Optimize float-to-int Wasm trunc instructions

core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "wasmi_core"
-version = "0.3.0"
+version = "0.4.0"
 edition = "2021"
 authors = ["Parity Technologies <admin@parity.io>"]
 license = "MIT/Apache-2.0"
@@ -13,7 +13,6 @@ keywords = ["wasm", "webassembly", "bytecode", "interpreter"]
 [dependencies]
 memory_units = "0.4.0"
 libm = "0.2.1"
-num-rational = { version = "0.4", default-features = false, features = ["num-bigint"] }
 num-traits = { version = "0.2.8", default-features = false }
 downcast-rs = { version = "1.2", default-features = false }
 
@@ -24,8 +23,6 @@ rand = "0.8.2"
 default = ["std"]
 # Use `no-default-features` for a `no_std` build.
 std = [
-    "num-rational/std",
-    "num-rational/num-bigint-std",
     "num-traits/std",
     "downcast-rs/std",
 ]

core/src/value.rs

@@ -500,18 +500,17 @@ impl WrapInto<F32> for F64 {
 }
 
 macro_rules! impl_try_truncate_into {
-    (@primitive $from: ident, $into: ident, $to_primitive:path) => {
+    (@primitive $from: ident, $into: ident, $to_primitive:path, $rmin:literal, $rmax:literal) => {
         impl TryTruncateInto<$into, TrapCode> for $from {
             #[inline]
             fn try_truncate_into(self) -> Result<$into, TrapCode> {
-                // Casting from a float to an integer will round the float towards zero
                 if self.is_nan() {
                     return Err(TrapCode::InvalidConversionToInt);
                 }
-                num_rational::BigRational::from_float(self)
-                    .map(|val| val.to_integer())
-                    .and_then(|val| $to_primitive(&val))
-                    .ok_or(TrapCode::IntegerOverflow)
+                if self <= $rmin || self >= $rmax {
+                    return Err(TrapCode::IntegerOverflow);
+                }
+                Ok(self as _)
             }
         }
 
@@ -548,14 +547,14 @@ macro_rules! impl_try_truncate_into {
     };
 }
 
-impl_try_truncate_into!(@primitive f32, i32, num_traits::cast::ToPrimitive::to_i32);
-impl_try_truncate_into!(@primitive f32, i64, num_traits::cast::ToPrimitive::to_i64);
-impl_try_truncate_into!(@primitive f64, i32, num_traits::cast::ToPrimitive::to_i32);
-impl_try_truncate_into!(@primitive f64, i64, num_traits::cast::ToPrimitive::to_i64);
-impl_try_truncate_into!(@primitive f32, u32, num_traits::cast::ToPrimitive::to_u32);
-impl_try_truncate_into!(@primitive f32, u64, num_traits::cast::ToPrimitive::to_u64);
-impl_try_truncate_into!(@primitive f64, u32, num_traits::cast::ToPrimitive::to_u32);
-impl_try_truncate_into!(@primitive f64, u64, num_traits::cast::ToPrimitive::to_u64);
+impl_try_truncate_into!(@primitive f32, i32, num_traits::cast::ToPrimitive::to_i32, -2147483904.0_f32, 2147483648.0_f32);
+impl_try_truncate_into!(@primitive f32, u32, num_traits::cast::ToPrimitive::to_u32,          -1.0_f32, 4294967296.0_f32);
+impl_try_truncate_into!(@primitive f64, i32, num_traits::cast::ToPrimitive::to_i32, -2147483649.0_f64, 2147483648.0_f64);
+impl_try_truncate_into!(@primitive f64, u32, num_traits::cast::ToPrimitive::to_u32,          -1.0_f64, 4294967296.0_f64);
+impl_try_truncate_into!(@primitive f32, i64, num_traits::cast::ToPrimitive::to_i64, -9223373136366403584.0_f32,  9223372036854775808.0_f32);
+impl_try_truncate_into!(@primitive f32, u64, num_traits::cast::ToPrimitive::to_u64,                   -1.0_f32, 18446744073709551616.0_f32);
+impl_try_truncate_into!(@primitive f64, i64, num_traits::cast::ToPrimitive::to_i64, -9223372036854777856.0_f64,  9223372036854775808.0_f64);
+impl_try_truncate_into!(@primitive f64, u64, num_traits::cast::ToPrimitive::to_u64,                   -1.0_f64, 18446744073709551616.0_f64);
 impl_try_truncate_into!(@wrapped F32, f32, i32);
 impl_try_truncate_into!(@wrapped F32, f32, i64);
 impl_try_truncate_into!(@wrapped F64, f64, i32);

wasmi_v1/Cargo.toml

@@ -13,7 +13,7 @@ exclude = [ "tests/*", "benches/*" ]
 
