From cf408ddff5f91c321133d4fdc6dd0ef299f4d210 Mon Sep 17 00:00:00 2001
From: chen ruixiang <ruixiang.chen@partner.bmw.com>
Date: Thu, 25 Aug 2022 16:59:18 +0800
Subject: [PATCH 1/3] Fix: add missed tests for simd with type usize and isize

---
 tests/compiler/simd.debug.wat   | 976 +++++++++++++++++++++++++++++++-
 tests/compiler/simd.release.wat | 316 ++++++++++-
 tests/compiler/simd.ts          | 136 +++++
 3 files changed, 1395 insertions(+), 33 deletions(-)

diff --git a/tests/compiler/simd.debug.wat b/tests/compiler/simd.debug.wat
index 43014db8e4..0188d8c644 100644
--- a/tests/compiler/simd.debug.wat
+++ b/tests/compiler/simd.debug.wat
@@ -3,13 +3,14 @@
  (type $none_=>_none (func))
  (type $i32_=>_i32 (func (param i32) (result i32)))
  (type $i32_=>_none (func (param i32)))
- (type $i32_i32_i32_=>_none (func (param i32 i32 i32)))
  (type $i32_i32_=>_i32 (func (param i32 i32) (result i32)))
+ (type $i32_i32_i32_=>_none (func (param i32 i32 i32)))
  (type $i32_i32_i32_=>_v128 (func (param i32 i32 i32) (result v128)))
  (type $i32_i32_i32_=>_i32 (func (param i32 i32 i32) (result i32)))
  (type $i32_i32_i32_i32_=>_none (func (param i32 i32 i32 i32)))
  (type $none_=>_i32 (func (result i32)))
  (type $i32_i32_=>_v128 (func (param i32 i32) (result v128)))
+ (type $i32_i32_i32_i32_i32_=>_i32 (func (param i32 i32 i32 i32 i32) (result i32)))
  (type $none_=>_v128 (func (result v128)))
  (type $v128_=>_v128 (func (param v128) (result v128)))
  (type $i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_=>_v128 (func (param i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32) (result v128)))
@@ -45,10 +46,13 @@
  (global $~lib/builtins/i32.MAX_VALUE i32 (i32.const 2147483647))
  (global $~lib/builtins/u32.MAX_VALUE i32 (i32.const -1))
  (global $~lib/builtins/u64.MAX_VALUE i64 (i64.const -1))
- (global $~lib/rt/__rtti_base i32 (i32.const 560))
- (global $~lib/memory/__data_end i32 (i32.const 604))
- (global $~lib/memory/__stack_pointer (mut i32) (i32.const 16988))
- (global $~lib/memory/__heap_base i32 (i32.const 16988))
+ (global $~lib/process/process.arch i32 (i32.const 576))
+ (global $~lib/native/ASC_SHRINK_LEVEL i32 (i32.const 0))
+ (global $~lib/builtins/usize.MAX_VALUE i32 (i32.const -1))
+ (global $~lib/rt/__rtti_base i32 (i32.const 624))
+ (global $~lib/memory/__data_end i32 (i32.const 668))
+ (global $~lib/memory/__stack_pointer (mut i32) (i32.const 17052))
+ (global $~lib/memory/__heap_base i32 (i32.const 17052))
  (memory $0 1)
  (data (i32.const 12) ",\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\10\00\00\00\01\00\00\00\02\00\00\00\03\00\00\00\04\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00")
  (data (i32.const 60) "<\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00(\00\00\00A\00l\00l\00o\00c\00a\00t\00i\00o\00n\00 \00t\00o\00o\00 \00l\00a\00r\00g\00e\00\00\00\00\00")
@@ -61,7 +65,9 @@
  (data (i32.const 396) "<\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00\1e\00\00\00~\00l\00i\00b\00/\00r\00t\00/\00t\00l\00s\00f\00.\00t\00s\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00")
  (data (i32.const 460) ",\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s\00\00\00")
  (data (i32.const 508) ",\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00\0e\00\00\00s\00i\00m\00d\00.\00t\00s\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00")
- (data (i32.const 560) "\05\00\00\00 \00\00\00\00\00\00\00 \00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\02\04\00\00\00\00\00\00\02\t\00\00\00\00\00\00")
+ (data (i32.const 556) "\1c\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00\0c\00\00\00w\00a\00s\00m\003\002\00")
+ (data (i32.const 588) "\1c\00\00\00\00\00\00\00\00\00\00\00\01\00\00\00\0c\00\00\00w\00a\00s\00m\006\004\00")
