let mut sb = settings::builder();
+        sb.set("enable_simd", "false").unwrap();
+        let shared = settings::Flags::new(sb);
+        let mut b = builder();
+        b.enable("supports_f").unwrap();
+        b.enable("supports_d").unwrap();
+        let f = Flags::new(&shared, b);
+        assert_eq!(f.full_float(), false);
+    }
+}

cranelift/codegen/src/isa/stack.rs

@@ -0,0 +1,94 @@
+//! Low-level details of stack accesses.
+//!
+//! The `ir::StackSlots` type deals with stack slots and stack frame layout. The `StackRef` type
+//! defined in this module expresses the low-level details of accessing a stack slot from an
+//! encoded instruction.
+
+use crate::ir::stackslot::{StackOffset, StackSlotKind, StackSlots};
+use crate::ir::StackSlot;
+
+/// A method for referencing a stack slot in the current stack frame.
+///
+/// Stack slots are addressed with a constant offset from a base register. The base can be the
+/// stack pointer, the frame pointer, or (in the future) a zone register pointing to an inner zone
+/// of a large stack frame.
+#[derive(Clone, Copy, Debug)]
+pub struct StackRef {
+    /// The base register to use for addressing.
+    pub base: StackBase,
+
+    /// Immediate offset from the base register to the first byte of the stack slot.
+    pub offset: StackOffset,
+}
+
+impl StackRef {
+    /// Get a reference to the stack slot `ss` using one of the base pointers in `mask`.
+    pub fn masked(ss: StackSlot, mask: StackBaseMask, frame: &StackSlots) -> Option<Self> {
+        // Try an SP-relative reference.
+        if mask.contains(StackBase::SP) {
+            return Some(Self::sp(ss, frame));
+        }
+
+        // No reference possible with this mask.
+        None
+    }
+
+    /// Get a reference to `ss` using the stack pointer as a base.
+    pub fn sp(ss: StackSlot, frame: &StackSlots) -> Self {
+        let size = frame
+            .frame_size
+            .expect("Stack layout must be computed before referencing stack slots");
+        let slot = &frame[ss];
+        let offset = if slot.kind == StackSlotKind::OutgoingArg {
+            // Outgoing argument slots have offsets relative to our stack pointer.
+            slot.offset.unwrap()
+        } else {
+            // All other slots have offsets relative to our caller's stack frame.
+            // Offset where SP is pointing. (All ISAs have stacks growing downwards.)
+            let sp_offset = -(size as StackOffset);
+            slot.offset.unwrap() - sp_offset
+        };
+        Self {
+            base: StackBase::SP,
+            offset,
+        }
+    }
+}
+
+/// Generic base register for referencing stack slots.
+///
+/// Most ISAs have a stack pointer and an optional frame pointer, so provide generic names for
+/// those two base pointers.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum StackBase {
+    /// Use the stack pointer.
+    SP = 0,
+
+    /// Use the frame pointer (if one is present).
+    FP = 1,
+
+    /// Use an explicit zone pointer in a general-purpose register.
+    ///
+    /// This feature is not yet implemented.
+    Zone = 2,
+}
+
+/// Bit mask of supported stack bases.
+///
+/// Many instruction encodings can use different base registers while others only work with the
+/// stack pointer, say. A `StackBaseMask` is a bit mask of supported stack bases for a given
+/// instruction encoding.
+///
+/// This behaves like a set of `StackBase` variants.
+///
+/// The internal representation as a `u8` is public because stack base masks are used in constant
+/// tables generated from the Python encoding definitions.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct StackBaseMask(pub u8);
+
+impl StackBaseMask {
+    /// Check if this mask contains the `base` variant.
+    pub fn contains(self, base: StackBase) -> bool {
+        self.0 & (1 << base as usize) != 0
+    }
+}

cranelift/codegen/src/isa/x86/abi.rs

@@ -0,0 +1,579 @@
+//! x86 ABI implementation.
+
+use super::registers::{FPR, GPR, RU};
+use crate::abi::{legalize_args, ArgAction, ArgAssigner, ValueConversion};
+use crate::cursor::{Cursor, CursorPosition, EncCursor};
+use crate::ir;
+use crate::ir::immediates::Imm64;
+use crate::ir::stackslot::{StackOffset, StackSize};
+use crate::ir::{
+    get_probestack_funcref, AbiParam, ArgumentExtension, ArgumentLoc, ArgumentPurpose, InstBuilder,
+    ValueLoc,
+};
+use crate::isa::{CallConv, RegClass, RegUnit, TargetIsa};
+use crate::regalloc::RegisterSet;
+use crate::result::CodegenResult;
+use crate::stack_layout::layout_stack;
+use core::i32;
+use target_lexicon::{PointerWidth, Triple};
+
+/// Argument registers for x86-64
+static ARG_GPRS: [RU; 6] = [RU::rdi, RU::rsi, RU::rdx, RU::rcx, RU::r8, RU::r9];
+
+/// Return value registers.
+static RET_GPRS: [RU; 3] = [RU::rax, RU::rdx, RU::rcx];
+
+/// Argument registers for x86-64, when using windows fastcall
+static ARG_GPRS_WIN_FASTCALL_X64: [RU; 4] = [RU::rcx, RU::rdx, RU::r8, RU::r9];
+
+/// Return value registers for x86-64, when using windows fastcall
+static RET_GPRS_WIN_FASTCALL_X64: [RU; 1] = [RU::rax];
+
+struct Args {
+    pointer_bytes: u8,
+    pointer_bits: u8,
+    pointer_type: ir::Type,
+    gpr: &'static [RU],
+    gpr_used: usize,
+    fpr_limit: usize,
+    fpr_used: usize,
+    offset: u32,
+    call_conv: CallConv,
+}
+
+impl Args {
+    fn new(bits: u8, gpr: &'static [RU], fpr_limit: usize, call_conv: CallConv) -> Self {
+        let offset = if let CallConv::WindowsFastcall = call_conv {
+            // [1] "The caller is responsible for allocating space for parameters to the callee,
+            // and must always allocate sufficient space to store four register parameters"
+            32
+        } else {
+            0
+        };
+
+        Self {
+            pointer_bytes: bits / 8,
+            pointer_bits: bits,
+            pointer_type: ir::Type::int(u16::from(bits)).unwrap(),
+            gpr,
+            gpr_used: 0,
+            fpr_limit,
+            fpr_used: 0,
+            offset,
+            call_conv,
+        }
+    }
+}
+
+impl ArgAssigner for Args {
+    fn assign(&mut self, arg: &AbiParam) -> ArgAction {
+        let ty = arg.value_type;
+
+        // Check for a legal type.
+        // We don't support SIMD yet, so break all vectors down.
+        if ty.is_vector() {
+            return ValueConversion::VectorSplit.into();
+        }
+
+        // Large integers and booleans are broken down to fit in a register.
+        if !ty.is_float() && ty.bits() > u16::from(self.pointer_bits) {
+            return ValueConversion::IntSplit.into();
+        }
+
+        // Small integers are extended to the size of a pointer register.
+        if ty.is_int() && ty.bits() < u16::from(self.pointer_bits) {
+            match arg.extension {
+                ArgumentExtension::None => {}
+                ArgumentExtension::Uext => return ValueConversion::Uext(self.pointer_type).into(),
+                ArgumentExtension::Sext => return ValueConversion::Sext(self.pointer_type).into(),
+            }
+        }
+
+        // Handle special-purpose arguments.
+        if ty.is_int() && self.call_conv == CallConv::Baldrdash {
+            match arg.purpose {
+                // This is SpiderMonkey's `WasmTlsReg`.
+                ArgumentPurpose::VMContext => {
+                    return ArgumentLoc::Reg(if self.pointer_bits == 64 {
+                        RU::r14
+                    } else {
+                        RU::rsi
+                    } as RegUnit)
+                    .into();
+                }
+                // This is SpiderMonkey's `WasmTableCallSigReg`.
+                ArgumentPurpose::SignatureId => return ArgumentLoc::Reg(RU::r10 as RegUnit).into(),
+                _ => {}
+            }
+        }
+
+        // Try to use a GPR.
+        if !ty.is_float() && self.gpr_used < self.gpr.len() {
+            let reg = self.gpr[self.gpr_used] as RegUnit;
+            self.gpr_used += 1;
+            return ArgumentLoc::Reg(reg).into();
+        }
+
+        // Try to use an FPR.
+        if ty.is_float() && self.fpr_used < self.fpr_limit {
+            let reg = FPR.unit(self.fpr_used);
+            self.fpr_used += 1;
+            return ArgumentLoc::Reg(reg).into();
+        }
+
+        // Assign a stack location.
+        let loc = ArgumentLoc::Stack(self.offset as i32);
+        self.offset += u32::from(self.pointer_bytes);
+        debug_assert!(self.offset <= i32::MAX as u32);
+        loc.into()
+    }
+}
+
+/// Legalize `sig`.
+pub fn legalize_signature(sig: &mut ir::Signature, triple: &Triple, _current: bool) {
+    let bits;
+    let mut args;
+
+    match triple.pointer_width().unwrap() {
+        PointerWidth::U16 => panic!(),
+        PointerWidth::U32 => {
+            bits = 32;
+            args = Args::new(bits, &[], 0, sig.call_conv);
+        }
+        PointerWidth::U64 => {
+            bits = 64;
+            args = if sig.call_conv == CallConv::WindowsFastcall {
+                Args::new(bits, &ARG_GPRS_WIN_FASTCALL_X64[..], 4, sig.call_conv)
+            } else {
+                Args::new(bits, &ARG_GPRS[..], 8, sig.call_conv)
+            };
+        }
+    }
+
+    legalize_args(&mut sig.params, &mut args);
+
+    let regs = if sig.call_conv == CallConv::WindowsFastcall {
+        &RET_GPRS_WIN_FASTCALL_X64[..]
+    } else {
+        &RET_GPRS[..]
+    };
+
+    let mut rets = Args::new(bits, regs, 2, sig.call_conv);
+    legalize_args(&mut sig.returns, &mut rets);
+}
+
+/// Get register class for a type appearing in a legalized signature.
