Implement iabs in ISLE (AArch64) (#4399)

* Implement `iabs` in ISLE (AArch64)

Converts the existing implementation of `iabs` for AArch64 into ISLE,
and fixes support for `iabs` on scalar values.

Copyright (c) 2022 Arm Limited.

* Improve scalar `iabs` implementation.

Also introduces `CSNeg` instruction.

Copyright (c) 2022 Arm Limited

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

@@ -187,6 +187,13 @@
         (rn Reg)
         (rm Reg))
 
+       ;; A conditional-select negation operation.
+       (CSNeg
+        (rd WritableReg)
+        (cond Cond)
+        (rn Reg)
+        (rm Reg))
+
        ;; A conditional-set operation.
        (CSet
         (rd WritableReg)
@@ -1534,11 +1541,15 @@
          (MInst.AluRRR (ALUOp.SubS) (operand_size ty) dst src1 src2)
          dst)))
 
+(decl cmp_imm (OperandSize Reg Imm12) ProducesFlags)
+(rule (cmp_imm size src1 src2)
+      (ProducesFlags.ProducesFlagsSideEffect
+       (MInst.AluRRImm12 (ALUOp.SubS) size (writable_zero_reg)
+        src1 src2)))
+
 (decl cmp64_imm (Reg Imm12) ProducesFlags)
 (rule (cmp64_imm src1 src2)
-      (ProducesFlags.ProducesFlagsSideEffect
-       (MInst.AluRRImm12 (ALUOp.SubS) (OperandSize.Size64) (writable_zero_reg)
-        src1 src2)))
+      (cmp_imm (OperandSize.Size64) src1 src2))
 
 ;; Helper for emitting `sbc` instructions.
 (decl sbc_paired (Type Reg Reg) ConsumesFlags)
@@ -1681,6 +1692,18 @@
          (MInst.CSel dst cond if_true if_false)
          dst)))
 
+;; Helper for generating a `CSNeg` instruction.
+;;
+;; Note that this doesn't actually emit anything, instead it produces a
+;; `ConsumesFlags` instruction which must be consumed with `with_flags*`
+;; helpers.
+(decl csneg (Cond Reg Reg) ConsumesFlags)
+(rule (csneg cond if_true if_false)
+      (let ((dst WritableReg (temp_writable_reg $I64)))
+        (ConsumesFlags.ConsumesFlagsReturnsReg
+         (MInst.CSNeg dst cond if_true if_false)
+         dst)))
+
 ;; Helpers for generating `add` instructions.
 
 (decl add (Type Reg Reg) Reg)
@@ -1769,6 +1792,17 @@
 (decl addp (Reg Reg VectorSize) Reg)
 (rule (addp x y size) (vec_rrr (VecALUOp.Addp) x y size))
 
+;; Helper for generating vector `abs` instructions.
+(decl vec_abs (Reg VectorSize) Reg)
+(rule (vec_abs x size) (vec_misc (VecMisc2.Abs) x size))
+
+;; Helper for generating instruction sequences to calculate a scalar absolute
+;; value.
+(decl abs (OperandSize Reg) Reg)
+(rule (abs size x)
+      (value_regs_get (with_flags (cmp_imm size x (u8_into_imm12 0))
+                                  (csneg (Cond.Gt) x x)) 0))
+
 ;; Helper for generating `addv` instructions.
 (decl addv (Reg VectorSize) Reg)
 (rule (addv x size) (vec_lanes (VecLanesOp.Addv) x size))

cranelift/codegen/src/isa/aarch64/inst/emit.rs

@@ -331,12 +331,16 @@ pub(crate) fn enc_adr(off: i32, rd: Writable<Reg>) -> u32 {
     (0b00010000 << 24) | (immlo << 29) | (immhi << 5) | machreg_to_gpr(rd.to_reg())
 }
 