+ (data (i32.const 624) "\05\00\00\00 \00\00\00\00\00\00\00 \00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\02\04\00\00\00\00\00\00\02\t\00\00\00\00\00\00")
  (table $0 1 1 funcref)
  (elem $0 (i32.const 1))
  (export "reexport" (func $simd/reexport))
@@ -6445,6 +6451,161 @@
   i32.ne
   drop
  )
+ (func $~lib/string/String#get:length (param $this i32) (result i32)
+  local.get $this
+  i32.const 20
+  i32.sub
+  i32.load offset=16
+  i32.const 1
+  i32.shr_u
+ )
+ (func $~lib/util/string/compareImpl (param $str1 i32) (param $index1 i32) (param $str2 i32) (param $index2 i32) (param $len i32) (result i32)
+  (local $ptr1 i32)
+  (local $ptr2 i32)
+  (local $var$7 i32)
+  (local $a i32)
+  (local $b i32)
+  local.get $str1
+  local.get $index1
+  i32.const 1
+  i32.shl
+  i32.add
+  local.set $ptr1
+  local.get $str2
+  local.get $index2
+  i32.const 1
+  i32.shl
+  i32.add
+  local.set $ptr2
+  i32.const 0
+  i32.const 2
+  i32.lt_s
+  drop
+  local.get $len
+  i32.const 4
+  i32.ge_u
+  if (result i32)
+   local.get $ptr1
+   i32.const 7
+   i32.and
+   local.get $ptr2
+   i32.const 7
+   i32.and
+   i32.or
+   i32.eqz
+  else
+   i32.const 0
+  end
+  if
+   block $do-break|0
+    loop $do-loop|0
+     local.get $ptr1
+     i64.load
+     local.get $ptr2
+     i64.load
+     i64.ne
+     if
+      br $do-break|0
+     end
+     local.get $ptr1
+     i32.const 8
+     i32.add
+     local.set $ptr1
+     local.get $ptr2
+     i32.const 8
+     i32.add
+     local.set $ptr2
+     local.get $len
+     i32.const 4
+     i32.sub
+     local.set $len
+     local.get $len
+     i32.const 4
+     i32.ge_u
+     br_if $do-loop|0
+    end
+   end
+  end
+  loop $while-continue|1
+   local.get $len
+   local.tee $var$7
+   i32.const 1
+   i32.sub
+   local.set $len
+   local.get $var$7
+   local.set $var$7
+   local.get $var$7
+   if
+    local.get $ptr1
+    i32.load16_u
+    local.set $a
+    local.get $ptr2
+    i32.load16_u
+    local.set $b
+    local.get $a
+    local.get $b
+    i32.ne
+    if
+     local.get $a
+     local.get $b
+     i32.sub
+     return
+    end
+    local.get $ptr1
+    i32.const 2
+    i32.add
+    local.set $ptr1
+    local.get $ptr2
+    i32.const 2
+    i32.add
+    local.set $ptr2
+    br $while-continue|1
+   end
+  end
+  i32.const 0
+ )
+ (func $~lib/string/String.__eq (param $left i32) (param $right i32) (result i32)
+  (local $leftLength i32)
+  local.get $left
+  local.get $right
+  i32.eq
+  if
+   i32.const 1
+   return
+  end
+  local.get $left
+  i32.const 0
+  i32.eq
+  if (result i32)
+   i32.const 1
+  else
+   local.get $right
+   i32.const 0
+   i32.eq
+  end
+  if
+   i32.const 0
+   return
+  end
+  local.get $left
+  call $~lib/string/String#get:length
+  local.set $leftLength
+  local.get $leftLength
+  local.get $right
+  call $~lib/string/String#get:length
+  i32.ne
+  if
+   i32.const 0
+   return
+  end
+  local.get $left
+  i32.const 0
+  local.get $right
+  i32.const 0
+  local.get $leftLength
+  call $~lib/util/string/compareImpl
+  i32.eqz
+ )
  (func $simd/test_const (result v128)
   (local $var$0 v128)
   v128.const i32x4 0x00000001 0x00000001 0x00000001 0x00000001
@@ -6522,6 +6683,8 @@
   call $simd/test_i64x2
   call $simd/test_f32x4
   call $simd/test_f64x2
+  call $simd/test_usize
+  call $simd/test_isize
   call $simd/test_const
   drop
   global.get $~lib/memory/__stack_pointer
@@ -6666,6 +6829,13 @@
  )
  (func $~lib/rt/__visit_globals (param $0 i32)
   (local $1 i32)
+  global.get $~lib/process/process.arch
+  local.tee $1
+  if
+   local.get $1
+   local.get $0
+   call $~lib/rt/itcms/__visit
+  end
   i32.const 272
   local.get $0