+pub fn regclass_for_abi_type(ty: ir::Type) -> RegClass {
+    if ty.is_int() || ty.is_bool() {
+        GPR
+    } else {
+        FPR
+    }
+}
+
+/// Get the set of allocatable registers for `func`.
+pub fn allocatable_registers(_func: &ir::Function, triple: &Triple) -> RegisterSet {
+    let mut regs = RegisterSet::new();
+    regs.take(GPR, RU::rsp as RegUnit);
+    regs.take(GPR, RU::rbp as RegUnit);
+
+    // 32-bit arch only has 8 registers.
+    if triple.pointer_width().unwrap() != PointerWidth::U64 {
+        for i in 8..16 {
+            regs.take(GPR, GPR.unit(i));
+            regs.take(FPR, FPR.unit(i));
+        }
+    }
+
+    regs
+}
+
+/// Get the set of callee-saved registers.
+fn callee_saved_gprs(isa: &TargetIsa, call_conv: CallConv) -> &'static [RU] {
+    match isa.triple().pointer_width().unwrap() {
+        PointerWidth::U16 => panic!(),
+        PointerWidth::U32 => &[RU::rbx, RU::rsi, RU::rdi],
+        PointerWidth::U64 => {
+            if call_conv == CallConv::WindowsFastcall {
+                // "registers RBX, RBP, RDI, RSI, RSP, R12, R13, R14, R15 are considered nonvolatile
+                //  and must be saved and restored by a function that uses them."
+                // as per https://msdn.microsoft.com/en-us/library/6t169e9c.aspx
+                // RSP & RSB are not listed below, since they are restored automatically during
+                // a function call. If that wasn't the case, function calls (RET) would not work.
+                &[
+                    RU::rbx,
+                    RU::rdi,
+                    RU::rsi,
+                    RU::r12,
+                    RU::r13,
+                    RU::r14,
+                    RU::r15,
+                ]
+            } else {
+                &[RU::rbx, RU::r12, RU::r13, RU::r14, RU::r15]
+            }
+        }
+    }
+}
+
+/// Get the set of callee-saved registers that are used.
+fn callee_saved_gprs_used(isa: &TargetIsa, func: &ir::Function) -> RegisterSet {
+    let mut all_callee_saved = RegisterSet::empty();
+    for reg in callee_saved_gprs(isa, func.signature.call_conv) {
+        all_callee_saved.free(GPR, *reg as RegUnit);
+    }
+
+    let mut used = RegisterSet::empty();
+    for value_loc in func.locations.values() {
+        // Note that `value_loc` here contains only a single unit of a potentially multi-unit
+        // register. We don't use registers that overlap each other in the x86 ISA, but in others
+        // we do. So this should not be blindly reused.
+        if let ValueLoc::Reg(ru) = *value_loc {
+            if !used.is_avail(GPR, ru) {
+                used.free(GPR, ru);
+            }
+        }
+    }
+
+    // regmove and regfill instructions may temporarily divert values into other registers,
+    // and these are not reflected in `func.locations`. Scan the function for such instructions
+    // and note which callee-saved registers they use.
+    //
+    // TODO: Consider re-evaluating how regmove/regfill/regspill work and whether it's possible
+    // to avoid this step.
+    for ebb in &func.layout {
+        for inst in func.layout.ebb_insts(ebb) {
+            match func.dfg[inst] {
+                ir::instructions::InstructionData::RegMove { dst, .. }
+                | ir::instructions::InstructionData::RegFill { dst, .. } => {
+                    if !used.is_avail(GPR, dst) {
+                        used.free(GPR, dst);
+                    }
+                }
+                _ => (),
+            }
+        }
+    }
+
+    used.intersect(&all_callee_saved);
+    used
+}
+
+pub fn prologue_epilogue(func: &mut ir::Function, isa: &TargetIsa) -> CodegenResult<()> {
+    match func.signature.call_conv {
+        // For now, just translate fast and cold as system_v.
+        CallConv::Fast | CallConv::Cold | CallConv::SystemV => {
+            system_v_prologue_epilogue(func, isa)
+        }
+        CallConv::WindowsFastcall => fastcall_prologue_epilogue(func, isa),
+        CallConv::Baldrdash => baldrdash_prologue_epilogue(func, isa),
+        CallConv::Probestack => unimplemented!("probestack calling convention"),
+    }
+}
+
+fn baldrdash_prologue_epilogue(func: &mut ir::Function, isa: &TargetIsa) -> CodegenResult<()> {
+    debug_assert!(
+        !isa.flags().probestack_enabled(),
+        "baldrdash does not expect cranelift to emit stack probes"
+    );
+
+    // Baldrdash on 32-bit x86 always aligns its stack pointer to 16 bytes.
+    let stack_align = 16;
+    let word_size = StackSize::from(isa.pointer_bytes());
+    let bytes = StackSize::from(isa.flags().baldrdash_prologue_words()) * word_size;
+
+    let mut ss = ir::StackSlotData::new(ir::StackSlotKind::IncomingArg, bytes);
+    ss.offset = Some(-(bytes as StackOffset));
+    func.stack_slots.push(ss);
+
+    layout_stack(&mut func.stack_slots, stack_align)?;
+    Ok(())
+}
+
+/// Implementation of the fastcall-based Win64 calling convention described at [1]
+/// [1] https://msdn.microsoft.com/en-us/library/ms235286.aspx
+fn fastcall_prologue_epilogue(func: &mut ir::Function, isa: &TargetIsa) -> CodegenResult<()> {
+    if isa.triple().pointer_width().unwrap() != PointerWidth::U64 {
+        panic!("TODO: windows-fastcall: x86-32 not implemented yet");
+    }
+
+    // [1] "The primary exceptions are the stack pointer and malloc or alloca memory,
+    // which are aligned to 16 bytes in order to aid performance"
+    let stack_align = 16;
+
+    let word_size = isa.pointer_bytes() as usize;
+    let reg_type = isa.pointer_type();
+
+    let csrs = callee_saved_gprs_used(isa, func);
+
+    // [1] "Space is allocated on the call stack as a shadow store for callees to save"
+    // This shadow store contains the parameters which are passed through registers (ARG_GPRS)
+    // and is eventually used by the callee to save & restore the values of the arguments.
+    //
+    // [2] https://blogs.msdn.microsoft.com/oldnewthing/20110302-00/?p=11333
+    // "Although the x64 calling convention reserves spill space for parameters,
+    //  you don’t have to use them as such"
+    //
+    // The reserved stack area is composed of:
+    //   return address + frame pointer + all callee-saved registers + shadow space
+    //
+    // Pushing the return address is an implicit function of the `call`
+    // instruction. Each of the others we will then push explicitly. Then we
+    // will adjust the stack pointer to make room for the rest of the required
+    // space for this frame.
+    const SHADOW_STORE_SIZE: i32 = 32;
+    let csr_stack_size = ((csrs.iter(GPR).len() + 2) * word_size) as i32;
+
+    // TODO: eventually use the 32 bytes (shadow store) as spill slot. This currently doesn't work
+    //       since cranelift does not support spill slots before incoming args
+
+    func.create_stack_slot(ir::StackSlotData {
+        kind: ir::StackSlotKind::IncomingArg,
+        size: csr_stack_size as u32,
+        offset: Some(-(SHADOW_STORE_SIZE + csr_stack_size)),
+    });
+
+    let total_stack_size = layout_stack(&mut func.stack_slots, stack_align)? as i32;
+    let local_stack_size = i64::from(total_stack_size - csr_stack_size);
+
+    // Add CSRs to function signature
+    let fp_arg = ir::AbiParam::special_reg(
+        reg_type,
+        ir::ArgumentPurpose::FramePointer,
+        RU::rbp as RegUnit,
+    );
+    func.signature.params.push(fp_arg);
+    func.signature.returns.push(fp_arg);
+
+    for csr in csrs.iter(GPR) {
+        let csr_arg = ir::AbiParam::special_reg(reg_type, ir::ArgumentPurpose::CalleeSaved, csr);
+        func.signature.params.push(csr_arg);
+        func.signature.returns.push(csr_arg);
+    }
+
+    // Set up the cursor and insert the prologue
+    let entry_ebb = func.layout.entry_block().expect("missing entry block");
+    let mut pos = EncCursor::new(func, isa).at_first_insertion_point(entry_ebb);
+    insert_common_prologue(&mut pos, local_stack_size, reg_type, &csrs, isa);
+
+    // Reset the cursor and insert the epilogue
+    let mut pos = pos.at_position(CursorPosition::Nowhere);
+    insert_common_epilogues(&mut pos, local_stack_size, reg_type, &csrs);
+
+    Ok(())
+}
+
+/// Insert a System V-compatible prologue and epilogue.
+fn system_v_prologue_epilogue(func: &mut ir::Function, isa: &TargetIsa) -> CodegenResult<()> {
+    // The original 32-bit x86 ELF ABI had a 4-byte aligned stack pointer, but
+    // newer versions use a 16-byte aligned stack pointer.
+    let stack_align = 16;
+    let pointer_width = isa.triple().pointer_width().unwrap();
+    let word_size = pointer_width.bytes() as usize;
+    let reg_type = ir::Type::int(u16::from(pointer_width.bits())).unwrap();
+
+    let csrs = callee_saved_gprs_used(isa, func);
+
+    // The reserved stack area is composed of:
+    //   return address + frame pointer + all callee-saved registers
+    //
+    // Pushing the return address is an implicit function of the `call`
+    // instruction. Each of the others we will then push explicitly. Then we
+    // will adjust the stack pointer to make room for the rest of the required
+    // space for this frame.
