x64: port atomic_cas to ISLE (#4223)

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

@@ -1058,6 +1058,11 @@
 (decl encode_fcmp_imm (FcmpImm) u8)
 (extern constructor encode_fcmp_imm encode_fcmp_imm)
 
+;;;; Registers ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(decl rax () WritableGpr)
+(extern constructor rax rax)
+
 ;;;; Newtypes for Different Register Classes ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (type Gpr (primitive Gpr))
@@ -2876,6 +2881,11 @@
 (rule (x64_mfence)
       (SideEffectNoResult.Inst (MInst.Fence (FenceKind.MFence))))
 
+(decl x64_cmpxchg (Type Gpr Gpr SyntheticAmode) Gpr)
+(rule (x64_cmpxchg ty expected replacement addr)
+      (let ((_ Unit (emit (MInst.LockCmpxchg ty replacement expected addr (rax)))))
+        (rax)))
+
 ;;;; Automatic conversions ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (convert Gpr InstOutput output_gpr)

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

@@ -2812,3 +2812,9 @@
       (side_effect (side_effect_concat
        (x64_movrm ty (to_amode flags address (zero_offset)) value)
        (x64_mfence))))
+
+;; Rules for `atomic_cas` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type (and (fits_in_64 ty) (ty_int _))
+                  (atomic_cas flags address expected replacement)))
+      (x64_cmpxchg ty expected replacement (to_amode flags address (zero_offset))))

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

@@ -2205,32 +2205,7 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
         }
 
         Opcode::AtomicCas => {
-            // This is very similar to, but not identical to, the `AtomicRmw` case.  As with
-            // `AtomicRmw`, there's no need to zero-extend narrow values here.
-            let dst = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-            let addr = lower_to_amode(ctx, inputs[0], 0);
-            let expected = put_input_in_reg(ctx, inputs[1]);
-            let replacement = put_input_in_reg(ctx, inputs[2]);
-            let ty_access = ty.unwrap();
-            assert!(is_valid_atomic_transaction_ty(ty_access));
-
-            // Move the expected value into %rax.  Because there's only one fixed register on
-            // the input side, we don't have to use `ensure_in_vreg`, as is necessary in the
-            // `AtomicRmw` case.
-            ctx.emit(Inst::gen_move(
-                Writable::from_reg(regs::rax()),
-                expected,
-                types::I64,
-            ));
-            ctx.emit(Inst::LockCmpxchg {
-                ty: ty_access,
-                mem: addr.into(),
-                replacement,
-                expected: regs::rax(),
-                dst_old: Writable::from_reg(regs::rax()),
-            });
-            // And finally, copy the old value at the location to its destination reg.
-            ctx.emit(Inst::gen_move(dst, regs::rax(), types::I64));
+            implemented_in_isle(ctx);
         }
 
         Opcode::AtomicLoad => {

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

@@ -547,6 +547,12 @@ where
     fn zero_offset(&mut self) -> Offset32 {
         Offset32::new(0)
     }
+
+    #[inline]
+    fn rax(&mut self) -> WritableGpr {
+        let gpr = Gpr::new(regs::rax()).unwrap();
+        WritableGpr::from_reg(gpr)
+    }
 }
 
 // Since x64 doesn't have 8x16 shifts and we must use a 16x8 shift instead, we