   call $~lib/rt/itcms/__visit
@@ -6752,14 +6922,804 @@
   global.get $~lib/memory/__data_end
   i32.lt_s
   if
-   i32.const 17008
-   i32.const 17056
+   i32.const 17072
+   i32.const 17120
    i32.const 1
    i32.const 1
    call $~lib/builtins/abort
    unreachable
   end
  )
+ (func $simd/test_usize
+  (local $0 i32)
+  (local $1 v128)
+  (local $2 i32)
+  global.get $~lib/memory/__stack_pointer
+  i32.const 8
+  i32.sub
+  global.set $~lib/memory/__stack_pointer
+  call $~stack_check
+  global.get $~lib/memory/__stack_pointer
+  i64.const 0
+  i64.store
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.mul
+  i32.const 42
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 2
+  i32x4.splat
+  i32x4.mul
+  i32.const 84
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 24
+  i32x4.splat
+  i32x4.add
+  i32.const 66
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32x4.extract_lane 0
+  i32.const 42
+  i32.eq
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 24
+  i32x4.replace_lane 0
+  i32x4.extract_lane 0
+  i32.const 24
+  i32.eq
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.add
+  i32.const 43
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.sub
+  i32.const 41
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.eq
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.eq
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.ne
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.ne
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  global.get $~lib/process/process.arch
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store
+  local.get $2
+  i32.const 576
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store offset=4
+  local.get $2
+  call $~lib/string/String.__eq
+  if
+   i32.const 1
+   i32x4.splat
+   i32.const 42
+   i32x4.splat
+   i32x4.le_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 1
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.le_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 1
+   i32x4.splat
+   i32.const 42
+   i32x4.splat
+   i32x4.lt_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 1
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.lt_u
+   i32.const 0
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 42
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.ge_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 1
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.ge_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 42
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.gt_u
+   i32.const -1
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+   i32.const 1
+   i32x4.splat
+   i32.const 1
+   i32x4.splat
+   i32x4.gt_u
+   i32.const 0
+   i32x4.splat
+   i8x16.eq
+   i8x16.all_true
+   i32.const 0
+   i32.ne
+   drop
+  end
+  i32.const 1
+  i32x4.splat
+  i32x4.neg
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  global.get $~lib/builtins/usize.MAX_VALUE
+  i32x4.splat
+  i32x4.neg
+  i32.const 1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.eq
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.shl
+  i32.const 2
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 2
+  i32x4.splat
+  i32.const 1
+  i32x4.shr_u
+  i32.const 1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32x4.all_true
+  i32.const 0
+  i32.ne
+  drop
+  global.get $~lib/process/process.arch
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store
+  local.get $2
+  i32.const 576
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store offset=4
+  local.get $2
+  call $~lib/string/String.__eq
+  if
+   global.get $~lib/builtins/usize.MAX_VALUE
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 15
+   i32.eq
+   drop
+   i32.const 0