+    let csr_stack_size = ((csrs.iter(GPR).len() + 2) * word_size) as i32;
+    func.create_stack_slot(ir::StackSlotData {
+        kind: ir::StackSlotKind::IncomingArg,
+        size: csr_stack_size as u32,
+        offset: Some(-csr_stack_size),
+    });
+
+    let total_stack_size = layout_stack(&mut func.stack_slots, stack_align)? as i32;
+    let local_stack_size = i64::from(total_stack_size - csr_stack_size);
+
+    // Add CSRs to function signature
+    let fp_arg = ir::AbiParam::special_reg(
+        reg_type,
+        ir::ArgumentPurpose::FramePointer,
+        RU::rbp as RegUnit,
+    );
+    func.signature.params.push(fp_arg);
+    func.signature.returns.push(fp_arg);
+
+    for csr in csrs.iter(GPR) {
+        let csr_arg = ir::AbiParam::special_reg(reg_type, ir::ArgumentPurpose::CalleeSaved, csr);
+        func.signature.params.push(csr_arg);
+        func.signature.returns.push(csr_arg);
+    }
+
+    // Set up the cursor and insert the prologue
+    let entry_ebb = func.layout.entry_block().expect("missing entry block");
+    let mut pos = EncCursor::new(func, isa).at_first_insertion_point(entry_ebb);
+    insert_common_prologue(&mut pos, local_stack_size, reg_type, &csrs, isa);
+
+    // Reset the cursor and insert the epilogue
+    let mut pos = pos.at_position(CursorPosition::Nowhere);
+    insert_common_epilogues(&mut pos, local_stack_size, reg_type, &csrs);
+
+    Ok(())
+}
+
+/// Insert the prologue for a given function.
+/// This is used by common calling conventions such as System V.
+fn insert_common_prologue(
+    pos: &mut EncCursor,
+    stack_size: i64,
+    reg_type: ir::types::Type,
+    csrs: &RegisterSet,
+    isa: &TargetIsa,
+) {
+    if stack_size > 0 {
+        // Check if there is a special stack limit parameter. If so insert stack check.
+        if let Some(stack_limit_arg) = pos.func.special_param(ArgumentPurpose::StackLimit) {
+            // Total stack size is the size of all stack area used by the function, including
+            // pushed CSRs, frame pointer.
+            // Also, the size of a return address, implicitly pushed by a x86 `call` instruction,
+            // also should be accounted for.
+            // TODO: Check if the function body actually contains a `call` instruction.
+            let word_size = isa.pointer_bytes();
+            let total_stack_size = (csrs.iter(GPR).len() + 1 + 1) as i64 * word_size as i64;
+
+            insert_stack_check(pos, total_stack_size, stack_limit_arg);
+        }
+    }
+
+    // Append param to entry EBB
+    let ebb = pos.current_ebb().expect("missing ebb under cursor");
+    let fp = pos.func.dfg.append_ebb_param(ebb, reg_type);
+    pos.func.locations[fp] = ir::ValueLoc::Reg(RU::rbp as RegUnit);
+
+    pos.ins().x86_push(fp);
+    pos.ins()
+        .copy_special(RU::rsp as RegUnit, RU::rbp as RegUnit);
+
+    for reg in csrs.iter(GPR) {
+        // Append param to entry EBB
+        let csr_arg = pos.func.dfg.append_ebb_param(ebb, reg_type);
+
+        // Assign it a location
+        pos.func.locations[csr_arg] = ir::ValueLoc::Reg(reg);
+
+        // Remember it so we can push it momentarily
+        pos.ins().x86_push(csr_arg);
+    }
+
+    // Allocate stack frame storage.
+    if stack_size > 0 {
+        if isa.flags().probestack_enabled()
+            && stack_size > (1 << isa.flags().probestack_size_log2())
+        {
+            // Emit a stack probe.
+            let rax = RU::rax as RegUnit;
+            let rax_val = ir::ValueLoc::Reg(rax);
+
+            // The probestack function expects its input in %rax.
+            let arg = pos.ins().iconst(reg_type, stack_size);
+            pos.func.locations[arg] = rax_val;
+
+            // Call the probestack function.
+            let callee = get_probestack_funcref(pos.func, reg_type, rax, isa);
+
+            // Make the call.
+            let call = if !isa.flags().is_pic()
+                && isa.triple().pointer_width().unwrap() == PointerWidth::U64
+                && !pos.func.dfg.ext_funcs[callee].colocated
+            {
+                // 64-bit non-PIC non-colocated calls need to be legalized to call_indirect.
+                // Use r11 as it may be clobbered under all supported calling conventions.
+                let r11 = RU::r11 as RegUnit;
+                let sig = pos.func.dfg.ext_funcs[callee].signature;
+                let addr = pos.ins().func_addr(reg_type, callee);
+                pos.func.locations[addr] = ir::ValueLoc::Reg(r11);
+                pos.ins().call_indirect(sig, addr, &[arg])
+            } else {
+                // Otherwise just do a normal call.
+                pos.ins().call(callee, &[arg])
+            };
+
+            // If the probestack function doesn't adjust sp, do it ourselves.
+            if !isa.flags().probestack_func_adjusts_sp() {
+                let result = pos.func.dfg.inst_results(call)[0];
+                pos.func.locations[result] = rax_val;
+                pos.ins().adjust_sp_down(result);
+            }
+        } else {
+            // Simply decrement the stack pointer.
+            pos.ins().adjust_sp_down_imm(Imm64::new(stack_size));
+        }
+    }
+}
+
+/// Insert a check that generates a trap if the stack pointer goes
+/// below a value in `stack_limit_arg`.
+fn insert_stack_check(pos: &mut EncCursor, stack_size: i64, stack_limit_arg: ir::Value) {
+    use crate::ir::condcodes::IntCC;
+
+    // Copy `stack_limit_arg` into a %rax and use it for calculating
+    // a SP threshold.
+    let stack_limit_copy = pos.ins().copy(stack_limit_arg);
+    pos.func.locations[stack_limit_copy] = ir::ValueLoc::Reg(RU::rax as RegUnit);
+    let sp_threshold = pos.ins().iadd_imm(stack_limit_copy, stack_size);
+    pos.func.locations[sp_threshold] = ir::ValueLoc::Reg(RU::rax as RegUnit);
+
+    // If the stack pointer currently reaches the SP threshold or below it then after opening
+    // the current stack frame, the current stack pointer will reach the limit.
+    let cflags = pos.ins().ifcmp_sp(sp_threshold);
+    pos.func.locations[cflags] = ir::ValueLoc::Reg(RU::rflags as RegUnit);
+    pos.ins().trapif(
+        IntCC::UnsignedGreaterThanOrEqual,
+        cflags,
+        ir::TrapCode::StackOverflow,
+    );
+}
+
+/// Find all `return` instructions and insert epilogues before them.
+fn insert_common_epilogues(
+    pos: &mut EncCursor,
+    stack_size: i64,
+    reg_type: ir::types::Type,
+    csrs: &RegisterSet,
+) {
+    while let Some(ebb) = pos.next_ebb() {
+        pos.goto_last_inst(ebb);
+        if let Some(inst) = pos.current_inst() {
+            if pos.func.dfg[inst].opcode().is_return() {
+                insert_common_epilogue(inst, stack_size, pos, reg_type, csrs);
+            }
+        }
+    }
+}
+
+/// Insert an epilogue given a specific `return` instruction.
+/// This is used by common calling conventions such as System V.
+fn insert_common_epilogue(
+    inst: ir::Inst,
+    stack_size: i64,
+    pos: &mut EncCursor,
+    reg_type: ir::types::Type,
+    csrs: &RegisterSet,
+) {
+    if stack_size > 0 {
+        pos.ins().adjust_sp_up_imm(Imm64::new(stack_size));
+    }
+
+    // Pop all the callee-saved registers, stepping backward each time to
+    // preserve the correct order.
+    let fp_ret = pos.ins().x86_pop(reg_type);
+    pos.prev_inst();
+
+    pos.func.locations[fp_ret] = ir::ValueLoc::Reg(RU::rbp as RegUnit);
+    pos.func.dfg.append_inst_arg(inst, fp_ret);
+
+    for reg in csrs.iter(GPR) {
+        let csr_ret = pos.ins().x86_pop(reg_type);
+        pos.prev_inst();
+
+        pos.func.locations[csr_ret] = ir::ValueLoc::Reg(reg);
+        pos.func.dfg.append_inst_arg(inst, csr_ret);
+    }
+}

cranelift/codegen/src/isa/x86/binemit.rs

@@ -0,0 +1,342 @@
+//! Emitting binary x86 machine code.
+
+use super::enc_tables::{needs_offset, needs_sib_byte};
+use super::registers::RU;
+use crate::binemit::{bad_encoding, CodeSink, Reloc};
+use crate::ir::condcodes::{CondCode, FloatCC, IntCC};
+use crate::ir::{Ebb, Function, Inst, InstructionData, JumpTable, Opcode, TrapCode};
+use crate::isa::{RegUnit, StackBase, StackBaseMask, StackRef};
+use crate::regalloc::RegDiversions;
+
+include!(concat!(env!("OUT_DIR"), "/binemit-x86.rs"));
+
+// Convert a stack base to the corresponding register.
+fn stk_base(base: StackBase) -> RegUnit {
+    let ru = match base {
+        StackBase::SP => RU::rsp,
+        StackBase::FP => RU::rbp,
+        StackBase::Zone => unimplemented!(),
+    };
+    ru as RegUnit
+}
+
+// Mandatory prefix bytes for Mp* opcodes.
+const PREFIX: [u8; 3] = [0x66, 0xf3, 0xf2];
+
+// Second byte for three-byte opcodes for mm=0b10 and mm=0b11.
+const OP3_BYTE2: [u8; 2] = [0x38, 0x3a];
+
+// A REX prefix with no bits set: 0b0100WRXB.
+const BASE_REX: u8 = 0b0100_0000;
+
+// Create a single-register REX prefix, setting the B bit to bit 3 of the register.
+// This is used for instructions that encode a register in the low 3 bits of the opcode and for
+// instructions that use the ModR/M `reg` field for something else.
+fn rex1(reg_b: RegUnit) -> u8 {
+    let b = ((reg_b >> 3) & 1) as u8;
+    BASE_REX | b
+}
+
+// Create a dual-register REX prefix, setting:
+//
+// REX.B = bit 3 of r/m register, or SIB base register when a SIB byte is present.
+// REX.R = bit 3 of reg register.
