x64 backend: implement 128-bit ops and misc fixes.

This implements all of the ops on I128 that are implemented by the
legacy x86 backend, and includes all that are required by at least one
major use-case (cg_clif rustc backend).

The sequences are open-coded where necessary; for e.g. the bit
operations, this can be somewhat complex, but these sequences have been
tested carefully. This PR also includes a drive-by fix of clz/ctz for 8-
and 16-bit cases where they were incorrect previously.

Also includes ridealong fixes developed while bringing up cg_clif
support, because they are difficult to completely separate due to
other refactors that occurred in this PR:

- fix REX prefix logic for some 8-bit instructions.

  When using an 8-bit register in 64-bit mode on x86-64, the REX prefix
  semantics are somewhat subtle: without the REX prefix, register numbers
  4--7 correspond to the second-to-lowest byte of the first four registers
  (AH, CH, BH, DH), whereas with the REX prefix, these register numbers
  correspond to the usual encoding (SPL, BPL, SIL, DIL). We could always
  emit a REX byte for instructions with 8-bit cases (this is harmless even
  if unneeded), but this would unnecessarily inflate code size; instead,
  the usual approach is to emit it only for these registers.

  This logic was present in some cases but missing for some other
  instructions: divide, not, negate, shifts.

  Fixes #2508.

- avoid unaligned SSE loads on some f64 ops.

  The implementations of several FP ops, such as fabs/fneg, used SSE
  instructions. This is not a problem per-se, except that load-op merging
  did not take *alignment* into account. Specifically, if an op on an f64
  loaded from memory happened to merge that load, and the instruction into
  which it was merged was an SSE instruction, then the SSE instruction
  imposes stricter (128-bit) alignment requirements than the load.f64 did.

  This PR simply forces any instruction lowerings that could use SSE
  instructions to implement non-SIMD operations to take inputs in
  registers only, and avoid load-op merging.

  Fixes #2507.

- two bugfixes exposed by cg_clif: urem/srem.i8, select.b1.

  - urem/srem.i8: the 8-bit form of the DIV instruction on x86-64 places
    the remainder in AH, not RDX, different from all the other width-forms
    of this instruction.

  - select.b1: we were not recognizing selects of boolean values as
    integer-typed operations, so we were generating XMM moves instead (!).

cranelift/codegen/src/isa/x64/abi.rs

@@ -138,42 +138,62 @@ impl ABIMachineSpec for X64ABIMachineSpec {
                 ),
             }
 
-            let intreg = in_int_reg(param.value_type);
-            let vecreg = in_vec_reg(param.value_type);
-            debug_assert!(intreg || vecreg);
-            debug_assert!(!(intreg && vecreg));
-
-            let (next_reg, candidate) = if intreg {
-                let candidate = match args_or_rets {
-                    ArgsOrRets::Args => get_intreg_for_arg_systemv(&call_conv, next_gpr),
-                    ArgsOrRets::Rets => get_intreg_for_retval_systemv(&call_conv, next_gpr, i),
-                };
-                debug_assert!(candidate
-                    .map(|r| r.get_class() == RegClass::I64)
-                    .unwrap_or(true));
-                (&mut next_gpr, candidate)
-            } else {
-                let candidate = match args_or_rets {
-                    ArgsOrRets::Args => get_fltreg_for_arg_systemv(&call_conv, next_vreg),
-                    ArgsOrRets::Rets => get_fltreg_for_retval_systemv(&call_conv, next_vreg, i),
-                };
-                debug_assert!(candidate
-                    .map(|r| r.get_class() == RegClass::V128)
-                    .unwrap_or(true));
-                (&mut next_vreg, candidate)
-            };
-
             if let Some(param) = try_fill_baldrdash_reg(call_conv, param) {
-                assert!(intreg);
                 ret.push(param);
-            } else if let Some(reg) = candidate {
+                continue;
+            }
+
+            // Find regclass(es) of the register(s) used to store a value of this type.
+            let (rcs, _) = Inst::rc_for_type(param.value_type)?;
+            let intreg = rcs[0] == RegClass::I64;
+            let num_regs = rcs.len();
+            assert!(num_regs <= 2);
+            if num_regs == 2 {
+                assert_eq!(rcs[0], rcs[1]);
+            }
+
+            let mut regs: SmallVec<[RealReg; 2]> = smallvec![];
+            for j in 0..num_regs {
+                let nextreg = if intreg {
+                    match args_or_rets {
+                        ArgsOrRets::Args => get_intreg_for_arg_systemv(&call_conv, next_gpr + j),
+                        ArgsOrRets::Rets => {
+                            get_intreg_for_retval_systemv(&call_conv, next_gpr + j, i + j)
+                        }
+                    }
+                } else {
+                    match args_or_rets {
+                        ArgsOrRets::Args => get_fltreg_for_arg_systemv(&call_conv, next_vreg + j),
+                        ArgsOrRets::Rets => {
+                            get_fltreg_for_retval_systemv(&call_conv, next_vreg + j, i + j)
+                        }
+                    }
+                };
+                if let Some(reg) = nextreg {
+                    regs.push(reg.to_real_reg());
+                } else {
+                    regs.clear();
+                    break;
+                }
+            }
+
+            if regs.len() > 0 {
+                let regs = match num_regs {
+                    1 => ValueRegs::one(regs[0]),
+                    2 => ValueRegs::two(regs[0], regs[1]),
+                    _ => panic!("More than two registers unexpected"),
+                };
                 ret.push(ABIArg::Reg(
-                    ValueRegs::one(reg.to_real_reg()),
+                    regs,
                     param.value_type,
                     param.extension,
                     param.purpose,
                 ));
-                *next_reg += 1;
+                if intreg {
+                    next_gpr += num_regs;
+                } else {
+                    next_vreg += num_regs;
+                }
             } else {
                 // Compute size. Every arg takes a minimum slot of 8 bytes. (16-byte
                 // stack alignment happens separately after all args.)
@@ -658,31 +678,6 @@ impl From<StackAMode> for SyntheticAmode {
     }
 }
 