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 0
+   i32.eq
+   drop
+  end
+  global.get $~lib/process/process.arch
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store
+  local.get $2
+  i32.const 608
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store offset=4
+  local.get $2
+  call $~lib/string/String.__eq
+  if
+   global.get $~lib/builtins/usize.MAX_VALUE
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 3
+   i32.eq
+   i32.eqz
+   if
+    i32.const 0
+    i32.const 528
+    i32.const 1137
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   i32.const 0
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 0
+   i32.eq
+   drop
+  end
+  i32.const 16
+  call $~lib/rt/tlsf/__alloc
+  local.set $0
+  local.get $0
+  i32.const 42
+  i32.store
+  i32.const 0
+  i32x4.splat
+  local.set $1
+  local.get $0
+  local.get $1
+  v128.load32_lane 0
+  local.set $1
+  local.get $1
+  i32x4.extract_lane 0
+  i32.const 42
+  i32.eq
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 528
+   i32.const 1153
+   i32.const 5
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  call $~lib/rt/tlsf/__free
+  global.get $~lib/memory/__stack_pointer
+  i32.const 8
+  i32.add
+  global.set $~lib/memory/__stack_pointer
+ )
+ (func $simd/test_isize
+  (local $0 i32)
+  (local $1 v128)
+  (local $2 i32)
+  global.get $~lib/memory/__stack_pointer
+  i32.const 8
+  i32.sub
+  global.set $~lib/memory/__stack_pointer
+  call $~stack_check
+  global.get $~lib/memory/__stack_pointer
+  i64.const 0
+  i64.store
+  i32.const 42
+  i32x4.splat
+  i32x4.extract_lane 0
+  i32.const 42
+  i32.eq
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 24
+  i32x4.replace_lane 0
+  i32x4.extract_lane 0
+  i32.const 24
+  i32.eq
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.add
+  i32.const 43
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.mul
+  i32.const 42
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.sub
+  i32.const 41
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.eq
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.eq
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.ne
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.ne
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.le_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.le_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.lt_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.lt_s
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.ge_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.ge_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.gt_s
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.splat
+  i32x4.gt_s
+  i32.const 0
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32.const 42
+  i32x4.splat
+  i32x4.eq
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32x4.neg
+  i32.const -1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const -1
+  i32x4.splat
+  i32x4.neg
+  i32.const 1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  global.get $~lib/builtins/usize.MAX_VALUE
+  i32x4.splat
+  i32x4.neg
+  i32.const 1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 42
+  i32x4.splat
+  i32x4.abs
+  i32.const 42
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32.const 1
+  i32x4.shl
+  i32.const 2
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 2
+  i32x4.splat
+  i32.const 1
+  i32x4.shr_s
+  i32.const 1
+  i32x4.splat
+  i8x16.eq
+  i8x16.all_true
+  i32.const 0
+  i32.ne
+  drop
+  i32.const 1
+  i32x4.splat
+  i32x4.all_true
+  i32.const 0
+  i32.ne
+  drop
+  global.get $~lib/process/process.arch
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store
+  local.get $2
+  i32.const 576