+fn rex2(rm: RegUnit, reg: RegUnit) -> u8 {
+    let b = ((rm >> 3) & 1) as u8;
+    let r = ((reg >> 3) & 1) as u8;
+    BASE_REX | b | (r << 2)
+}
+
+// Create a three-register REX prefix, setting:
+//
+// REX.B = bit 3 of r/m register, or SIB base register when a SIB byte is present.
+// REX.R = bit 3 of reg register.
+// REX.X = bit 3 of SIB index register.
+fn rex3(rm: RegUnit, reg: RegUnit, index: RegUnit) -> u8 {
+    let b = ((rm >> 3) & 1) as u8;
+    let r = ((reg >> 3) & 1) as u8;
+    let x = ((index >> 3) & 1) as u8;
+    BASE_REX | b | (x << 1) | (r << 2)
+}
+
+// Emit a REX prefix.
+//
+// The R, X, and B bits are computed from registers using the functions above. The W bit is
+// extracted from `bits`.
+fn rex_prefix<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(rex & 0xf8, BASE_REX);
+    let w = ((bits >> 15) & 1) as u8;
+    sink.put1(rex | (w << 3));
+}
+
+// Emit a single-byte opcode with no REX prefix.
+fn put_op1<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x8f00, 0, "Invalid encoding bits for Op1*");
+    debug_assert_eq!(rex, BASE_REX, "Invalid registers for REX-less Op1 encoding");
+    sink.put1(bits as u8);
+}
+
+// Emit a single-byte opcode with REX prefix.
+fn put_rexop1<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x0f00, 0, "Invalid encoding bits for Op1*");
+    rex_prefix(bits, rex, sink);
+    sink.put1(bits as u8);
+}
+
+// Emit two-byte opcode: 0F XX
+fn put_op2<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x8f00, 0x0400, "Invalid encoding bits for Op2*");
+    debug_assert_eq!(rex, BASE_REX, "Invalid registers for REX-less Op2 encoding");
+    sink.put1(0x0f);
+    sink.put1(bits as u8);
+}
+
+// Emit two-byte opcode: 0F XX with REX prefix.
+fn put_rexop2<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x0f00, 0x0400, "Invalid encoding bits for RexOp2*");
+    rex_prefix(bits, rex, sink);
+    sink.put1(0x0f);
+    sink.put1(bits as u8);
+}
+
+// Emit single-byte opcode with mandatory prefix.
+fn put_mp1<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x8c00, 0, "Invalid encoding bits for Mp1*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    debug_assert_eq!(rex, BASE_REX, "Invalid registers for REX-less Mp1 encoding");
+    sink.put1(bits as u8);
+}
+
+// Emit single-byte opcode with mandatory prefix and REX.
+fn put_rexmp1<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x0c00, 0, "Invalid encoding bits for Mp1*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    rex_prefix(bits, rex, sink);
+    sink.put1(bits as u8);
+}
+
+// Emit two-byte opcode (0F XX) with mandatory prefix.
+fn put_mp2<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x8c00, 0x0400, "Invalid encoding bits for Mp2*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    debug_assert_eq!(rex, BASE_REX, "Invalid registers for REX-less Mp2 encoding");
+    sink.put1(0x0f);
+    sink.put1(bits as u8);
+}
+
+// Emit two-byte opcode (0F XX) with mandatory prefix and REX.
+fn put_rexmp2<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x0c00, 0x0400, "Invalid encoding bits for Mp2*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    rex_prefix(bits, rex, sink);
+    sink.put1(0x0f);
+    sink.put1(bits as u8);
+}
+
+// Emit three-byte opcode (0F 3[8A] XX) with mandatory prefix.
+fn put_mp3<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x8800, 0x0800, "Invalid encoding bits for Mp3*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    debug_assert_eq!(rex, BASE_REX, "Invalid registers for REX-less Mp3 encoding");
+    let mm = (bits >> 10) & 3;
+    sink.put1(0x0f);
+    sink.put1(OP3_BYTE2[(mm - 2) as usize]);
+    sink.put1(bits as u8);
+}
+
+// Emit three-byte opcode (0F 3[8A] XX) with mandatory prefix and REX
+fn put_rexmp3<CS: CodeSink + ?Sized>(bits: u16, rex: u8, sink: &mut CS) {
+    debug_assert_eq!(bits & 0x0800, 0x0800, "Invalid encoding bits for Mp3*");
+    let pp = (bits >> 8) & 3;
+    sink.put1(PREFIX[(pp - 1) as usize]);
+    rex_prefix(bits, rex, sink);
+    let mm = (bits >> 10) & 3;
+    sink.put1(0x0f);
+    sink.put1(OP3_BYTE2[(mm - 2) as usize]);
+    sink.put1(bits as u8);
+}
+
+/// Emit a ModR/M byte for reg-reg operands.
+fn modrm_rr<CS: CodeSink + ?Sized>(rm: RegUnit, reg: RegUnit, sink: &mut CS) {
+    let reg = reg as u8 & 7;
+    let rm = rm as u8 & 7;
+    let mut b = 0b11000000;
+    b |= reg << 3;
+    b |= rm;
+    sink.put1(b);
+}
+
+/// Emit a ModR/M byte where the reg bits are part of the opcode.
+fn modrm_r_bits<CS: CodeSink + ?Sized>(rm: RegUnit, bits: u16, sink: &mut CS) {
+    let reg = (bits >> 12) as u8 & 7;
+    let rm = rm as u8 & 7;
+    let mut b = 0b11000000;
+    b |= reg << 3;
+    b |= rm;
+    sink.put1(b);
+}
+
+/// Emit a mode 00 ModR/M byte. This is a register-indirect addressing mode with no offset.
+/// Registers %rsp and %rbp are invalid for `rm`, %rsp indicates a SIB byte, and %rbp indicates an
+/// absolute immediate 32-bit address.
+fn modrm_rm<CS: CodeSink + ?Sized>(rm: RegUnit, reg: RegUnit, sink: &mut CS) {
+    let reg = reg as u8 & 7;
+    let rm = rm as u8 & 7;
+    let mut b = 0b00000000;
+    b |= reg << 3;
+    b |= rm;
+    sink.put1(b);
+}
+
+/// Emit a mode 00 Mod/RM byte, with a rip-relative displacement in 64-bit mode. Effective address
+/// is calculated by adding displacement to 64-bit rip of next instruction. See intel Sw dev manual
+/// section 2.2.1.6.
+fn modrm_riprel<CS: CodeSink + ?Sized>(reg: RegUnit, sink: &mut CS) {
+    modrm_rm(0b101, reg, sink)
+}
+
+/// Emit a mode 01 ModR/M byte. This is a register-indirect addressing mode with 8-bit
+/// displacement.
+/// Register %rsp is invalid for `rm`. It indicates the presence of a SIB byte.
+fn modrm_disp8<CS: CodeSink + ?Sized>(rm: RegUnit, reg: RegUnit, sink: &mut CS) {
+    let reg = reg as u8 & 7;
+    let rm = rm as u8 & 7;
+    let mut b = 0b01000000;
+    b |= reg << 3;
+    b |= rm;
+    sink.put1(b);
+}
+
+/// Emit a mode 10 ModR/M byte. This is a register-indirect addressing mode with 32-bit
+/// displacement.
+/// Register %rsp is invalid for `rm`. It indicates the presence of a SIB byte.
+fn modrm_disp32<CS: CodeSink + ?Sized>(rm: RegUnit, reg: RegUnit, sink: &mut CS) {
+    let reg = reg as u8 & 7;
+    let rm = rm as u8 & 7;
+    let mut b = 0b10000000;
+    b |= reg << 3;
+    b |= rm;
+    sink.put1(b);
+}
+
+/// Emit a mode 00 ModR/M with a 100 RM indicating a SIB byte is present.
+fn modrm_sib<CS: CodeSink + ?Sized>(reg: RegUnit, sink: &mut CS) {
+    modrm_rm(0b100, reg, sink);
+}
+
+/// Emit a mode 01 ModR/M with a 100 RM indicating a SIB byte and 8-bit
+/// displacement are present.
+fn modrm_sib_disp8<CS: CodeSink + ?Sized>(reg: RegUnit, sink: &mut CS) {
+    modrm_disp8(0b100, reg, sink);
+}
+
+/// Emit a mode 10 ModR/M with a 100 RM indicating a SIB byte and 32-bit
+/// displacement are present.
+fn modrm_sib_disp32<CS: CodeSink + ?Sized>(reg: RegUnit, sink: &mut CS) {
+    modrm_disp32(0b100, reg, sink);
+}
+
+/// Emit a SIB byte with a base register and no scale+index.
+fn sib_noindex<CS: CodeSink + ?Sized>(base: RegUnit, sink: &mut CS) {
+    let base = base as u8 & 7;
+    // SIB        SS_III_BBB.
+    let mut b = 0b00_100_000;
+    b |= base;
+    sink.put1(b);
+}
+
+/// Emit a SIB byte with a scale, base, and index.
+fn sib<CS: CodeSink + ?Sized>(scale: u8, index: RegUnit, base: RegUnit, sink: &mut CS) {
+    // SIB        SS_III_BBB.
+    debug_assert_eq!(scale & !0x03, 0, "Scale out of range");
+    let scale = scale & 3;
+    let index = index as u8 & 7;
+    let base = base as u8 & 7;
+    let b: u8 = (scale << 6) | (index << 3) | base;
+    sink.put1(b);
+}
+
+/// Get the low 4 bits of an opcode for an integer condition code.
+///
+/// Add this offset to a base opcode for:
+///
+/// ---- 0x70: Short conditional branch.
+/// 0x0f 0x80: Long conditional branch.
+/// 0x0f 0x90: SetCC.
+///
+fn icc2opc(cond: IntCC) -> u16 {
+    use crate::ir::condcodes::IntCC::*;
+    match cond {
+        // 0x0 = Overflow.
+        // 0x1 = !Overflow.
+        UnsignedLessThan => 0x2,
+        UnsignedGreaterThanOrEqual => 0x3,
+        Equal => 0x4,
+        NotEqual => 0x5,
+        UnsignedLessThanOrEqual => 0x6,
+        UnsignedGreaterThan => 0x7,
+        // 0x8 = Sign.
+        // 0x9 = !Sign.
+        // 0xa = Parity even.