-fn in_int_reg(ty: types::Type) -> bool {
-    match ty {
-        types::I8
-        | types::I16
-        | types::I32
-        | types::I64
-        | types::B1
-        | types::B8
-        | types::B16
-        | types::B32
-        | types::B64
-        | types::R64 => true,
-        types::R32 => panic!("unexpected 32-bits refs on x64!"),
-        _ => false,
-    }
-}
-
-fn in_vec_reg(ty: types::Type) -> bool {
-    match ty {
-        types::F32 | types::F64 => true,
-        _ if ty.is_vector() => true,
-        _ => false,
-    }
-}
-
 fn get_intreg_for_arg_systemv(call_conv: &CallConv, idx: usize) -> Option<Reg> {
     match call_conv {
         CallConv::Fast

cranelift/codegen/src/isa/x64/inst/args.rs

@@ -346,23 +346,35 @@ impl PrettyPrintSized for RegMem {
 #[derive(Copy, Clone, PartialEq)]
 pub enum AluRmiROpcode {
     Add,
+    Adc,
     Sub,
+    Sbb,
     And,
     Or,
     Xor,
     /// The signless, non-extending (N x N -> N, for N in {32,64}) variant.
     Mul,
+    /// 8-bit form of And. Handled separately as we don't have full 8-bit op
+    /// support (we just use wider instructions). Used only with some sequences
+    /// with SETcc.
+    And8,
+    /// 8-bit form of Or.
+    Or8,
 }
 
 impl fmt::Debug for AluRmiROpcode {
     fn fmt(&self, fmt: &mut fmt::Formatter) -> fmt::Result {
         let name = match self {
             AluRmiROpcode::Add => "add",
+            AluRmiROpcode::Adc => "adc",
             AluRmiROpcode::Sub => "sub",
+            AluRmiROpcode::Sbb => "sbb",
             AluRmiROpcode::And => "and",
             AluRmiROpcode::Or => "or",
             AluRmiROpcode::Xor => "xor",
             AluRmiROpcode::Mul => "imul",
+            AluRmiROpcode::And8 => "and",
+            AluRmiROpcode::Or8 => "or",
         };
         write!(fmt, "{}", name)
     }
@@ -374,6 +386,16 @@ impl fmt::Display for AluRmiROpcode {
     }
 }
 
+impl AluRmiROpcode {
+    /// Is this a special-cased 8-bit ALU op?
+    pub fn is_8bit(self) -> bool {
+        match self {
+            AluRmiROpcode::And8 | AluRmiROpcode::Or8 => true,
+            _ => false,
+        }
+    }
+}
+
 #[derive(Clone, PartialEq)]
 pub enum UnaryRmROpcode {
     /// Bit-scan reverse.
@@ -1010,7 +1032,7 @@ impl fmt::Display for ExtMode {
 }
 
 /// These indicate the form of a scalar shift/rotate: left, signed right, unsigned right.
-#[derive(Clone)]
+#[derive(Clone, Copy)]
 pub enum ShiftKind {
     ShiftLeft,
     /// Inserts zeros in the most significant bits.