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store offset=4
+  local.get $2
+  call $~lib/string/String.__eq
+  if
+   global.get $~lib/builtins/usize.MAX_VALUE
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 15
+   i32.eq
+   drop
+   i32.const 0
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 0
+   i32.eq
+   drop
+  end
+  global.get $~lib/process/process.arch
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store
+  local.get $2
+  i32.const 608
+  local.set $2
+  global.get $~lib/memory/__stack_pointer
+  local.get $2
+  i32.store offset=4
+  local.get $2
+  call $~lib/string/String.__eq
+  if
+   global.get $~lib/builtins/usize.MAX_VALUE
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 3
+   i32.eq
+   i32.eqz
+   if
+    i32.const 0
+    i32.const 528
+    i32.const 1197
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   i32.const 0
+   i32x4.splat
+   i32x4.bitmask
+   i32.const 0
+   i32.eq
+   drop
+  end
+  i32.const 16
+  call $~lib/rt/tlsf/__alloc
+  local.set $0
+  local.get $0
+  i32.const 42
+  i32.store
+  i32.const 0
+  i32x4.splat
+  local.set $1
+  local.get $0
+  local.get $1
+  v128.load32_lane 0
+  local.set $1
+  local.get $1
+  i32x4.extract_lane 0
+  i32.const 42
+  i32.eq
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 528
+   i32.const 1213
+   i32.const 5
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  call $~lib/rt/tlsf/__free
+  global.get $~lib/memory/__stack_pointer
+  i32.const 8
+  i32.add
+  global.set $~lib/memory/__stack_pointer
+ )
  (func $~lib/rt/__newArray (param $0 i32) (param $1 i32) (param $2 i32) (param $3 i32) (result i32)
   (local $4 i32)
   (local $5 i32)
diff --git a/tests/compiler/simd.release.wat b/tests/compiler/simd.release.wat
index 9ce20cf5a2..277f80324a 100644
--- a/tests/compiler/simd.release.wat
+++ b/tests/compiler/simd.release.wat
@@ -4,10 +4,10 @@
  (type $i32_i32_i32_=>_v128 (func (param i32 i32 i32) (result v128)))
  (type $i32_i32_=>_none (func (param i32 i32)))
  (type $i32_i32_=>_i32 (func (param i32 i32) (result i32)))
+ (type $i32_=>_i32 (func (param i32) (result i32)))
  (type $i32_i32_i32_i32_=>_none (func (param i32 i32 i32 i32)))
  (type $i32_i32_i32_=>_none (func (param i32 i32 i32)))
  (type $none_=>_i32 (func (result i32)))
- (type $i32_=>_i32 (func (param i32) (result i32)))
  (type $v128_=>_v128 (func (param v128) (result v128)))
  (type $i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_=>_v128 (func (param i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32) (result v128)))
  (type $i32_i32_i32_i32_i32_i32_i32_i32_=>_v128 (func (param i32 i32 i32 i32 i32 i32 i32 i32) (result v128)))
@@ -30,7 +30,7 @@
  (global $~lib/rt/itcms/fromSpace (mut i32) (i32.const 0))
  (global $~lib/rt/tlsf/ROOT (mut i32) (i32.const 0))
  (global $simd/vec (mut v128) (v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000))
- (global $~lib/memory/__stack_pointer (mut i32) (i32.const 18012))
+ (global $~lib/memory/__stack_pointer (mut i32) (i32.const 18076))
  (memory $0 1)
  (data (i32.const 1036) ",")
  (data (i32.const 1052) "\10\00\00\00\01\00\00\00\02\00\00\00\03\00\00\00\04")
@@ -48,8 +48,12 @@
  (data (i32.const 1496) "\01\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s")
  (data (i32.const 1532) ",")
  (data (i32.const 1544) "\01\00\00\00\0e\00\00\00s\00i\00m\00d\00.\00t\00s")
- (data (i32.const 1584) "\05\00\00\00 \00\00\00\00\00\00\00 ")
- (data (i32.const 1612) "\02\04\00\00\00\00\00\00\02\t")
+ (data (i32.const 1580) "\1c")
+ (data (i32.const 1592) "\01\00\00\00\0c\00\00\00w\00a\00s\00m\003\002")
+ (data (i32.const 1612) "\1c")
+ (data (i32.const 1624) "\01\00\00\00\0c\00\00\00w\00a\00s\00m\006\004")
+ (data (i32.const 1648) "\05\00\00\00 \00\00\00\00\00\00\00 ")
