Port icmp to ISLE (AArch64) (#4898)

* Port `icmp` to ISLE (AArch64)

Ported the existing implementation of `icmp` (and, by extension, the
`lower_icmp` function) to ISLE for AArch64.

Copyright (c) 2022 Arm Limited

* Allow 'producer chains', eliminating `Nop0`s

Copyright (c) 2022 Arm Limited

cranelift/codegen/src/isa/aarch64/inst.isle

@@ -1960,6 +1960,12 @@
        (MInst.AluRRImm12 (ALUOp.AddS) size (writable_zero_reg)
         src1 src2)))
 
+(decl cmp (OperandSize Reg Reg) ProducesFlags)
+(rule (cmp size src1 src2)
+      (ProducesFlags.ProducesFlagsSideEffect
+       (MInst.AluRRR (ALUOp.SubS) size (writable_zero_reg)
+        src1 src2)))
+
 (decl cmp_imm (OperandSize Reg Imm12) ProducesFlags)
 (rule (cmp_imm size src1 src2)
       (ProducesFlags.ProducesFlagsSideEffect
@@ -1970,6 +1976,12 @@
 (rule (cmp64_imm src1 src2)
       (cmp_imm (OperandSize.Size64) src1 src2))
 
+(decl cmp_extend (OperandSize Reg Reg ExtendOp) ProducesFlags)
+(rule (cmp_extend size src1 src2 extend)
+      (ProducesFlags.ProducesFlagsSideEffect
+       (MInst.AluRRRExtend (ALUOp.SubS) size (writable_zero_reg)
+        src1 src2 extend)))
+
 ;; Helper for emitting `sbc` instructions.
 (decl sbc_paired (Type Reg Reg) ConsumesFlags)
 (rule (sbc_paired ty src1 src2)
@@ -2199,6 +2211,13 @@
          (MInst.CSNeg dst cond if_true if_false)
          dst)))
 
+;; Helper for generating `MInst.CCmp` instructions.
+;; Creates a new `ProducesFlags` from the supplied `ProducesFlags` followed
+;; immediately by the `MInst.CCmp` instruction.
+(decl ccmp (OperandSize Reg Reg NZCV Cond ProducesFlags) ProducesFlags)
+(rule (ccmp size rn rm nzcv cond inst_input)
+      (produces_flags_append inst_input (MInst.CCmp size rn rm nzcv cond)))
+
 ;; Helper for generating `MInst.CCmpImm` instructions.
 (decl ccmp_imm (OperandSize u8 Reg UImm5 NZCV Cond) ConsumesFlags)
 (rule (ccmp_imm size 1 rn imm nzcv cond)
@@ -2845,6 +2864,11 @@
 ;; TODO: Port lower_fp_condcode() to ISLE.
 (extern constructor fp_cond_code fp_cond_code)
 