cranelift/codegen/src/isa/x64/inst/emit.rs

@@ -83,6 +83,14 @@ impl RexFlags {
         self
     }
 
+    #[inline(always)]
+    fn always_emit_if_8bit_needed(&mut self, reg: u8) -> &mut Self {
+        if reg >= 4 && reg <= 7 {
+            self.always_emit();
+        }
+        self
+    }
+
     #[inline(always)]
     fn must_clear_w(&self) -> bool {
         (self.0 & 1) != 0
@@ -527,7 +535,7 @@ pub(crate) fn emit(
             src,
             dst: reg_g,
         } => {
-            let rex = if *is_64 {
+            let mut rex = if *is_64 {
                 RexFlags::set_w()
             } else {
                 RexFlags::clear_w()
@@ -581,17 +589,26 @@ pub(crate) fn emit(
                     }
                 }
             } else {
-                let (opcode_r, opcode_m, subopcode_i) = match op {
+                let (opcode_r, opcode_m, subopcode_i, is_8bit) = match op {
-                    AluRmiROpcode::Add => (0x01, 0x03, 0),
+                    AluRmiROpcode::Add => (0x01, 0x03, 0, false),
-                    AluRmiROpcode::Sub => (0x29, 0x2B, 5),
+                    AluRmiROpcode::Adc => (0x11, 0x03, 0, false),
-                    AluRmiROpcode::And => (0x21, 0x23, 4),
+                    AluRmiROpcode::Sub => (0x29, 0x2B, 5, false),
-                    AluRmiROpcode::Or => (0x09, 0x0B, 1),
+                    AluRmiROpcode::Sbb => (0x19, 0x2B, 5, false),
-                    AluRmiROpcode::Xor => (0x31, 0x33, 6),
+                    AluRmiROpcode::And => (0x21, 0x23, 4, false),
+                    AluRmiROpcode::Or => (0x09, 0x0B, 1, false),
+                    AluRmiROpcode::Xor => (0x31, 0x33, 6, false),
+                    AluRmiROpcode::And8 => (0x20, 0x22, 4, true),
+                    AluRmiROpcode::Or8 => (0x08, 0x0A, 1, true),
                     AluRmiROpcode::Mul => panic!("unreachable"),
                 };
+                assert!(!(is_8bit && *is_64));
 
                 match src {
                     RegMemImm::Reg { reg: reg_e } => {
+                        if is_8bit {
+                            rex.always_emit_if_8bit_needed(int_reg_enc(*reg_e));
+                            rex.always_emit_if_8bit_needed(int_reg_enc(reg_g.to_reg()));
+                        }
                         // GCC/llvm use the swapped operand encoding (viz., the R/RM vs RM/R
                         // duality). Do this too, so as to be able to compare generated machine
                         // code easily.
@@ -604,11 +621,12 @@ pub(crate) fn emit(
                             reg_g.to_reg(),
                             rex,
                         );
-                        // NB: if this is ever extended to handle byte size ops, be sure to retain
-                        // redundant REX prefixes.
                     }
 
                     RegMemImm::Mem { addr } => {
+                        if is_8bit {
+                            rex.always_emit_if_8bit_needed(int_reg_enc(reg_g.to_reg()));
+                        }
                         // Here we revert to the "normal" G-E ordering.
                         let amode = addr.finalize(state, sink);
                         emit_std_reg_mem(
@@ -625,6 +643,7 @@ pub(crate) fn emit(
                     }
 
                     RegMemImm::Imm { simm32 } => {
+                        assert!(!is_8bit);
                         let use_imm8 = low8_will_sign_extend_to_32(*simm32);
                         let opcode = if use_imm8 { 0x83 } else { 0x81 };
                         // And also here we use the "normal" G-E ordering.
@@ -685,8 +704,13 @@ pub(crate) fn emit(
         }
 
         Inst::Not { size, src } => {
+            let src = int_reg_enc(src.to_reg());
             let (opcode, prefix, rex_flags) = match size {
-                1 => (0xF6, LegacyPrefixes::None, RexFlags::clear_w()),
+                1 => (
+                    0xF6,
+                    LegacyPrefixes::None,
+                    *RexFlags::clear_w().always_emit_if_8bit_needed(src),
+                ),
                 2 => (0xF7, LegacyPrefixes::_66, RexFlags::clear_w()),
                 4 => (0xF7, LegacyPrefixes::None, RexFlags::clear_w()),
                 8 => (0xF7, LegacyPrefixes::None, RexFlags::set_w()),
@@ -694,13 +718,17 @@ pub(crate) fn emit(
             };
 
             let subopcode = 2;
-            let src = int_reg_enc(src.to_reg());
             emit_std_enc_enc(sink, prefix, opcode, 1, subopcode, src, rex_flags)
         }
 
         Inst::Neg { size, src } => {
+            let src = int_reg_enc(src.to_reg());
             let (opcode, prefix, rex_flags) = match size {
-                1 => (0xF6, LegacyPrefixes::None, RexFlags::clear_w()),
+                1 => (
+                    0xF6,
+                    LegacyPrefixes::None,
+                    *RexFlags::clear_w().always_emit_if_8bit_needed(src),
+                ),
                 2 => (0xF7, LegacyPrefixes::_66, RexFlags::clear_w()),
                 4 => (0xF7, LegacyPrefixes::None, RexFlags::clear_w()),
                 8 => (0xF7, LegacyPrefixes::None, RexFlags::set_w()),
@@ -708,7 +736,6 @@ pub(crate) fn emit(
             };
 
             let subopcode = 3;
-            let src = int_reg_enc(src.to_reg());
             emit_std_enc_enc(sink, prefix, opcode, 1, subopcode, src, rex_flags)
         }
 
