From bda1720eaf09ab859325bdbdeccc7bdac6683e0b Mon Sep 17 00:00:00 2001
From: Haoyu Zhang <haoyu_z@outlook.com>
Date: Wed, 1 Nov 2023 02:59:01 -0400
Subject: [PATCH] [v3i/simd]: Implement load_lane instructions

---
 src/engine/Value.v3            |  1 +
 src/engine/v3/V3Interpreter.v3 | 23 +++++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/src/engine/Value.v3 b/src/engine/Value.v3
index 00a3c136b..0b6ce4582 100644
--- a/src/engine/Value.v3
+++ b/src/engine/Value.v3
@@ -141,6 +141,7 @@ component Values {
 	def v_w_8(v: Value) -> u8	{ return u8.view(Value.I64.!(v).val); }
 	def v_w_16(v: Value) -> u16	{ return u16.view(Value.I64.!(v).val); }
 	def v_w_32(v: Value) -> u32	{ return u32.view(Value.I64.!(v).val); }
+	def v_w_64(v: Value) -> u64	{ return u64.view(Value.I64.!(v).val); }
 	def v_v128(v: Value) -> (u64, u64) {
 		var b = Value.V128.!(v);
 		return (b.low, b.high);
diff --git a/src/engine/v3/V3Interpreter.v3 b/src/engine/v3/V3Interpreter.v3
index 7b3cfd16a..5e89c1e22 100644
--- a/src/engine/v3/V3Interpreter.v3
+++ b/src/engine/v3/V3Interpreter.v3
@@ -914,6 +914,10 @@ component V3Interpreter {
 				if (t != TrapReason.NONE) trap(t);
 			}
 			V128_LOAD => doLoad(16, DataReaders.read_range_u128, Value.V128);
+			V128_LOAD_64_LANE => doLoadLane(8, DataReaders.read_range_u64, Value.I64, Values.v_w_64);
+			V128_LOAD_32_LANE => doLoadLane(4, DataReaders.read_range_u32, Value.I32, Values.v_u);
+			V128_LOAD_16_LANE => doLoadLane(2, DataReaders.read_range_u32_u16, Value.I32, Values.v_u_16);
+			V128_LOAD_8_LANE => doLoadLane(1, DataReaders.read_range_u32_u8, Value.I32, Values.v_u_8);
 			V128_STORE => doStore(16, DataWriters.write_range_u128, Values.v_v128);
 			V128_CONST => {
 				var low = codeptr.read_u64();
@@ -1048,6 +1052,25 @@ component V3Interpreter {
 		if (t.reason != TrapReason.NONE) trap(t.reason);
 		else push(box(read(t.result)));
 	}
+	def doLoadLane<T, R>(size: byte, read: Range<byte> -> T, box: T -> Value, unbox: Value -> R) {
+		var v = Values.v_v128(pop());
+		doLoad(size, read, box);
+		var val = u64.!(unbox(pop()));
+		var idx = codeptr.read1();
+		var low = v.0, high = v.1;
+		// Calculate number of lanes in low/high : (bytes(v128) / size) / 2
+		def half_lanes = (16 / size) >> 1;
+		if (idx < half_lanes) { // Update a lane in low
+			var shift = byte.view(idx * size * 8);
+			val <<= shift;
+			low |= val;
+		} else { // Update a lane in high
+			var shift = byte.view((idx - half_lanes) * size * 8);
+			val <<= shift;
+			high |= val;
+		}
+		push(Value.V128(low, high));
+	}
 	def doStore<T>(size: byte, write: (Range<byte>, T) -> void, unbox: Value -> T) {
 		var memarg = codeptr.read_MemArg();
 		var memory = frame.func.instance.memories[memarg.memory_index];