+;; Lower an integer cond code.
+(decl cond_code (IntCC) Cond)
+;; TODO: Port lower_condcode() to ISLE.
+(extern constructor cond_code cond_code)
+
 ;; Generate comparison to zero operator from input condition code
 (decl float_cc_cmp_zero_to_vec_misc_op (FloatCC) VecMisc2)
 (extern constructor float_cc_cmp_zero_to_vec_misc_op float_cc_cmp_zero_to_vec_misc_op)
@@ -3280,3 +3304,169 @@
       (let ((dst WritableReg (temp_writable_reg $I64))
             (_ Unit (emit (MInst.ElfTlsGetAddr name dst))))
         dst))
+
+;; Helpers for lowering `icmp` sequences.
+;; `lower_icmp` contains shared functionality for lowering `icmp`
+;; sequences, which `lower_icmp_into_{reg,flags}` extend from.
+(decl lower_icmp (IntCC Value Value Type) ProducesFlags)
+(decl lower_icmp_into_reg (IntCC Value Value Type Type) ValueRegs)
+(decl lower_icmp_into_flags (IntCC Value Value Type) ProducesFlags)
+;; For most cases, `lower_icmp_into_flags` is the same as `lower_icmp`,
+;; except for some I128 cases (see below).
+(rule -1 (lower_icmp_into_flags cond x y ty) (lower_icmp cond x y ty))
+
+;; Vectors.
+;; `icmp` into flags for vectors is invalid.
+(rule (lower_icmp_into_reg cond x y in_ty @ (multi_lane _ _) _out_ty)
+      (let ((cond Cond (cond_code cond))
+            (rn Reg (put_in_reg x))
+            (rm Reg (put_in_reg y)))
+       (vec_cmp rn rm in_ty cond)))
+
+;; Determines the appropriate extend op given the value type and whether it is signed.
+(decl lower_icmp_extend (Type bool) ExtendOp)
+(rule (lower_icmp_extend $I8 $true) (ExtendOp.SXTB))
+(rule (lower_icmp_extend $I16 $true) (ExtendOp.SXTH))
+(rule (lower_icmp_extend $I8 $false) (ExtendOp.UXTB))
+(rule (lower_icmp_extend $I16 $false) (ExtendOp.UXTH))
+
+;; Integers <= 64-bits.
+(rule (lower_icmp_into_reg cond rn rm in_ty out_ty)
+      (if (ty_int_bool_ref_scalar_64 in_ty))
+      (let ((cc Cond (cond_code cond)))
+       (with_flags
+        (lower_icmp cond rn rm in_ty)
+        (materialize_bool_result (ty_bits out_ty) cc))))
+
+(rule 1 (lower_icmp cond rn rm (fits_in_16 ty))
+      (if (signed_cond_code cond))
+      (let ((rn Reg (put_in_reg_sext32 rn)))
+      (cmp_extend (operand_size ty) rn rm (lower_icmp_extend ty $true))))
+(rule (lower_icmp cond rn (imm12_from_value rm) (fits_in_16 ty))
+      (let ((rn Reg (put_in_reg_zext32 rn)))
+      (cmp_imm (operand_size ty) rn rm)))
+(rule -1 (lower_icmp cond rn rm (fits_in_16 ty))
+      (let ((rn Reg (put_in_reg_zext32 rn)))
+      (cmp_extend (operand_size ty) rn rm (lower_icmp_extend ty $false))))
+(rule -2 (lower_icmp cond rn (imm12_from_value rm) ty)
+      (if (ty_int_bool_ref_scalar_64 ty))
+      (cmp_imm (operand_size ty) rn rm))
+(rule -3 (lower_icmp cond rn rm ty)
+      (if (ty_int_bool_ref_scalar_64 ty))
+      (cmp (operand_size ty) rn rm))
+
+;; 128-bit integers.
+(rule (lower_icmp_into_reg cond @ (IntCC.Equal) rn rm $I128 out_ty)
+      (let ((cc Cond (cond_code cond)))
+       (with_flags
+        (lower_icmp cond rn rm $I128)
+        (materialize_bool_result (ty_bits out_ty) cc))))
+(rule (lower_icmp_into_reg cond @ (IntCC.NotEqual) rn rm $I128 out_ty)
+      (let ((cc Cond (cond_code cond)))
+       (with_flags
+        (lower_icmp cond rn rm $I128)
+        (materialize_bool_result (ty_bits out_ty) cc))))
+
+;; cmp lhs_lo, rhs_lo
+;; ccmp lhs_hi, rhs_hi, #0, eq
+(decl lower_icmp_i128_eq_ne (Value Value) ProducesFlags)
+(rule (lower_icmp_i128_eq_ne lhs rhs)
+      (let ((lhs_lo Reg (value_regs_get lhs 0))
+            (lhs_hi Reg (value_regs_get lhs 1))
+            (rhs_lo Reg (value_regs_get rhs 0))
+            (rhs_hi Reg (value_regs_get rhs 1))
+            (cmp_inst ProducesFlags (cmp (OperandSize.Size64) lhs_lo rhs_lo)))
+       (ccmp (OperandSize.Size64) lhs_hi rhs_hi
+        (nzcv $false $false $false $false) (Cond.Eq) cmp_inst)))
+
+(rule (lower_icmp (IntCC.Equal) lhs rhs $I128)
+      (lower_icmp_i128_eq_ne lhs rhs))
+(rule (lower_icmp (IntCC.NotEqual) lhs rhs $I128)
+      (lower_icmp_i128_eq_ne lhs rhs))