-fn enc_csel(rd: Writable<Reg>, rn: Reg, rm: Reg, cond: Cond) -> u32 {
+fn enc_csel(rd: Writable<Reg>, rn: Reg, rm: Reg, cond: Cond, op: u32, o2: u32) -> u32 {
+    debug_assert_eq!(op & 0b1, op);
+    debug_assert_eq!(o2 & 0b1, o2);
     0b100_11010100_00000_0000_00_00000_00000
+        | (op << 30)
         | (machreg_to_gpr(rm) << 16)
+        | (cond.bits() << 12)
+        | (o2 << 10)
         | (machreg_to_gpr(rn) << 5)
         | machreg_to_gpr(rd.to_reg())
-        | (cond.bits() << 12)
 }
 
 fn enc_fcsel(rd: Writable<Reg>, rn: Reg, rm: Reg, cond: Cond, size: ScalarSize) -> u32 {
@@ -348,18 +352,6 @@ fn enc_fcsel(rd: Writable<Reg>, rn: Reg, rm: Reg, cond: Cond, size: ScalarSize)
         | (cond.bits() << 12)
 }
 
-fn enc_cset(rd: Writable<Reg>, cond: Cond) -> u32 {
-    0b100_11010100_11111_0000_01_11111_00000
-        | machreg_to_gpr(rd.to_reg())
-        | (cond.invert().bits() << 12)
-}
-
-fn enc_csetm(rd: Writable<Reg>, cond: Cond) -> u32 {
-    0b110_11010100_11111_0000_00_11111_00000
-        | machreg_to_gpr(rd.to_reg())
-        | (cond.invert().bits() << 12)
-}
-
 fn enc_ccmp_imm(size: OperandSize, rn: Reg, imm: UImm5, nzcv: NZCV, cond: Cond) -> u32 {
     0b0_1_1_11010010_00000_0000_10_00000_0_0000
         | size.sf_bit() << 31
@@ -1352,15 +1344,21 @@ impl MachInstEmit for Inst {
                 let rd = allocs.next_writable(rd);
                 let rn = allocs.next(rn);
                 let rm = allocs.next(rm);
-                sink.put4(enc_csel(rd, rn, rm, cond));
+                sink.put4(enc_csel(rd, rn, rm, cond, 0, 0));
+            }
+            &Inst::CSNeg { rd, rn, rm, cond } => {
+                let rd = allocs.next_writable(rd);
+                let rn = allocs.next(rn);
+                let rm = allocs.next(rm);
+                sink.put4(enc_csel(rd, rn, rm, cond, 1, 1));
             }
             &Inst::CSet { rd, cond } => {
                 let rd = allocs.next_writable(rd);
-                sink.put4(enc_cset(rd, cond));
+                sink.put4(enc_csel(rd, zero_reg(), zero_reg(), cond.invert(), 0, 1));
             }
             &Inst::CSetm { rd, cond } => {
                 let rd = allocs.next_writable(rd);
-                sink.put4(enc_csetm(rd, cond));
+                sink.put4(enc_csel(rd, zero_reg(), zero_reg(), cond.invert(), 1, 0));
             }
             &Inst::CCmpImm {
                 size,

cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs

@@ -2129,6 +2129,16 @@ fn test_aarch64_binemit() {
         "8A218E9A",
         "csel x10, x12, x14, hs",
     ));
+    insns.push((
+        Inst::CSNeg {
+            rd: writable_xreg(10),
+            rn: xreg(12),
+            rm: xreg(14),
+            cond: Cond::Hs,
+        },
+        "8A258EDA",
+        "csneg x10, x12, x14, hs",
+    ));
     insns.push((
         Inst::CSet {
             rd: writable_xreg(15),