@@ -717,7 +744,7 @@ pub(crate) fn emit(
             signed,
             divisor,
         } => {
-            let (opcode, prefix, rex_flags) = match size {
+            let (opcode, prefix, mut rex_flags) = match size {
                 1 => (0xF6, LegacyPrefixes::None, RexFlags::clear_w()),
                 2 => (0xF7, LegacyPrefixes::_66, RexFlags::clear_w()),
                 4 => (0xF7, LegacyPrefixes::None, RexFlags::clear_w()),
@@ -732,6 +759,9 @@ pub(crate) fn emit(
             match divisor {
                 RegMem::Reg { reg } => {
                     let src = int_reg_enc(*reg);
+                    if *size == 1 {
+                        rex_flags.always_emit_if_8bit_needed(src);
+                    }
                     emit_std_enc_enc(sink, prefix, opcode, 1, subopcode, src, rex_flags)
                 }
                 RegMem::Mem { addr: src } => {
@@ -987,9 +1017,7 @@ pub(crate) fn emit(
                         ExtMode::BL | ExtMode::BQ => {
                             // A redundant REX prefix must be emitted for certain register inputs.
                             let enc_src = int_reg_enc(*src);
-                            if enc_src >= 4 && enc_src <= 7 {
+                            rex_flags.always_emit_if_8bit_needed(enc_src);
-                                rex_flags.always_emit();
-                            };
                         }
                         _ => {}
                     }
@@ -1084,9 +1112,7 @@ pub(crate) fn emit(
                         ExtMode::BL | ExtMode::BQ => {
                             // A redundant REX prefix must be emitted for certain register inputs.
                             let enc_src = int_reg_enc(*src);
-                            if enc_src >= 4 && enc_src <= 7 {
+                            rex_flags.always_emit_if_8bit_needed(enc_src);
-                                rex_flags.always_emit();
-                            };
                         }
                         _ => {}
                     }
@@ -1130,9 +1156,7 @@ pub(crate) fn emit(
                     let mut rex = RexFlags::clear_w();
 
                     let enc_src = int_reg_enc(*src);
-                    if enc_src >= 4 && enc_src <= 7 {
+                    rex.always_emit_if_8bit_needed(enc_src);
-                        rex.always_emit();
-                    };
 
                     // MOV r8, r/m8 is (REX.W==0) 88 /r
                     emit_std_reg_mem(
@@ -1215,7 +1239,11 @@ pub(crate) fn emit(
             match num_bits {
                 None => {
                     let (opcode, prefix, rex_flags) = match size {
-                        1 => (0xD2, LegacyPrefixes::None, RexFlags::clear_w()),
+                        1 => (
+                            0xD2,
+                            LegacyPrefixes::None,
+                            *RexFlags::clear_w().always_emit_if_8bit_needed(enc_dst),
+                        ),
                         2 => (0xD3, LegacyPrefixes::_66, RexFlags::clear_w()),
                         4 => (0xD3, LegacyPrefixes::None, RexFlags::clear_w()),
                         8 => (0xD3, LegacyPrefixes::None, RexFlags::set_w()),
@@ -1231,7 +1259,11 @@ pub(crate) fn emit(
 
                 Some(num_bits) => {
                     let (opcode, prefix, rex_flags) = match size {
-                        1 => (0xC0, LegacyPrefixes::None, RexFlags::clear_w()),
+                        1 => (
+                            0xC0,
+                            LegacyPrefixes::None,
+                            *RexFlags::clear_w().always_emit_if_8bit_needed(enc_dst),
+                        ),
                         2 => (0xC1, LegacyPrefixes::_66, RexFlags::clear_w()),
                         4 => (0xC1, LegacyPrefixes::None, RexFlags::clear_w()),
                         8 => (0xC1, LegacyPrefixes::None, RexFlags::set_w()),
@@ -1330,9 +1362,7 @@ pub(crate) fn emit(
                     let mut rex = RexFlags::clear_w();
                     // Here, a redundant REX prefix changes the meaning of the instruction.
                     let enc_g = int_reg_enc(*reg_g);
-                    if enc_g >= 4 && enc_g <= 7 {
+                    rex.always_emit_if_8bit_needed(enc_g);
-                        rex.always_emit();
-                    }
                     rex
                 }
                 _ => panic!("x64::Inst::Cmp_RMI_R::emit: unreachable"),
@@ -1343,9 +1373,7 @@ pub(crate) fn emit(
                     if *size == 1 {
                         // Check whether the E register forces the use of a redundant REX.
                         let enc_e = int_reg_enc(*reg_e);
-                        if enc_e >= 4 && enc_e <= 7 {
+                        rex.always_emit_if_8bit_needed(enc_e);
-                            rex.always_emit();
-                        }
                     }
 