+
+;; cmp      lhs_lo, rhs_lo
+;; cset     tmp1, unsigned_cond
+;; cmp      lhs_hi, rhs_hi
+;; cset     tmp2, cond
+;; csel     dst, tmp1, tmp2, eq
+(rule -1 (lower_icmp_into_reg cond lhs rhs $I128 out_ty)
+      (let ((unsigned_cond Cond (cond_code (intcc_unsigned cond)))
+            (cond Cond (cond_code cond))
+            (lhs_lo Reg (value_regs_get lhs 0))
+            (lhs_hi Reg (value_regs_get lhs 1))
+            (rhs_lo Reg (value_regs_get rhs 0))
+            (rhs_hi Reg (value_regs_get rhs 1))
+            (tmp1 ValueRegs
+             (with_flags (cmp (OperandSize.Size64) lhs_lo rhs_lo)
+                         (materialize_bool_result
+                          (ty_bits out_ty) unsigned_cond)))
+            (tmp1 Reg (value_regs_get tmp1 0))
+            (dst ValueRegs
+                  (with_flags (cmp (OperandSize.Size64) lhs_hi rhs_hi)
+                   (lower_icmp_i128_consumer cond (ty_bits out_ty)
+                    tmp1 lhs_hi rhs_hi))))
+       dst))
+
+(decl lower_icmp_i128_consumer (Cond u8 Reg Reg Reg) ConsumesFlags)
+(rule (lower_icmp_i128_consumer cond 1 tmp1 lhs_hi rhs_hi)
+      (let ((tmp2 WritableReg (temp_writable_reg $I64))
+            (dst WritableReg (temp_writable_reg $I64)))
+       (ConsumesFlags.ConsumesFlagsTwiceReturnsValueRegs
+        (MInst.CSet tmp2 cond)
+        (MInst.CSel dst (Cond.Eq) tmp1 tmp2)
+        (value_reg dst))))
+(rule (lower_icmp_i128_consumer cond 128 tmp1 lhs_hi rhs_hi)
+      (let ((tmp2 WritableReg (temp_writable_reg $I64))
+            (dst WritableReg (temp_writable_reg $I64)))
+       (ConsumesFlags.ConsumesFlagsTwiceReturnsValueRegs
+        (MInst.CSetm tmp2 cond)
+        (MInst.CSel dst (Cond.Eq) tmp1 tmp2)
+        (value_regs dst dst))))
+(rule -1 (lower_icmp_i128_consumer cond _out_ty_bits tmp1 lhs_hi rhs_hi)
+      (let ((tmp2 WritableReg (temp_writable_reg $I64))
+            (dst WritableReg (temp_writable_reg $I64)))
+       (ConsumesFlags.ConsumesFlagsTwiceReturnsValueRegs
+        (MInst.CSetm tmp2 cond)
+        (MInst.CSel dst (Cond.Eq) tmp1 tmp2)
+        (value_reg dst))))
+
+;; Exceptional `lower_icmp_into_flags` rules.
+;; We need to guarantee that the flags for `cond` are correct, so we
+;; compare `dst` with 1.
+(rule (lower_icmp_into_flags cond @ (IntCC.SignedGreaterThanOrEqual) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0))
+            (tmp Reg (imm $I64 (ImmExtend.Sign) 1))) ;; mov tmp, #1
+       (cmp (OperandSize.Size64) dst tmp)))
+(rule (lower_icmp_into_flags cond @ (IntCC.UnsignedGreaterThanOrEqual) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0))
+            (tmp Reg (imm $I64 (ImmExtend.Zero) 1)))
+       (cmp (OperandSize.Size64) dst tmp)))
+(rule (lower_icmp_into_flags cond @ (IntCC.SignedLessThanOrEqual) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0))
+            (tmp Reg (imm $I64 (ImmExtend.Sign) 1)))
+       (cmp (OperandSize.Size64) tmp dst)))
+(rule (lower_icmp_into_flags cond @ (IntCC.UnsignedLessThanOrEqual) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0))
+            (tmp Reg (imm $I64 (ImmExtend.Zero) 1)))
+       (cmp (OperandSize.Size64) tmp dst)))
+;; For strict comparisons, we compare with 0.
+(rule (lower_icmp_into_flags cond @ (IntCC.SignedGreaterThan) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0)))
+       (cmp (OperandSize.Size64) dst (zero_reg))))
+(rule (lower_icmp_into_flags cond @ (IntCC.UnsignedGreaterThan) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0)))
+       (cmp (OperandSize.Size64) dst (zero_reg))))
+(rule (lower_icmp_into_flags cond @ (IntCC.SignedLessThan) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0)))
+       (cmp (OperandSize.Size64) (zero_reg) dst)))
+(rule (lower_icmp_into_flags cond @ (IntCC.UnsignedLessThan) lhs rhs $I128)
+      (let ((dst ValueRegs (lower_icmp_into_reg cond lhs rhs $I128 $B1))
+            (dst Reg (value_regs_get dst 0)))
+       (cmp (OperandSize.Size64) (zero_reg) dst)))