cranelift/codegen/src/isa/aarch64/inst/mod.rs

@@ -669,6 +669,11 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
             collector.reg_use(rn);
             collector.reg_use(rm);
         }
+        &Inst::CSNeg { rd, rn, rm, .. } => {
+            collector.reg_def(rd);
+            collector.reg_use(rn);
+            collector.reg_use(rm);
+        }
         &Inst::CSet { rd, .. } | &Inst::CSetm { rd, .. } => {
             collector.reg_def(rd);
         }
@@ -1510,6 +1515,13 @@ impl Inst {
                 let cond = cond.pretty_print(0, allocs);
                 format!("csel {}, {}, {}, {}", rd, rn, rm, cond)
             }
+            &Inst::CSNeg { rd, rn, rm, cond } => {
+                let rd = pretty_print_ireg(rd.to_reg(), OperandSize::Size64, allocs);
+                let rn = pretty_print_ireg(rn, OperandSize::Size64, allocs);
+                let rm = pretty_print_ireg(rm, OperandSize::Size64, allocs);
+                let cond = cond.pretty_print(0, allocs);
+                format!("csneg {}, {}, {}, {}", rd, rn, rm, cond)
+            }
             &Inst::CSet { rd, cond } => {
                 let rd = pretty_print_ireg(rd.to_reg(), OperandSize::Size64, allocs);
                 let cond = cond.pretty_print(0, allocs);

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

@@ -156,6 +156,17 @@
 (rule (lower (has_type ty (iadd_pairwise x y)))
       (addp x y (vector_size ty)))
 
+;;;; Rules for `iabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (iabs x)))
+      (vec_abs x (vector_size ty)))
+
+(rule (lower (has_type $I64 (iabs x)))
+      (abs (OperandSize.Size64) x))
+
+(rule (lower (has_type (fits_in_32 ty) (iabs x)))
+      (abs (OperandSize.Size32) (put_in_reg_sext32 x)))
+
 ;;;; Rules for `isub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; `i64` and smaller

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

@@ -1941,17 +1941,7 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             panic!("ALU+imm and ALU+carry ops should not appear here!");
         }
 
-        Opcode::Iabs => {
-            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-            let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
-            let ty = ty.unwrap();
-            ctx.emit(Inst::VecMisc {
-                op: VecMisc2::Abs,
-                rd,
-                rn,
-                size: VectorSize::from_ty(ty),
-            });
-        }
+        Opcode::Iabs => implemented_in_isle(ctx),
         Opcode::AvgRound => {
             let ty = ty.unwrap();
 

cranelift/filetests/filetests/isa/aarch64/iabs.clif

@@ -0,0 +1,119 @@
+test compile precise-output
+set unwind_info=false
+target aarch64
+
+function %f1(i8x16) -> i8x16 {
+block0(v0: i8x16):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.16b, v0.16b
+;   ret
+
+function %f2(i8x8) -> i8x8 {
+block0(v0: i8x8):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.8b, v0.8b
+;   ret
+
+function %f3(i16x8) -> i16x8 {
+block0(v0: i16x8):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.8h, v0.8h
+;   ret
+
+function %f4(i16x4) -> i16x4 {
+block0(v0: i16x4):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.4h, v0.4h
+;   ret
+
+function %f5(i32x4) -> i32x4 {
+block0(v0: i32x4):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.4s, v0.4s
+;   ret
+
+function %f6(i32x2) -> i32x2 {
+block0(v0: i32x2):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.2s, v0.2s
+;   ret
+
+function %f7(i64x2) -> i64x2 {
+block0(v0: i64x2):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   abs v0.2d, v0.2d
+;   ret
+
+function %f8(i8) -> i8 {
+block0(v0: i8):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   sxtb w3, w0
+;   subs wzr, w3, #0
+;   csneg x0, x3, x3, gt
+;   ret
+
+function %f9(i16) -> i16 {
+block0(v0: i16):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   sxth w3, w0
+;   subs wzr, w3, #0
+;   csneg x0, x3, x3, gt
+;   ret
+
+function %f10(i32) -> i32 {
+block0(v0: i32):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   subs wzr, w0, #0
+;   csneg x0, x0, x0, gt
+;   ret
+
+function %f11(i64) -> i64 {
+block0(v0: i64):
+  v1 = iabs v0
+  return v1
+}
+
+; block0:
+;   subs xzr, x0, #0
+;   csneg x0, x0, x0, gt
+;   ret

cranelift/filetests/filetests/runtests/iabs.clif

@@ -1,5 +1,7 @@
 test interpret
-; aarch64 & x86_64 only support vector iabs
+test run
+target aarch64
+; x86_64 only supports vector iabs
 
 function %iabs_i8(i8) -> i8 {
 block0(v0: i8):