 [dependencies]
 wasmparser = { version = "0.90", package = "wasmparser-nostd", default-features = false }
-wasmi_core = { version = "0.3", path = "../core", default-features = false }
+wasmi_core = { version = "0.4", path = "../core", default-features = false }
 spin = { version = "0.9", default-features = false, features = ["mutex", "spin_mutex"] }
 
 [dev-dependencies]

wasmi_v1/benches/benches.rs

@@ -11,6 +11,7 @@ use criterion::{criterion_group, criterion_main, Bencher, Criterion};
 use std::{slice, time::Duration};
 use wasmi as v1;
 use wasmi::core::Value;
+use wasmi_core::{F32, F64};
 
 const WASM_KERNEL: &str =
     "benches/wasm/wasm_kernel/target/wasm32-unknown-unknown/release/wasm_kernel.wasm";
@@ -46,6 +47,7 @@ criterion_group! {
         bench_execute_rev_comp_v1,
         bench_execute_regex_redux_v1,
         bench_execute_count_until_v1,
+        bench_execute_trunc_f2i,
         bench_execute_global_bump,
         bench_execute_fac_recursive_v1,
         bench_execute_fac_opt_v1,
@@ -254,6 +256,24 @@ fn bench_execute_count_until_v1(c: &mut Criterion) {
     });
 }
 
+fn bench_execute_trunc_f2i(c: &mut Criterion) {
+    const ITERATIONS: i32 = 25_000;
+    c.bench_function("execute/trunc_f2i/v1", |b| {
+        let (mut store, instance) = load_instance_from_wat_v1(include_bytes!("wat/trunc_f2i.wat"));
+        let count_until = instance
+            .get_export(&store, "trunc_f2i")
+            .and_then(v1::Extern::into_func)
+            .unwrap();
+        let count_until = count_until.typed::<(i32, F32, F64), (), _>(&store).unwrap();
+
+        b.iter(|| {
+            count_until
+                .call(&mut store, (ITERATIONS, F32::from(42.0), F64::from(69.0)))
+                .unwrap();
+        })
+    });
+}
+
 fn bench_execute_global_bump(c: &mut Criterion) {
     const BUMP_AMOUNT: i32 = 100_000;
     c.bench_function("execute/global_bump/v1", |b| {

wasmi_v1/benches/wat/trunc_f2i.wat

@@ -0,0 +1,45 @@
+(module
+    (func (export "trunc_f2i") (param $n i32) (param $input32 f32) (param $input64 f64) (result)
+        (local $i i32)
+        (block $exit
+            (if
+                (i32.le_u
+                    (local.get $n)
+                    (i32.const 0)
+                )
+                (unreachable) ;; trap if $n <= 0
+            )
+            (local.set $i (local.get $n)) ;; i = n
+            (loop $continue
+                (drop
+                    (i32.trunc_f32_s (local.get $input32)) ;; <- under test
+                )
+                (drop
+                    (i32.trunc_f32_u (local.get $input32)) ;; <- under test
+                )
+                (drop
+                    (i64.trunc_f32_s (local.get $input32)) ;; <- under test
+                )
+                (drop
+                    (i64.trunc_f64_u (local.get $input64)) ;; <- under test
+                )
+                (drop
+                    (i32.trunc_f64_s (local.get $input64)) ;; <- under test
+                )
+                (drop
+                    (i32.trunc_f64_u (local.get $input64)) ;; <- under test
+                )
+                (drop
+                    (i64.trunc_f64_s (local.get $input64)) ;; <- under test
+                )
+                (drop
+                    (i64.trunc_f64_u (local.get $input64)) ;; <- under test
+                )
+                (local.set $i ;; i -= 1
+                    (i32.sub (local.get $i) (i32.const 1))
+                )
+                (br_if $continue (local.get $i)) ;; continue if i != 0
+            )
+        )
+    )
+)