cranelift/codegen/src/isa/aarch64/lower.isle

@@ -1730,6 +1730,9 @@
             (vec_size VectorSize (vector_size ty)))
           (value_reg (int_cmp_zero_swap cond rn vec_size))))
 
+(rule -1 (lower (has_type out_ty (icmp cond x @ (value_type in_ty) y)))
+      (lower_icmp_into_reg cond x y in_ty out_ty))
+
 ;;;; Rules for `trap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (rule (lower (trap trap_code))

cranelift/codegen/src/isa/aarch64/lower/isle.rs

@@ -6,10 +6,10 @@ use generated_code::Context;
 
 // Types that the generated ISLE code uses via `use super::*`.
 use super::{
-    fp_reg, lower_constant_f128, lower_constant_f32, lower_constant_f64, lower_fp_condcode,
-    stack_reg, writable_link_reg, writable_zero_reg, zero_reg, AMode, ASIMDFPModImm,
-    ASIMDMovModImm, BranchTarget, CallIndInfo, CallInfo, Cond, CondBrKind, ExtendOp, FPUOpRI,
-    FPUOpRIMod, FloatCC, Imm12, ImmLogic, ImmShift, Inst as MInst, IntCC, JTSequenceInfo,
+    fp_reg, lower_condcode, lower_constant_f128, lower_constant_f32, lower_constant_f64,
+    lower_fp_condcode, stack_reg, writable_link_reg, writable_zero_reg, zero_reg, AMode,
+    ASIMDFPModImm, ASIMDMovModImm, BranchTarget, CallIndInfo, CallInfo, Cond, CondBrKind, ExtendOp,
+    FPUOpRI, FPUOpRIMod, FloatCC, Imm12, ImmLogic, ImmShift, Inst as MInst, IntCC, JTSequenceInfo,
     MachLabel, MemLabel, MoveWideConst, MoveWideOp, NarrowValueMode, Opcode, OperandSize,
     PairAMode, Reg, SImm9, ScalarSize, ShiftOpAndAmt, UImm12Scaled, UImm5, VecMisc2, VectorSize,
     NZCV,
@@ -517,6 +517,10 @@ impl Context for IsleContext<'_, '_, MInst, Flags, IsaFlags, 6> {
         lower_fp_condcode(*cc)
     }
 
+    fn cond_code(&mut self, cc: &condcodes::IntCC) -> Cond {
+        lower_condcode(*cc)
+    }
+
     fn preg_sp(&mut self) -> PReg {
         super::regs::stack_reg().to_real_reg().unwrap().into()
     }

cranelift/codegen/src/isa/aarch64/lower_inst.rs

@@ -293,11 +293,7 @@ pub(crate) fn lower_insn_to_regs(
             panic!("Should never reach ifcmp as isel root!");
         }
 
-        Opcode::Icmp => {
-            let condcode = ctx.data(insn).cond_code().unwrap();
-            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-            lower_icmp(ctx, insn, condcode, IcmpOutput::Register(rd))?;
-        }
+        Opcode::Icmp => implemented_in_isle(ctx),
 
         Opcode::Fcmp => implemented_in_isle(ctx),
 

cranelift/codegen/src/isa/x64/inst.isle

@@ -1362,12 +1362,6 @@
 (decl intcc_without_eq (IntCC) IntCC)
 (extern constructor intcc_without_eq intcc_without_eq)
 
-;; This is a direct import of `IntCC::unsigned`.
-;; Get the corresponding IntCC with the signed component removed.
-;; For conditions without a signed component, this is a no-op.
-(decl intcc_unsigned (IntCC) IntCC)
-(extern constructor intcc_unsigned intcc_unsigned)
-
 ;;;; Helpers for Getting Particular Physical Registers ;;;;;;;;;;;;;;;;;;;;;;;;;
 ;;
 ;; These should only be used for legalization purposes, when we can't otherwise

cranelift/codegen/src/isa/x64/lower/isle.rs

@@ -596,11 +596,6 @@ impl Context for IsleContext<'_, '_, MInst, Flags, IsaFlags, 6> {
         x.without_equal()
     }
 
-    #[inline]
-    fn intcc_unsigned(&mut self, x: &IntCC) -> IntCC {
-        x.unsigned()
-    }
-
     #[inline]
     fn intcc_to_cc(&mut self, intcc: &IntCC) -> CC {
         CC::from_intcc(*intcc)