+        // 0xb = Parity odd.
+        SignedLessThan => 0xc,
+        SignedGreaterThanOrEqual => 0xd,
+        SignedLessThanOrEqual => 0xe,
+        SignedGreaterThan => 0xf,
+    }
+}
+
+/// Get the low 4 bits of an opcode for a floating point condition code.
+///
+/// The ucomiss/ucomisd instructions set the FLAGS bits CF/PF/CF like this:
+///
+///    ZPC OSA
+/// UN 111 000
+/// GT 000 000
+/// LT 001 000
+/// EQ 100 000
+///
+/// Not all floating point condition codes are supported.
+fn fcc2opc(cond: FloatCC) -> u16 {
+    use crate::ir::condcodes::FloatCC::*;
+    match cond {
+        Ordered                    => 0xb, // EQ|LT|GT => *np (P=0)
+        Unordered                  => 0xa, // UN       => *p  (P=1)
+        OrderedNotEqual            => 0x5, // LT|GT    => *ne (Z=0),
+        UnorderedOrEqual           => 0x4, // UN|EQ    => *e  (Z=1)
+        GreaterThan                => 0x7, // GT       => *a  (C=0&Z=0)
+        GreaterThanOrEqual         => 0x3, // GT|EQ    => *ae (C=0)
+        UnorderedOrLessThan        => 0x2, // UN|LT    => *b  (C=1)
+        UnorderedOrLessThanOrEqual => 0x6, // UN|LT|EQ => *be (Z=1|C=1)
+        Equal |                            // EQ
+        NotEqual |                         // UN|LT|GT
+        LessThan |                         // LT
+        LessThanOrEqual |                  // LT|EQ
+        UnorderedOrGreaterThan |           // UN|GT
+        UnorderedOrGreaterThanOrEqual      // UN|GT|EQ
+        => panic!("{} not supported", cond),
+    }
+}
+
+/// Emit a single-byte branch displacement to `destination`.
+fn disp1<CS: CodeSink + ?Sized>(destination: Ebb, func: &Function, sink: &mut CS) {
+    let delta = func.offsets[destination].wrapping_sub(sink.offset() + 1);
+    sink.put1(delta as u8);
+}
+
+/// Emit a four-byte branch displacement to `destination`.
+fn disp4<CS: CodeSink + ?Sized>(destination: Ebb, func: &Function, sink: &mut CS) {
+    let delta = func.offsets[destination].wrapping_sub(sink.offset() + 4);
+    sink.put4(delta);
+}
+
+/// Emit a four-byte displacement to jump table `jt`.
+fn jt_disp4<CS: CodeSink + ?Sized>(jt: JumpTable, func: &Function, sink: &mut CS) {
+    let delta = func.jt_offsets[jt].wrapping_sub(sink.offset() + 4);
+    sink.put4(delta);
+}

cranelift/codegen/src/isa/x86/enc_tables.rs

@@ -0,0 +1,778 @@
+//! Encoding tables for x86 ISAs.
+
+use super::registers::*;
+use crate::bitset::BitSet;
+use crate::cursor::{Cursor, FuncCursor};
+use crate::flowgraph::ControlFlowGraph;
+use crate::ir::condcodes::IntCC;
+use crate::ir::{self, Function, Inst, InstBuilder};
+use crate::isa;
+use crate::isa::constraints::*;
+use crate::isa::enc_tables::*;
+use crate::isa::encoding::base_size;
+use crate::isa::encoding::RecipeSizing;
+use crate::isa::RegUnit;
+use crate::regalloc::RegDiversions;
+
+include!(concat!(env!("OUT_DIR"), "/encoding-x86.rs"));
+include!(concat!(env!("OUT_DIR"), "/legalize-x86.rs"));
+
+pub fn needs_sib_byte(reg: RegUnit) -> bool {
+    reg == RU::r12 as RegUnit || reg == RU::rsp as RegUnit
+}
+pub fn needs_offset(reg: RegUnit) -> bool {
+    reg == RU::r13 as RegUnit || reg == RU::rbp as RegUnit
+}
+pub fn needs_sib_byte_or_offset(reg: RegUnit) -> bool {
+    needs_sib_byte(reg) || needs_offset(reg)
+}
+
+fn additional_size_if(
+    op_index: usize,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+    condition_func: fn(RegUnit) -> bool,
+) -> u8 {
+    let addr_reg = divert.reg(func.dfg.inst_args(inst)[op_index], &func.locations);
+    if condition_func(addr_reg) {
+        1
+    } else {
+        0
+    }
+}
+
+fn size_plus_maybe_offset_for_in_reg_0(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(0, inst, divert, func, needs_offset)
+}
+fn size_plus_maybe_offset_for_in_reg_1(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(1, inst, divert, func, needs_offset)
+}
+fn size_plus_maybe_sib_for_in_reg_0(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(0, inst, divert, func, needs_sib_byte)
+}
+fn size_plus_maybe_sib_for_in_reg_1(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(1, inst, divert, func, needs_sib_byte)
+}
+fn size_plus_maybe_sib_or_offset_for_in_reg_0(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(0, inst, divert, func, needs_sib_byte_or_offset)
+}
+fn size_plus_maybe_sib_or_offset_for_in_reg_1(
+    sizing: &RecipeSizing,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    sizing.base_size + additional_size_if(1, inst, divert, func, needs_sib_byte_or_offset)
+}
+
+/// Expand the `sdiv` and `srem` instructions using `x86_sdivmodx`.
+fn expand_sdivrem(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    isa: &isa::TargetIsa,
+) {
+    let (x, y, is_srem) = match func.dfg[inst] {
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Sdiv,
+            args,
+        } => (args[0], args[1], false),
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Srem,
+            args,
+        } => (args[0], args[1], true),
+        _ => panic!("Need sdiv/srem: {}", func.dfg.display_inst(inst, None)),
+    };
+    let avoid_div_traps = isa.flags().avoid_div_traps();
+    let old_ebb = func.layout.pp_ebb(inst);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+    pos.func.dfg.clear_results(inst);
+
+    // If we can tolerate native division traps, sdiv doesn't need branching.
+    if !avoid_div_traps && !is_srem {
+        let xhi = pos.ins().sshr_imm(x, i64::from(ty.lane_bits()) - 1);
+        pos.ins().with_result(result).x86_sdivmodx(x, xhi, y);
+        pos.remove_inst();
+        return;
+    }
+
+    // EBB handling the -1 divisor case.
+    let minus_one = pos.func.dfg.make_ebb();
+
+    // Final EBB with one argument representing the final result value.
+    let done = pos.func.dfg.make_ebb();
+
+    // Move the `inst` result value onto the `done` EBB.
+    pos.func.dfg.attach_ebb_param(done, result);
+
+    // Start by checking for a -1 divisor which needs to be handled specially.
+    let is_m1 = pos.ins().ifcmp_imm(y, -1);
+    pos.ins().brif(IntCC::Equal, is_m1, minus_one, &[]);
+
+    // Put in an explicit division-by-zero trap if the environment requires it.
+    if avoid_div_traps {
+        pos.ins().trapz(y, ir::TrapCode::IntegerDivisionByZero);
+    }
+
+    // Now it is safe to execute the `x86_sdivmodx` instruction which will still trap on division
+    // by zero.
+    let xhi = pos.ins().sshr_imm(x, i64::from(ty.lane_bits()) - 1);
+    let (quot, rem) = pos.ins().x86_sdivmodx(x, xhi, y);
+    let divres = if is_srem { rem } else { quot };
+    pos.ins().jump(done, &[divres]);
+
+    // Now deal with the -1 divisor case.
+    pos.insert_ebb(minus_one);
+    let m1_result = if is_srem {
+        // x % -1 = 0.
+        pos.ins().iconst(ty, 0)
+    } else {
+        // Explicitly check for overflow: Trap when x == INT_MIN.
+        debug_assert!(avoid_div_traps, "Native trapping divide handled above");
+        let f = pos.ins().ifcmp_imm(x, -1 << (ty.lane_bits() - 1));
+        pos.ins()
+            .trapif(IntCC::Equal, f, ir::TrapCode::IntegerOverflow);
+        // x / -1 = -x.
+        pos.ins().irsub_imm(x, 0)
+    };
+
+    // Recycle the original instruction as a jump.
+    pos.func.dfg.replace(inst).jump(done, &[m1_result]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, minus_one);
+    cfg.recompute_ebb(pos.func, done);
+}
+
+/// Expand the `udiv` and `urem` instructions using `x86_udivmodx`.
+fn expand_udivrem(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    _cfg: &mut ControlFlowGraph,
+    isa: &isa::TargetIsa,
+) {
+    let (x, y, is_urem) = match func.dfg[inst] {
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Udiv,
+            args,
+        } => (args[0], args[1], false),
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Urem,
+            args,
+        } => (args[0], args[1], true),
+        _ => panic!("Need udiv/urem: {}", func.dfg.display_inst(inst, None)),
+    };
+    let avoid_div_traps = isa.flags().avoid_div_traps();
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+    pos.func.dfg.clear_results(inst);
+
+    // Put in an explicit division-by-zero trap if the environment requires it.
+    if avoid_div_traps {
+        pos.ins().trapz(y, ir::TrapCode::IntegerDivisionByZero);
+    }
+
+    // Now it is safe to execute the `x86_udivmodx` instruction.
+    let xhi = pos.ins().iconst(ty, 0);
+    let reuse = if is_urem {
+        [None, Some(result)]
+    } else {
+        [Some(result), None]
+    };
+    pos.ins().with_results(reuse).x86_udivmodx(x, xhi, y);
+    pos.remove_inst();
+}
+
+/// Expand the `fmin` and `fmax` instructions using the x86 `x86_fmin` and `x86_fmax`
+/// instructions.