                     // Use the swapped operands encoding for CMP, to stay consistent with the output of
@@ -2761,9 +2789,7 @@ pub(crate) fn emit(
                 types::I8 => {
                     let mut rex_flags = RexFlags::clear_w();
                     let enc_src = int_reg_enc(*src);
-                    if enc_src >= 4 && enc_src <= 7 {
+                    rex_flags.always_emit_if_8bit_needed(enc_src);
-                        rex_flags.always_emit();
-                    };
                     (LegacyPrefixes::_F0, rex_flags, 0x0FB0)
                 }
                 types::I16 => (LegacyPrefixes::_66F0, RexFlags::clear_w(), 0x0FB1),

cranelift/codegen/src/isa/x64/inst/emit_tests.rs

@@ -1025,6 +1025,56 @@ fn test_x64_emit() {
         "4C09FA",
         "orq     %r15, %rdx",
     ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::And8, RegMemImm::reg(r15), w_rdx),
+        "4420FA",
+        "andb    %r15b, %dl",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::And8, RegMemImm::reg(rax), w_rsi),
+        "4020C6",
+        "andb    %al, %sil",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::And8, RegMemImm::reg(rax), w_rbx),
+        "20C3",
+        "andb    %al, %bl",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(
+            false,
+            AluRmiROpcode::And8,
+            RegMemImm::mem(Amode::imm_reg(0, rax)),
+            w_rbx,
+        ),
+        "2218",
+        "andb    0(%rax), %bl",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::Or8, RegMemImm::reg(r15), w_rdx),
+        "4408FA",
+        "orb     %r15b, %dl",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::Or8, RegMemImm::reg(rax), w_rsi),
+        "4008C6",
+        "orb     %al, %sil",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(false, AluRmiROpcode::Or8, RegMemImm::reg(rax), w_rbx),
+        "08C3",
+        "orb     %al, %bl",
+    ));
+    insns.push((
+        Inst::alu_rmi_r(
+            false,
+            AluRmiROpcode::Or8,
+            RegMemImm::mem(Amode::imm_reg(0, rax)),
+            w_rbx,
+        ),
+        "0A18",
+        "orb     0(%rax), %bl",
+    ));
     insns.push((
         Inst::alu_rmi_r(true, AluRmiROpcode::Xor, RegMemImm::reg(r15), w_rdx),
         "4C31FA",
@@ -1193,6 +1243,16 @@ fn test_x64_emit() {
         "66F7D7",
         "notw    %di",
     ));
+    insns.push((
+        Inst::not(1, Writable::from_reg(regs::rdi())),
+        "40F6D7",
+        "notb    %dil",
+    ));
+    insns.push((
+        Inst::not(1, Writable::from_reg(regs::rax())),
+        "F6D0",
+        "notb    %al",
+    ));
 
     // ========================================================
     // Neg
@@ -1216,6 +1276,16 @@ fn test_x64_emit() {
         "66F7DF",
         "negw    %di",
     ));
+    insns.push((
+        Inst::neg(1, Writable::from_reg(regs::rdi())),
+        "40F6DF",
+        "negb    %dil",
+    ));
+    insns.push((
+        Inst::neg(1, Writable::from_reg(regs::rax())),
+        "F6D8",
+        "negb    %al",
+    ));
 
     // ========================================================
     // Div
@@ -1239,6 +1309,16 @@ fn test_x64_emit() {
         "48F7F7",
         "div     %rdi",
     ));
+    insns.push((
+        Inst::div(1, false, RegMem::reg(regs::rax())),
+        "F6F0",
+        "div     %al",
+    ));
+    insns.push((
+        Inst::div(1, false, RegMem::reg(regs::rsi())),
+        "40F6F6",
+        "div     %sil",
+    ));
 
     // ========================================================
     // MulHi
@@ -2352,9 +2432,14 @@ fn test_x64_emit() {
     ));
     insns.push((
         Inst::shift_r(1, ShiftKind::RotateRight, None, w_rsi),
-        "D2CE",
+        "40D2CE",
         "rorb    %cl, %sil",
     ));
+    insns.push((
+        Inst::shift_r(1, ShiftKind::RotateRight, None, w_rax),
+        "D2C8",
+        "rorb    %cl, %al",
+    ));
     insns.push((
         Inst::shift_r(1, ShiftKind::RotateRight, Some(5), w_r15),
         "41C0CF05",

cranelift/codegen/src/isa/x64/inst/mod.rs

@@ -1243,6 +1243,14 @@ impl PrettyPrint for Inst {
             (if is_64 { "q" } else { "l" }).to_string()
         }
 
