Enable the simd_conversions test for AArch64

Copyright (c) 2021, Arm Limited.

build.rs

@@ -224,8 +224,7 @@ fn ignore(testsuite: &str, testname: &str, strategy: &str) -> bool {
             ("simd", _) if platform_is_s390x() => return true,
 
             // These are new instructions that are not really implemented in any backend.
-            ("simd", "simd_conversions")
+            ("simd", "simd_i16x8_extadd_pairwise_i8x16")
-            | ("simd", "simd_i16x8_extadd_pairwise_i8x16")
             | ("simd", "simd_i32x4_extadd_pairwise_i16x8") => return true,
 
             _ => {}

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

@@ -2128,6 +2128,8 @@ impl MachInstEmit for Inst {
                     VecRRNarrowOp::Uqxtn16 => (0b1, 0b00, 0b10100),
                     VecRRNarrowOp::Uqxtn32 => (0b1, 0b01, 0b10100),
                     VecRRNarrowOp::Uqxtn64 => (0b1, 0b10, 0b10100),
+                    VecRRNarrowOp::Fcvtn32 => (0b0, 0b00, 0b10110),
+                    VecRRNarrowOp::Fcvtn64 => (0b0, 0b01, 0b10110),
                 };
 
                 sink.put4(enc_vec_rr_misc(

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

@@ -2611,6 +2611,28 @@ fn test_aarch64_binemit() {
         "uqxtn2 v11.4s, v12.2d",
     ));
 
+    insns.push((
+        Inst::VecRRNarrow {
+            op: VecRRNarrowOp::Fcvtn32,
+            rd: writable_vreg(0),
+            rn: vreg(0),
+            high_half: false,
+        },
+        "0068210E",
+        "fcvtn v0.4h, v0.4s",
+    ));
+
+    insns.push((
+        Inst::VecRRNarrow {
+            op: VecRRNarrowOp::Fcvtn64,
+            rd: writable_vreg(31),
+            rn: vreg(30),
+            high_half: true,
+        },
+        "DF6B614E",
+        "fcvtn2 v31.4s, v30.2d",
+    ));
+
     insns.push((
         Inst::VecRRPair {
             op: VecPairOp::Addp,

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

@@ -396,6 +396,10 @@ pub enum VecRRNarrowOp {
     Uqxtn32,
     /// Unsigned saturating extract narrow, 64-bit elements
     Uqxtn64,
+    /// Floating-point convert to lower precision narrow, 32-bit elements
+    Fcvtn32,
+    /// Floating-point convert to lower precision narrow, 64-bit elements
+    Fcvtn64,
 }
 
 /// A vector operation on a pair of elements with one register.
@@ -4073,6 +4077,18 @@ impl Inst {
                     (VecRRNarrowOp::Uqxtn64, true) => {
                         ("uqxtn2", VectorSize::Size32x4, VectorSize::Size64x2)
                     }
+                    (VecRRNarrowOp::Fcvtn32, false) => {
+                        ("fcvtn", VectorSize::Size16x4, VectorSize::Size32x4)
+                    }
+                    (VecRRNarrowOp::Fcvtn32, true) => {
+                        ("fcvtn2", VectorSize::Size16x8, VectorSize::Size32x4)
+                    }
+                    (VecRRNarrowOp::Fcvtn64, false) => {
+                        ("fcvtn", VectorSize::Size32x2, VectorSize::Size64x2)
+                    }
+                    (VecRRNarrowOp::Fcvtn64, true) => {
+                        ("fcvtn2", VectorSize::Size32x4, VectorSize::Size64x2)
+                    }
                 };
                 let rd = show_vreg_vector(rd.to_reg(), mb_rru, rd_size);
                 let rn = show_vreg_vector(rn, mb_rru, size);

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

@@ -3555,12 +3555,62 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             });
         }
 
-        Opcode::ConstAddr
+        Opcode::FcvtLowFromSint => {
-        | Opcode::FcvtLowFromSint
+            let ty = ty.unwrap();
-        | Opcode::Fvdemote
+
-        | Opcode::FvpromoteLow
+            if ty != F64X2 {
-        | Opcode::Vconcat
+                return Err(CodegenError::Unsupported(format!(
-        | Opcode::Vsplit => unimplemented!("lowering {}", op),
+                    "FcvtLowFromSint: Unsupported type: {:?}",
+                    ty
+                )));
+            }
+
+            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
+            let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
+
+            ctx.emit(Inst::VecExtend {
+                t: VecExtendOp::Sxtl32,
+                rd,
+                rn,
+                high_half: false,
+            });
+            ctx.emit(Inst::VecMisc {
+                op: VecMisc2::Scvtf,
+                rd,
+                rn: rd.to_reg(),
+                size: VectorSize::Size64x2,
+            });
+        }
+
+        Opcode::FvpromoteLow => {
+            debug_assert_eq!(ty.unwrap(), F64X2);
+
+            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
+            let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
+
+            ctx.emit(Inst::VecRRLong {
+                op: VecRRLongOp::Fcvtl32,
+                rd,
+                rn,
+                high_half: false,
+            });
+        }
+
+        Opcode::Fvdemote => {
+            debug_assert_eq!(ty.unwrap(), F32X4);
+
+            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
+            let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
+
+            ctx.emit(Inst::VecRRNarrow {
+                op: VecRRNarrowOp::Fcvtn64,
+                rd,
+                rn,
+                high_half: false,
+            });
+        }
+
+        Opcode::ConstAddr | Opcode::Vconcat | Opcode::Vsplit => unimplemented!("lowering {}", op),
     }
 
     Ok(())