+fn expand_minmax(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::FloatCC;
+
+    let (x, y, x86_opc, bitwise_opc) = match func.dfg[inst] {
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Fmin,
+            args,
+        } => (args[0], args[1], ir::Opcode::X86Fmin, ir::Opcode::Bor),
+        ir::InstructionData::Binary {
+            opcode: ir::Opcode::Fmax,
+            args,
+        } => (args[0], args[1], ir::Opcode::X86Fmax, ir::Opcode::Band),
+        _ => panic!("Expected fmin/fmax: {}", func.dfg.display_inst(inst, None)),
+    };
+    let old_ebb = func.layout.pp_ebb(inst);
+
+    // We need to handle the following conditions, depending on how x and y compare:
+    //
+    // 1. LT or GT: The native `x86_opc` min/max instruction does what we need.
+    // 2. EQ: We need to use `bitwise_opc` to make sure that
+    //    fmin(0.0, -0.0) -> -0.0 and fmax(0.0, -0.0) -> 0.0.
+    // 3. UN: We need to produce a quiet NaN that is canonical if the inputs are canonical.
+
+    // EBB handling case 3) where one operand is NaN.
+    let uno_ebb = func.dfg.make_ebb();
+
+    // EBB that handles the unordered or equal cases 2) and 3).
+    let ueq_ebb = func.dfg.make_ebb();
+
+    // Final EBB with one argument representing the final result value.
+    let done = func.dfg.make_ebb();
+
+    // The basic blocks are laid out to minimize branching for the common cases:
+    //
+    // 1) One branch not taken, one jump.
+    // 2) One branch taken.
+    // 3) Two branches taken, one jump.
+
+    // Move the `inst` result value onto the `done` EBB.
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+    func.dfg.clear_results(inst);
+    func.dfg.attach_ebb_param(done, result);
+
+    // Test for case 1) ordered and not equal.
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+    let cmp_ueq = pos.ins().fcmp(FloatCC::UnorderedOrEqual, x, y);
+    pos.ins().brnz(cmp_ueq, ueq_ebb, &[]);
+
+    // Handle the common ordered, not equal (LT|GT) case.
+    let one_inst = pos.ins().Binary(x86_opc, ty, x, y).0;
+    let one_result = pos.func.dfg.first_result(one_inst);
+    pos.ins().jump(done, &[one_result]);
+
+    // Case 3) Unordered.
+    // We know that at least one operand is a NaN that needs to be propagated. We simply use an
+    // `fadd` instruction which has the same NaN propagation semantics.
+    pos.insert_ebb(uno_ebb);
+    let uno_result = pos.ins().fadd(x, y);
+    pos.ins().jump(done, &[uno_result]);
+
+    // Case 2) or 3).
+    pos.insert_ebb(ueq_ebb);
+    // Test for case 3) (UN) one value is NaN.
+    // TODO: When we get support for flag values, we can reuse the above comparison.
+    let cmp_uno = pos.ins().fcmp(FloatCC::Unordered, x, y);
+    pos.ins().brnz(cmp_uno, uno_ebb, &[]);
+
+    // We are now in case 2) where x and y compare EQ.
+    // We need a bitwise operation to get the sign right.
+    let bw_inst = pos.ins().Binary(bitwise_opc, ty, x, y).0;
+    let bw_result = pos.func.dfg.first_result(bw_inst);
+    // This should become a fall-through for this second most common case.
+    // Recycle the original instruction as a jump.
+    pos.func.dfg.replace(inst).jump(done, &[bw_result]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, ueq_ebb);
+    cfg.recompute_ebb(pos.func, uno_ebb);
+    cfg.recompute_ebb(pos.func, done);
+}
+
+/// x86 has no unsigned-to-float conversions. We handle the easy case of zero-extending i32 to
+/// i64 with a pattern, the rest needs more code.
+fn expand_fcvt_from_uint(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::IntCC;
+
+    let x;
+    match func.dfg[inst] {
+        ir::InstructionData::Unary {
+            opcode: ir::Opcode::FcvtFromUint,
+            arg,
+        } => x = arg,
+        _ => panic!("Need fcvt_from_uint: {}", func.dfg.display_inst(inst, None)),
+    }
+    let xty = func.dfg.value_type(x);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+
+    // Conversion from unsigned 32-bit is easy on x86-64.
+    // TODO: This should be guarded by an ISA check.
+    if xty == ir::types::I32 {
+        let wide = pos.ins().uextend(ir::types::I64, x);
+        pos.func.dfg.replace(inst).fcvt_from_sint(ty, wide);
+        return;
+    }
+
+    let old_ebb = pos.func.layout.pp_ebb(inst);
+
+    // EBB handling the case where x < 0.
+    let neg_ebb = pos.func.dfg.make_ebb();
+
+    // Final EBB with one argument representing the final result value.
+    let done = pos.func.dfg.make_ebb();
+
+    // Move the `inst` result value onto the `done` EBB.
+    pos.func.dfg.clear_results(inst);
+    pos.func.dfg.attach_ebb_param(done, result);
+
+    // If x as a signed int is not negative, we can use the existing `fcvt_from_sint` instruction.
+    let is_neg = pos.ins().icmp_imm(IntCC::SignedLessThan, x, 0);
+    pos.ins().brnz(is_neg, neg_ebb, &[]);
+
+    // Easy case: just use a signed conversion.
+    let posres = pos.ins().fcvt_from_sint(ty, x);
+    pos.ins().jump(done, &[posres]);
+
+    // Now handle the negative case.
+    pos.insert_ebb(neg_ebb);
+
+    // Divide x by two to get it in range for the signed conversion, keep the LSB, and scale it
+    // back up on the FP side.
+    let ihalf = pos.ins().ushr_imm(x, 1);
+    let lsb = pos.ins().band_imm(x, 1);
+    let ifinal = pos.ins().bor(ihalf, lsb);
+    let fhalf = pos.ins().fcvt_from_sint(ty, ifinal);
+    let negres = pos.ins().fadd(fhalf, fhalf);
+
+    // Recycle the original instruction as a jump.
+    pos.func.dfg.replace(inst).jump(done, &[negres]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, neg_ebb);
+    cfg.recompute_ebb(pos.func, done);
+}
+
+fn expand_fcvt_to_sint(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::{FloatCC, IntCC};
+    use crate::ir::immediates::{Ieee32, Ieee64};
+
+    let x = match func.dfg[inst] {
+        ir::InstructionData::Unary {
+            opcode: ir::Opcode::FcvtToSint,
+            arg,
+        } => arg,
+        _ => panic!("Need fcvt_to_sint: {}", func.dfg.display_inst(inst, None)),
+    };
+    let old_ebb = func.layout.pp_ebb(inst);
+    let xty = func.dfg.value_type(x);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    // Final EBB after the bad value checks.
+    let done = func.dfg.make_ebb();
+
+    // The `x86_cvtt2si` performs the desired conversion, but it doesn't trap on NaN or overflow.
+    // It produces an INT_MIN result instead.
+    func.dfg.replace(inst).x86_cvtt2si(ty, x);
+
+    let mut pos = FuncCursor::new(func).after_inst(inst);
+    pos.use_srcloc(inst);
+
+    let is_done = pos
+        .ins()
+        .icmp_imm(IntCC::NotEqual, result, 1 << (ty.lane_bits() - 1));
+    pos.ins().brnz(is_done, done, &[]);
+
+    // We now have the following possibilities:
+    //
+    // 1. INT_MIN was actually the correct conversion result.
+    // 2. The input was NaN -> trap bad_toint
+    // 3. The input was out of range -> trap int_ovf
+    //
+
+    // Check for NaN.
+    let is_nan = pos.ins().fcmp(FloatCC::Unordered, x, x);
+    pos.ins()
+        .trapnz(is_nan, ir::TrapCode::BadConversionToInteger);
+
+    // Check for case 1: INT_MIN is the correct result.
+    // Determine the smallest floating point number that would convert to INT_MIN.
+    let mut overflow_cc = FloatCC::LessThan;
+    let output_bits = ty.lane_bits();
+    let flimit = match xty {
+        ir::types::F32 =>
+        // An f32 can represent `i16::min_value() - 1` exactly with precision to spare, so
+        // there are values less than -2^(N-1) that convert correctly to INT_MIN.
+        {
+            pos.ins().f32const(if output_bits < 32 {
+                overflow_cc = FloatCC::LessThanOrEqual;
+                Ieee32::fcvt_to_sint_negative_overflow(output_bits)
+            } else {
+                Ieee32::pow2(output_bits - 1).neg()
+            })
+        }
+        ir::types::F64 =>
+        // An f64 can represent `i32::min_value() - 1` exactly with precision to spare, so
+        // there are values less than -2^(N-1) that convert correctly to INT_MIN.
+        {
+            pos.ins().f64const(if output_bits < 64 {
+                overflow_cc = FloatCC::LessThanOrEqual;
+                Ieee64::fcvt_to_sint_negative_overflow(output_bits)
+            } else {
+                Ieee64::pow2(output_bits - 1).neg()
+            })
+        }
+        _ => panic!("Can't convert {}", xty),
+    };
+    let overflow = pos.ins().fcmp(overflow_cc, x, flimit);
+    pos.ins().trapnz(overflow, ir::TrapCode::IntegerOverflow);
+
+    // Finally, we could have a positive value that is too large.
+    let fzero = match xty {
+        ir::types::F32 => pos.ins().f32const(Ieee32::with_bits(0)),
+        ir::types::F64 => pos.ins().f64const(Ieee64::with_bits(0)),
+        _ => panic!("Can't convert {}", xty),
+    };
+    let overflow = pos.ins().fcmp(FloatCC::GreaterThanOrEqual, x, fzero);
+    pos.ins().trapnz(overflow, ir::TrapCode::IntegerOverflow);
+
+    pos.ins().jump(done, &[]);
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, done);
+}
+
+fn expand_fcvt_to_sint_sat(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::{FloatCC, IntCC};
+    use crate::ir::immediates::{Ieee32, Ieee64};
+
+    let x = match func.dfg[inst] {
+        ir::InstructionData::Unary {
+            opcode: ir::Opcode::FcvtToSintSat,
+            arg,
+        } => arg,
+        _ => panic!(
+            "Need fcvt_to_sint_sat: {}",
+            func.dfg.display_inst(inst, None)
+        ),
+    };
+
+    let old_ebb = func.layout.pp_ebb(inst);
+    let xty = func.dfg.value_type(x);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    // Final EBB after the bad value checks.