+        fn suffix_lqb(is_64: bool, is_8: bool) -> String {
+            match (is_64, is_8) {
+                (_, true) => "b".to_string(),
+                (true, false) => "q".to_string(),
+                (false, false) => "l".to_string(),
+            }
+        }
+
         fn size_lq(is_64: bool) -> u8 {
             if is_64 {
                 8
@@ -1251,6 +1259,16 @@ impl PrettyPrint for Inst {
             }
         }
 
+        fn size_lqb(is_64: bool, is_8: bool) -> u8 {
+            if is_8 {
+                1
+            } else if is_64 {
+                8
+            } else {
+                4
+            }
+        }
+
         fn suffix_bwlq(size: u8) -> String {
             match size {
                 1 => "b".to_string(),
@@ -1271,9 +1289,9 @@ impl PrettyPrint for Inst {
                 dst,
             } => format!(
                 "{} {}, {}",
-                ljustify2(op.to_string(), suffix_lq(*is_64)),
+                ljustify2(op.to_string(), suffix_lqb(*is_64, op.is_8bit())),
-                src.show_rru_sized(mb_rru, size_lq(*is_64)),
+                src.show_rru_sized(mb_rru, size_lqb(*is_64, op.is_8bit())),
-                show_ireg_sized(dst.to_reg(), mb_rru, size_lq(*is_64)),
+                show_ireg_sized(dst.to_reg(), mb_rru, size_lqb(*is_64, op.is_8bit())),
             ),
 
             Inst::UnaryRmR { src, dst, op, size } => format!(
@@ -2065,6 +2083,17 @@ impl Amode {
             }
         }
     }
+
+    /// Offset the amode by a fixed offset.
+    pub(crate) fn offset(&self, offset: u32) -> Self {
+        let mut ret = self.clone();
+        match &mut ret {
+            &mut Amode::ImmReg { ref mut simm32, .. } => *simm32 += offset,
+            &mut Amode::ImmRegRegShift { ref mut simm32, .. } => *simm32 += offset,
+            _ => panic!("Cannot offset amode: {:?}", self),
+        }
+        ret
+    }
 }
 
 impl RegMemImm {
@@ -2548,77 +2577,88 @@ impl MachInst for Inst {
         ty: Type,
         mut alloc_tmp: F,
     ) -> SmallVec<[Self; 4]> {
-        // We don't support 128-bit constants.
-        assert!(value <= u64::MAX as u128);
         let mut ret = SmallVec::new();
-        let to_reg = to_regs
+        if ty == types::I128 {
-            .only_reg()
+            ret.push(Inst::imm(
-            .expect("multi-reg values not supported on x64");
+                OperandSize::Size64,
-        if ty == types::F32 {
+                value as u64,
-            if value == 0 {
+                to_regs.regs()[0],
-                ret.push(Inst::xmm_rm_r(
+            ));
-                    SseOpcode::Xorps,
+            ret.push(Inst::imm(
-                    RegMem::reg(to_reg.to_reg()),
+                OperandSize::Size64,
-                    to_reg,
+                (value >> 64) as u64,
-                ));
+                to_regs.regs()[1],
-            } else {
+            ));
-                let tmp = alloc_tmp(types::I32);
-                ret.push(Inst::imm(OperandSize::Size32, value as u64, tmp));
-
-                ret.push(Inst::gpr_to_xmm(
-                    SseOpcode::Movd,
-                    RegMem::reg(tmp.to_reg()),
-                    OperandSize::Size32,
-                    to_reg,
-                ));
-            }
-        } else if ty == types::F64 {
-            if value == 0 {
-                ret.push(Inst::xmm_rm_r(
-                    SseOpcode::Xorpd,
-                    RegMem::reg(to_reg.to_reg()),
-                    to_reg,
-                ));
-            } else {
-                let tmp = alloc_tmp(types::I64);
-                ret.push(Inst::imm(OperandSize::Size64, value as u64, tmp));
-
-                ret.push(Inst::gpr_to_xmm(
-                    SseOpcode::Movq,
-                    RegMem::reg(tmp.to_reg()),
-                    OperandSize::Size64,
-                    to_reg,
-                ));
-            }
         } else {
-            // Must be an integer type.
+            let to_reg = to_regs
-            debug_assert!(
+                .only_reg()
-                ty == types::B1
+                .expect("multi-reg values not supported on x64");
-                    || ty == types::I8
+            if ty == types::F32 {
-                    || ty == types::B8
+                if value == 0 {
-                    || ty == types::I16
+                    ret.push(Inst::xmm_rm_r(
-                    || ty == types::B16
+                        SseOpcode::Xorps,
-                    || ty == types::I32
+                        RegMem::reg(to_reg.to_reg()),
-                    || ty == types::B32
+                        to_reg,
-                    || ty == types::I64
+                    ));
-                    || ty == types::B64
+                } else {
-                    || ty == types::R32
+                    let tmp = alloc_tmp(types::I32);
-                    || ty == types::R64
+                    ret.push(Inst::imm(OperandSize::Size32, value as u64, tmp));
-            );
+
-            if value == 0 {
+                    ret.push(Inst::gpr_to_xmm(
-                ret.push(Inst::alu_rmi_r(
+                        SseOpcode::Movd,
-                    ty == types::I64,
+                        RegMem::reg(tmp.to_reg()),
-                    AluRmiROpcode::Xor,
+                        OperandSize::Size32,
-                    RegMemImm::reg(to_reg.to_reg()),
+                        to_reg,
-                    to_reg,
+                    ));
-                ));
+                }
+            } else if ty == types::F64 {
+                if value == 0 {
+                    ret.push(Inst::xmm_rm_r(
+                        SseOpcode::Xorpd,
+                        RegMem::reg(to_reg.to_reg()),
+                        to_reg,
+                    ));
+                } else {
+                    let tmp = alloc_tmp(types::I64);
+                    ret.push(Inst::imm(OperandSize::Size64, value as u64, tmp));
+
+                    ret.push(Inst::gpr_to_xmm(
+                        SseOpcode::Movq,
+                        RegMem::reg(tmp.to_reg()),
+                        OperandSize::Size64,
+                        to_reg,
+                    ));
+                }
             } else {
-                let value = value as u64;
+                // Must be an integer type.
-                ret.push(Inst::imm(
+                debug_assert!(
-                    OperandSize::from_bytes(ty.bytes()),
+                    ty == types::B1
-                    value.into(),
+                        || ty == types::I8
-                    to_reg,
+                        || ty == types::B8
-                ));
+                        || ty == types::I16
+                        || ty == types::B16
+                        || ty == types::I32
+                        || ty == types::B32
+                        || ty == types::I64
+                        || ty == types::B64
+                        || ty == types::R32
+                        || ty == types::R64
+                );
+                if value == 0 {
+                    ret.push(Inst::alu_rmi_r(
+                        ty == types::I64,
+                        AluRmiROpcode::Xor,
+                        RegMemImm::reg(to_reg.to_reg()),
+                        to_reg,
+                    ));
+                } else {
+                    let value = value as u64;
+                    ret.push(Inst::imm(
+                        OperandSize::from_bytes(ty.bytes()),
+                        value.into(),
+                        to_reg,
+                    ));
+                }
             }
         }
         ret