+    let done_ebb = func.dfg.make_ebb();
+    func.dfg.clear_results(inst);
+    func.dfg.attach_ebb_param(done_ebb, result);
+
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+
+    // The `x86_cvtt2si` performs the desired conversion, but it doesn't trap on NaN or
+    // overflow. It produces an INT_MIN result instead.
+    let cvtt2si = pos.ins().x86_cvtt2si(ty, x);
+
+    let is_done = pos
+        .ins()
+        .icmp_imm(IntCC::NotEqual, cvtt2si, 1 << (ty.lane_bits() - 1));
+    pos.ins().brnz(is_done, done_ebb, &[cvtt2si]);
+
+    // We now have the following possibilities:
+    //
+    // 1. INT_MIN was actually the correct conversion result.
+    // 2. The input was NaN -> replace the result value with 0.
+    // 3. The input was out of range -> saturate the result to the min/max value.
+
+    // Check for NaN, which is truncated to 0.
+    let zero = pos.ins().iconst(ty, 0);
+    let is_nan = pos.ins().fcmp(FloatCC::Unordered, x, x);
+    pos.ins().brnz(is_nan, done_ebb, &[zero]);
+
+    // Check for case 1: INT_MIN is the correct result.
+    // Determine the smallest floating point number that would convert to INT_MIN.
+    let mut overflow_cc = FloatCC::LessThan;
+    let output_bits = ty.lane_bits();
+    let flimit = match xty {
+        ir::types::F32 =>
+        // An f32 can represent `i16::min_value() - 1` exactly with precision to spare, so
+        // there are values less than -2^(N-1) that convert correctly to INT_MIN.
+        {
+            pos.ins().f32const(if output_bits < 32 {
+                overflow_cc = FloatCC::LessThanOrEqual;
+                Ieee32::fcvt_to_sint_negative_overflow(output_bits)
+            } else {
+                Ieee32::pow2(output_bits - 1).neg()
+            })
+        }
+        ir::types::F64 =>
+        // An f64 can represent `i32::min_value() - 1` exactly with precision to spare, so
+        // there are values less than -2^(N-1) that convert correctly to INT_MIN.
+        {
+            pos.ins().f64const(if output_bits < 64 {
+                overflow_cc = FloatCC::LessThanOrEqual;
+                Ieee64::fcvt_to_sint_negative_overflow(output_bits)
+            } else {
+                Ieee64::pow2(output_bits - 1).neg()
+            })
+        }
+        _ => panic!("Can't convert {}", xty),
+    };
+
+    let overflow = pos.ins().fcmp(overflow_cc, x, flimit);
+    let min_imm = match ty {
+        ir::types::I32 => i32::min_value() as i64,
+        ir::types::I64 => i64::min_value(),
+        _ => panic!("Don't know the min value for {}", ty),
+    };
+    let min_value = pos.ins().iconst(ty, min_imm);
+    pos.ins().brnz(overflow, done_ebb, &[min_value]);
+
+    // Finally, we could have a positive value that is too large.
+    let fzero = match xty {
+        ir::types::F32 => pos.ins().f32const(Ieee32::with_bits(0)),
+        ir::types::F64 => pos.ins().f64const(Ieee64::with_bits(0)),
+        _ => panic!("Can't convert {}", xty),
+    };
+
+    let max_imm = match ty {
+        ir::types::I32 => i32::max_value() as i64,
+        ir::types::I64 => i64::max_value(),
+        _ => panic!("Don't know the max value for {}", ty),
+    };
+    let max_value = pos.ins().iconst(ty, max_imm);
+
+    let overflow = pos.ins().fcmp(FloatCC::GreaterThanOrEqual, x, fzero);
+    pos.ins().brnz(overflow, done_ebb, &[max_value]);
+
+    // Recycle the original instruction.
+    pos.func.dfg.replace(inst).jump(done_ebb, &[cvtt2si]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done_ebb);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, done_ebb);
+}
+
+fn expand_fcvt_to_uint(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::{FloatCC, IntCC};
+    use crate::ir::immediates::{Ieee32, Ieee64};
+
+    let x = match func.dfg[inst] {
+        ir::InstructionData::Unary {
+            opcode: ir::Opcode::FcvtToUint,
+            arg,
+        } => arg,
+        _ => panic!("Need fcvt_to_uint: {}", func.dfg.display_inst(inst, None)),
+    };
+
+    let old_ebb = func.layout.pp_ebb(inst);
+    let xty = func.dfg.value_type(x);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    // EBB handling numbers >= 2^(N-1).
+    let large = func.dfg.make_ebb();
+
+    // Final EBB after the bad value checks.
+    let done = func.dfg.make_ebb();
+
+    // Move the `inst` result value onto the `done` EBB.
+    func.dfg.clear_results(inst);
+    func.dfg.attach_ebb_param(done, result);
+
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+
+    // Start by materializing the floating point constant 2^(N-1) where N is the number of bits in
+    // the destination integer type.
+    let pow2nm1 = match xty {
+        ir::types::F32 => pos.ins().f32const(Ieee32::pow2(ty.lane_bits() - 1)),
+        ir::types::F64 => pos.ins().f64const(Ieee64::pow2(ty.lane_bits() - 1)),
+        _ => panic!("Can't convert {}", xty),
+    };
+    let is_large = pos.ins().ffcmp(x, pow2nm1);
+    pos.ins()
+        .brff(FloatCC::GreaterThanOrEqual, is_large, large, &[]);
+
+    // We need to generate a specific trap code when `x` is NaN, so reuse the flags from the
+    // previous comparison.
+    pos.ins().trapff(
+        FloatCC::Unordered,
+        is_large,
+        ir::TrapCode::BadConversionToInteger,
+    );
+
+    // Now we know that x < 2^(N-1) and not NaN.
+    let sres = pos.ins().x86_cvtt2si(ty, x);
+    let is_neg = pos.ins().ifcmp_imm(sres, 0);
+    pos.ins()
+        .brif(IntCC::SignedGreaterThanOrEqual, is_neg, done, &[sres]);
+    pos.ins().trap(ir::TrapCode::IntegerOverflow);
+
+    // Handle the case where x >= 2^(N-1) and not NaN.
+    pos.insert_ebb(large);
+    let adjx = pos.ins().fsub(x, pow2nm1);
+    let lres = pos.ins().x86_cvtt2si(ty, adjx);
+    let is_neg = pos.ins().ifcmp_imm(lres, 0);
+    pos.ins()
+        .trapif(IntCC::SignedLessThan, is_neg, ir::TrapCode::IntegerOverflow);
+    let lfinal = pos.ins().iadd_imm(lres, 1 << (ty.lane_bits() - 1));
+
+    // Recycle the original instruction as a jump.
+    pos.func.dfg.replace(inst).jump(done, &[lfinal]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, large);
+    cfg.recompute_ebb(pos.func, done);
+}
+
+fn expand_fcvt_to_uint_sat(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    cfg: &mut ControlFlowGraph,
+    _isa: &isa::TargetIsa,
+) {
+    use crate::ir::condcodes::{FloatCC, IntCC};
+    use crate::ir::immediates::{Ieee32, Ieee64};
+
+    let x = match func.dfg[inst] {
+        ir::InstructionData::Unary {
+            opcode: ir::Opcode::FcvtToUintSat,
+            arg,
+        } => arg,
+        _ => panic!(
+            "Need fcvt_to_uint_sat: {}",
+            func.dfg.display_inst(inst, None)
+        ),
+    };
+
+    let old_ebb = func.layout.pp_ebb(inst);
+    let xty = func.dfg.value_type(x);
+    let result = func.dfg.first_result(inst);
+    let ty = func.dfg.value_type(result);
+
+    // EBB handling numbers >= 2^(N-1).
+    let large = func.dfg.make_ebb();
+
+    // Final EBB after the bad value checks.
+    let done = func.dfg.make_ebb();
+
+    // Move the `inst` result value onto the `done` EBB.
+    func.dfg.clear_results(inst);
+    func.dfg.attach_ebb_param(done, result);
+
+    let mut pos = FuncCursor::new(func).at_inst(inst);
+    pos.use_srcloc(inst);
+
+    // Start by materializing the floating point constant 2^(N-1) where N is the number of bits in
+    // the destination integer type.
+    let pow2nm1 = match xty {
+        ir::types::F32 => pos.ins().f32const(Ieee32::pow2(ty.lane_bits() - 1)),
+        ir::types::F64 => pos.ins().f64const(Ieee64::pow2(ty.lane_bits() - 1)),
+        _ => panic!("Can't convert {}", xty),
+    };
+    let zero = pos.ins().iconst(ty, 0);
+    let is_large = pos.ins().ffcmp(x, pow2nm1);
+    pos.ins()
+        .brff(FloatCC::GreaterThanOrEqual, is_large, large, &[]);
+
+    // We need to generate zero when `x` is NaN, so reuse the flags from the previous comparison.
+    pos.ins().brff(FloatCC::Unordered, is_large, done, &[zero]);
+
+    // Now we know that x < 2^(N-1) and not NaN. If the result of the cvtt2si is positive, we're
+    // done; otherwise saturate to the minimum unsigned value, that is 0.
+    let sres = pos.ins().x86_cvtt2si(ty, x);
+    let is_neg = pos.ins().ifcmp_imm(sres, 0);
+    pos.ins()
+        .brif(IntCC::SignedGreaterThanOrEqual, is_neg, done, &[sres]);
+    pos.ins().jump(done, &[zero]);
+
+    // Handle the case where x >= 2^(N-1) and not NaN.
+    pos.insert_ebb(large);
+    let adjx = pos.ins().fsub(x, pow2nm1);
+    let lres = pos.ins().x86_cvtt2si(ty, adjx);
+    let max_value = pos.ins().iconst(
+        ty,
+        match ty {
+            ir::types::I32 => u32::max_value() as i64,
+            ir::types::I64 => u64::max_value() as i64,
+            _ => panic!("Can't convert {}", ty),
+        },
+    );
+    let is_neg = pos.ins().ifcmp_imm(lres, 0);
+    pos.ins()
+        .brif(IntCC::SignedLessThan, is_neg, done, &[max_value]);
+    let lfinal = pos.ins().iadd_imm(lres, 1 << (ty.lane_bits() - 1));
+
+    // Recycle the original instruction as a jump.
+    pos.func.dfg.replace(inst).jump(done, &[lfinal]);
+
+    // Finally insert a label for the completion.
+    pos.next_inst();
+    pos.insert_ebb(done);
+
+    cfg.recompute_ebb(pos.func, old_ebb);
+    cfg.recompute_ebb(pos.func, large);
+    cfg.recompute_ebb(pos.func, done);
+}

cranelift/codegen/src/isa/x86/mod.rs

@@ -0,0 +1,145 @@
+//! x86 Instruction Set Architectures.
+
+mod abi;
+mod binemit;
+mod enc_tables;
+mod registers;
+pub mod settings;
+
+use super::super::settings as shared_settings;
+#[cfg(feature = "testing_hooks")]
+use crate::binemit::CodeSink;
+use crate::binemit::{emit_function, MemoryCodeSink};
+use crate::ir;
+use crate::isa::enc_tables::{self as shared_enc_tables, lookup_enclist, Encodings};
+use crate::isa::Builder as IsaBuilder;
+use crate::isa::{EncInfo, RegClass, RegInfo, TargetIsa};
+use crate::regalloc;
+use crate::result::CodegenResult;
+use crate::timing;
+use core::fmt;
+use std::boxed::Box;
+use target_lexicon::{PointerWidth, Triple};
+
+#[allow(dead_code)]
+struct Isa {
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    isa_flags: settings::Flags,
+    cpumode: &'static [shared_enc_tables::Level1Entry<u16>],
+}
+
+/// Get an ISA builder for creating x86 targets.
+pub fn isa_builder(triple: Triple) -> IsaBuilder {
+    IsaBuilder {
+        triple,
+        setup: settings::builder(),
+        constructor: isa_constructor,
+    }
+}
+
+fn isa_constructor(
+    triple: Triple,
+    shared_flags: shared_settings::Flags,
+    builder: shared_settings::Builder,
+) -> Box<TargetIsa> {
+    let level1 = match triple.pointer_width().unwrap() {
+        PointerWidth::U16 => unimplemented!("x86-16"),
+        PointerWidth::U32 => &enc_tables::LEVEL1_I32[..],
+        PointerWidth::U64 => &enc_tables::LEVEL1_I64[..],
+    };
+    Box::new(Isa {
+        triple,
+        isa_flags: settings::Flags::new(&shared_flags, builder),
+        shared_flags,
+        cpumode: level1,
+    })
+}
+
+impl TargetIsa for Isa {
+    fn name(&self) -> &'static str {
+        "x86"
+    }
+
+    fn triple(&self) -> &Triple {
+        &self.triple
+    }
+
+    fn flags(&self) -> &shared_settings::Flags {
+        &self.shared_flags
+    }
+
+    fn uses_cpu_flags(&self) -> bool {
+        true
+    }
+
+    fn uses_complex_addresses(&self) -> bool {
+        true
+    }
+
+    fn register_info(&self) -> RegInfo {
+        registers::INFO.clone()
+    }
+
+    fn encoding_info(&self) -> EncInfo {
+        enc_tables::INFO.clone()
+    }
+
+    fn legal_encodings<'a>(
+        &'a self,
+        func: &'a ir::Function,
+        inst: &'a ir::InstructionData,
+        ctrl_typevar: ir::Type,
+    ) -> Encodings<'a> {
+        lookup_enclist(
+            ctrl_typevar,
+            inst,
+            func,
+            self.cpumode,
+            &enc_tables::LEVEL2[..],
+            &enc_tables::ENCLISTS[..],
+            &enc_tables::LEGALIZE_ACTIONS[..],
+            &enc_tables::RECIPE_PREDICATES[..],
+            &enc_tables::INST_PREDICATES[..],
+            self.isa_flags.predicate_view(),
+        )
+    }
+
+    fn legalize_signature(&self, sig: &mut ir::Signature, current: bool) {
+        abi::legalize_signature(sig, &self.triple, current)
+    }
+
+    fn regclass_for_abi_type(&self, ty: ir::Type) -> RegClass {
+        abi::regclass_for_abi_type(ty)
+    }
+
+    fn allocatable_registers(&self, func: &ir::Function) -> regalloc::RegisterSet {
+        abi::allocatable_registers(func, &self.triple)
+    }
+
+    #[cfg(feature = "testing_hooks")]
+    fn emit_inst(
+        &self,
+        func: &ir::Function,
+        inst: ir::Inst,
+        divert: &mut regalloc::RegDiversions,
+        sink: &mut CodeSink,
+    ) {
+        binemit::emit_inst(func, inst, divert, sink)
+    }
+
+    fn emit_function_to_memory(&self, func: &ir::Function, sink: &mut MemoryCodeSink) {
+        emit_function(func, binemit::emit_inst, sink)
+    }
+
+    fn prologue_epilogue(&self, func: &mut ir::Function) -> CodegenResult<()> {
+        let _tt = timing::prologue_epilogue();
+        abi::prologue_epilogue(func, self)
+    }
+}
+
+impl fmt::Display for Isa {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{}\n{}", self.shared_flags, self.isa_flags)
+    }
+}

cranelift/codegen/src/isa/x86/registers.rs

@@ -0,0 +1,63 @@
+//! x86 register descriptions.
+
+use crate::isa::registers::{RegBank, RegClass, RegClassData, RegInfo, RegUnit};
+
+include!(concat!(env!("OUT_DIR"), "/registers-x86.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::isa::RegUnit;
+    use std::string::{String, ToString};
+
+    #[test]
+    fn unit_encodings() {
+        // The encoding of integer registers is not alphabetical.
+        assert_eq!(INFO.parse_regunit("rax"), Some(0));
+        assert_eq!(INFO.parse_regunit("rbx"), Some(3));
+        assert_eq!(INFO.parse_regunit("rcx"), Some(1));
+        assert_eq!(INFO.parse_regunit("rdx"), Some(2));
+        assert_eq!(INFO.parse_regunit("rsi"), Some(6));
+        assert_eq!(INFO.parse_regunit("rdi"), Some(7));
+        assert_eq!(INFO.parse_regunit("rbp"), Some(5));
+        assert_eq!(INFO.parse_regunit("rsp"), Some(4));
+        assert_eq!(INFO.parse_regunit("r8"), Some(8));
+        assert_eq!(INFO.parse_regunit("r15"), Some(15));
+
+        assert_eq!(INFO.parse_regunit("xmm0"), Some(16));
+        assert_eq!(INFO.parse_regunit("xmm15"), Some(31));
+    }
+
+    #[test]
+    fn unit_names() {
+        fn uname(ru: RegUnit) -> String {
+            INFO.display_regunit(ru).to_string()
+        }
+
+        assert_eq!(uname(0), "%rax");
+        assert_eq!(uname(3), "%rbx");
+        assert_eq!(uname(1), "%rcx");
+        assert_eq!(uname(2), "%rdx");
+        assert_eq!(uname(6), "%rsi");
+        assert_eq!(uname(7), "%rdi");
+        assert_eq!(uname(5), "%rbp");
+        assert_eq!(uname(4), "%rsp");
+        assert_eq!(uname(8), "%r8");
+        assert_eq!(uname(15), "%r15");
+        assert_eq!(uname(16), "%xmm0");
+        assert_eq!(uname(31), "%xmm15");
+    }
+
+    #[test]
+    fn regclasses() {
+        assert_eq!(GPR.intersect_index(GPR), Some(GPR.into()));
+        assert_eq!(GPR.intersect_index(ABCD), Some(ABCD.into()));
+        assert_eq!(GPR.intersect_index(FPR), None);
+        assert_eq!(ABCD.intersect_index(GPR), Some(ABCD.into()));
+        assert_eq!(ABCD.intersect_index(ABCD), Some(ABCD.into()));
+        assert_eq!(ABCD.intersect_index(FPR), None);
+        assert_eq!(FPR.intersect_index(FPR), Some(FPR.into()));
+        assert_eq!(FPR.intersect_index(GPR), None);
+        assert_eq!(FPR.intersect_index(ABCD), None);
+    }
+}

cranelift/codegen/src/isa/x86/settings.rs

@@ -0,0 +1,52 @@
+//! x86 Settings.
+
+use crate::settings::{self, detail, Builder};
+use core::fmt;
+
+// Include code generated by `cranelift-codegen/meta-python/gen_settings.py`. This file contains a public
+// `Flags` struct with an impl for all of the settings defined in
+// `cranelift-codegen/meta-python/isa/x86/settings.py`.
+include!(concat!(env!("OUT_DIR"), "/settings-x86.rs"));
+
+#[cfg(test)]
+mod tests {
+    use super::{builder, Flags};
+    use crate::settings::{self, Configurable};
+
+    #[test]
+    fn presets() {
+        let shared = settings::Flags::new(settings::builder());
+
+        // Nehalem has SSE4.1 but not BMI1.
+        let mut b0 = builder();
+        b0.enable("nehalem").unwrap();
+        let f0 = Flags::new(&shared, b0);
+        assert_eq!(f0.has_sse41(), true);
+        assert_eq!(f0.has_bmi1(), false);
+
+        let mut b1 = builder();
+        b1.enable("haswell").unwrap();
+        let f1 = Flags::new(&shared, b1);
+        assert_eq!(f1.has_sse41(), true);
+        assert_eq!(f1.has_bmi1(), true);
+    }
+    #[test]
+    fn display_presets() {
+        // Spot check that the flags Display impl does not cause a panic
+        let shared = settings::Flags::new(settings::builder());
+
+        let b0 = builder();
+        let f0 = Flags::new(&shared, b0);
+        let _ = format!("{}", f0);
+
+        let mut b1 = builder();
+        b1.enable("nehalem").unwrap();
+        let f1 = Flags::new(&shared, b1);
+        let _ = format!("{}", f1);
+
+        let mut b2 = builder();
+        b2.enable("haswell").unwrap();
+        let f2 = Flags::new(&shared, b2);
+        let _ = format!("{}", f2);
+    }
+}