cranelift/filetests/filetests/isa/x64/bitops-i128-run.clif

@@ -0,0 +1,27 @@
+test run
+target x86_64
+feature "experimental_x64"
+
+function %ctz(i64, i64) -> i8 {
+block0(v0: i64, v1: i64):
+    v2 = iconcat v0, v1
+    v3 = ctz.i128 v2
+    v4 = ireduce.i8 v3
+    return v4
+}
+; run: %ctz(0x00000000_00000000, 0x00000001_00000000) == 96
+; run: %ctz(0x00000000_00010000, 0x00000001_00000000) == 16
+; run: %ctz(0x00000000_00010000, 0x00000000_00000000) == 16
+; run: %ctz(0x00000000_00000000, 0x00000000_00000000) == 128
+
+function %clz(i64, i64) -> i8 {
+block0(v0: i64, v1: i64):
+    v2 = iconcat v0, v1
+    v3 = clz.i128 v2
+    v4 = ireduce.i8 v3
+    return v4
+}
+; run: %clz(0x00000000_00000000, 0x00000001_00000000) == 31
+; run: %clz(0x00000000_00010000, 0x00000001_00000000) == 31
+; run: %clz(0x00000000_00010000, 0x00000000_00000000) == 111
+; run: %clz(0x00000000_00000000, 0x00000000_00000000) == 128

cranelift/filetests/filetests/isa/x64/bitrev-i128-run.clif

@@ -0,0 +1,47 @@
+test run
+target x86_64
+feature "experimental_x64"
+
+function %reverse_bits_zero() -> b1 {
+block0:
+    v0 = iconst.i64 0
+    v1 = iconcat v0, v0
+    v2 = bitrev.i128 v1
+    v3 = icmp eq v2, v1
+    return v3
+}
+; run
+
+function %reverse_bits_one() -> b1 {
+block0:
+    v0 = iconst.i64 0
+    v1 = iconst.i64 1
+    v2 = iconcat v0, v1
+
+    v3 = bitrev.i128 v2
+
+    v4 = iconst.i64 0x8000_0000_0000_0000
+    v5 = iconst.i64 0
+    v6 = iconcat v4, v5
+
+    v7 = icmp eq v3, v6
+    return v7
+}
+; run
+
+function %reverse_bits() -> b1 {
+block0:
+    v0 = iconst.i64 0x06AD_8667_69EC_41BA
+    v1 = iconst.i64 0x6C83_D81A_6E28_83AB
+    v2 = iconcat v0, v1
+
+    v3 = bitrev.i128 v2
+
+    v4 = iconst.i64 0xD5C11476581BC136
+    v5 = iconst.i64 0x5D823796E661B560
+    v6 = iconcat v4, v5
+
+    v7 = icmp eq v3, v6
+    return v7
+}
+; run

cranelift/filetests/filetests/isa/x64/floating-point.clif

@@ -0,0 +1,26 @@
+test compile
+target x86_64
+feature "experimental_x64"
+
+function %f(f64) -> f64 {
+block0(v0: f64):
+    v1 = fabs.f64 v0
+    return v1
+}
+; check:  movabsq $$9223372036854775807, %rsi
+; nextln: movq    %rsi, %xmm1
+; nextln: andpd   %xmm0, %xmm1
+; nextln: movaps  %xmm1, %xmm0
+
+
+function %f(i64) -> f64 {
+block0(v0: i64):
+    v1 = load.f64 v0
+    v2 = fabs.f64 v1
+    return v2
+}
+; check:  movsd   0(%rdi), %xmm0
+; nextln: movabsq $$9223372036854775807, %rsi
+; nextln: movq    %rsi, %xmm1
+; nextln: andpd   %xmm0, %xmm1
+; nextln: movaps  %xmm1, %xmm0

cranelift/filetests/filetests/isa/x64/select-i128.clif

@@ -0,0 +1,29 @@
+test compile
+target x86_64
+feature "experimental_x64"
+
+function %f0(i32, i128, i128) -> i128 {
+; check:   pushq   %rbp
+; nextln:  movq    %rsp, %rbp
+
+block0(v0: i32, v1: i128, v2: i128):
+
+    v3 = iconst.i32 42
+    v4 = icmp.i32 eq v0, v3
+; nextln: movl    $$42, %eax
+; nextln: cmpl    %eax, %edi
+
+    v5 = select.i128 v4, v1, v2
+; nextln: cmovzq  %rsi, %rcx
+; nextln: cmovzq  %rdx, %r8
+
+    return v5
+; nextln: movq    %rcx, %rax
+; nextln: movq    %r8, %rdx
+
+; nextln: movq    %rbp, %rsp
+; nextln: popq    %rbp
+; nextln: ret
+ 
+}
+

cranelift/filetests/filetests/isa/x64/shift-i128-run.clif

@@ -0,0 +1,106 @@
+test run
+target x86_64
+feature "experimental_x64"
+
+function %ishl1() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconcat v0, v0
+    v2 = iconst.i32 2
+    v3 = ishl.i128 v1, v2
+    v4 = iconst.i64 0x04040404_04040404
+    v5 = iconcat v4, v4
+    v6 = icmp eq v3, v5
+    return v6
+}
+; run
+
+function %ishl2() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconst.i64 0x01010101_01010101
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 9
+    v4 = ishl.i128 v2, v3
+    v5 = iconst.i64 0x02020202_02020200
+    v6 = iconst.i64 0x02020202_02020202
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run
+
+function %ishl3() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconst.i64 0xffffffff_ffffffff
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 66
+    v4 = ishl.i128 v2, v3
+    v5 = iconst.i64 0x00000000_00000000
+    v6 = iconst.i64 0x04040404_04040404
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run
+
+function %ushr1() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconst.i64 0x01010101_01010101
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 2
+    v4 = ushr.i128 v2, v3
+    v5 = iconst.i64 0x40404040_40404040
+    v6 = iconst.i64 0x00404040_40404040
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run
+
+function %ushr2() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconst.i64 0x01010101_01010101
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 66
+    v4 = ushr.i128 v2, v3
+    v5 = iconst.i64 0x00404040_40404040
+    v6 = iconst.i64 0x00000000_00000000
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run
+
+function %sshr1() -> b1 {
+block0:
+    v0 = iconst.i64 0x01010101_01010101
+    v1 = iconst.i64 0x81010101_01010101
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 2
+    v4 = sshr.i128 v2, v3
+    v5 = iconst.i64 0x40404040_40404040
+    v6 = iconst.i64 0xe0404040_40404040
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run
+
+function %sshr2() -> b1 {
+block0:
+    v0 = iconst.i64 0x12345678_9abcdef0
+    v1 = iconst.i64 0x80101010_10101010
+    v2 = iconcat v0, v1
+    v3 = iconst.i32 66
+    v4 = sshr.i128 v2, v3
+    v5 = iconst.i64 0xe0040404_04040404
+    v6 = iconst.i64 0xffffffff_ffffffff
+    v7 = iconcat v5, v6
+    v8 = icmp eq v4, v7
+    return v8
+}
+; run