+ (data (i32.const 1676) "\02\04\00\00\00\00\00\00\02\t")
  (export "reexport" (func $simd/reexport))
  (export "test_vars_i8x16_partial" (func $simd/test_vars_i8x16_partial))
  (export "test_vars_i8x16_full" (func $simd/test_vars_i8x16_full))
@@ -69,6 +73,8 @@
  (func $~lib/rt/itcms/visitRoots
   (local $0 i32)
   (local $1 i32)
+  i32.const 1600
+  call $byn-split-outlined-A$~lib/rt/itcms/__visit
   i32.const 1296
   call $byn-split-outlined-A$~lib/rt/itcms/__visit
   i32.const 1104
@@ -144,7 +150,7 @@
    if
     i32.const 0
     local.get $0
-    i32.const 18012
+    i32.const 18076
     i32.lt_u
     local.get $0
     i32.load offset=8
@@ -195,7 +201,7 @@
    i32.const 1
   else
    local.get $1
-   i32.const 1584
+   i32.const 1648
    i32.load
    i32.gt_u
    if
@@ -209,7 +215,7 @@
    local.get $1
    i32.const 3
    i32.shl
-   i32.const 1588
+   i32.const 1652
    i32.add
    i32.load
    i32.const 32
@@ -774,10 +780,10 @@
   if
    unreachable
   end
-  i32.const 18016
+  i32.const 18080
   i32.const 0
   i32.store
-  i32.const 19584
+  i32.const 19648
   i32.const 0
   i32.store
   loop $for-loop|0
@@ -788,7 +794,7 @@
     local.get $0
     i32.const 2
     i32.shl
-    i32.const 18016
+    i32.const 18080
     i32.add
     i32.const 0
     i32.store offset=4
@@ -806,7 +812,7 @@
       i32.add
       i32.const 2
       i32.shl
-      i32.const 18016
+      i32.const 18080
       i32.add
       i32.const 0
       i32.store offset=96
@@ -824,20 +830,20 @@
     br $for-loop|0
    end
   end
-  i32.const 18016
-  i32.const 19588
+  i32.const 18080
+  i32.const 19652
   memory.size
   i32.const 16
   i32.shl
   call $~lib/rt/tlsf/addMemory
-  i32.const 18016
+  i32.const 18080
   global.set $~lib/rt/tlsf/ROOT
  )
  (func $~lib/rt/tlsf/__free (param $0 i32)
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 18012
+  i32.const 18076
   i32.lt_u
   if
    return
@@ -967,7 +973,7 @@
      local.set $0
      loop $while-continue|0
       local.get $0
-      i32.const 18012
+      i32.const 18076
       i32.lt_u
       if
        local.get $0
@@ -1067,7 +1073,7 @@
      unreachable
     end
     local.get $0
-    i32.const 18012
+    i32.const 18076
     i32.lt_u
     if
      local.get $0
@@ -1845,6 +1851,117 @@
   local.get $0
   call $~lib/rt/tlsf/__free
  )
+ (func $~lib/string/String.__eq (param $0 i32) (result i32)
+  (local $1 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  local.get $0
+  i32.const 1600
+  i32.eq
+  if
+   i32.const 1
+   return
+  end
+  local.get $0
+  i32.eqz
+  if
+   i32.const 0
+   return
+  end
+  i32.const 1596
+  i32.load
+  i32.const 1
+  i32.shr_u
+  local.tee $2
+  local.get $0
+  i32.const 20
+  i32.sub
+  i32.load offset=16
+  i32.const 1
+  i32.shr_u
+  i32.ne
+  if
+   i32.const 0
+   return
+  end
+  i32.const 1600
+  local.set $3
+  local.get $0
+  local.tee $1
+  i32.const 7
+  i32.and
+  i32.eqz
+  local.get $2
+  local.tee $0
+  i32.const 4
+  i32.ge_u
+  i32.and
+  if
+   loop $do-loop|0
+    local.get $3
+    i64.load
+    local.get $1
+    i64.load
+    i64.eq
+    if
+     local.get $3
+     i32.const 8
+     i32.add
+     local.set $3
+     local.get $1
+     i32.const 8
+     i32.add
+     local.set $1
+     local.get $0
+     i32.const 4
+     i32.sub
+     local.tee $0
+     i32.const 4
+     i32.ge_u
+     br_if $do-loop|0
+    end
+   end
+  end
+  block $__inlined_func$~lib/util/string/compareImpl
+   loop $while-continue|1
+    local.get $0
+    local.tee $2
+    i32.const 1
+    i32.sub
+    local.set $0
+    local.get $2
+    if
+     local.get $3
+     i32.load16_u
+     local.tee $4
+     local.get $1
+     i32.load16_u
+     local.tee $5
+     i32.sub
+     local.set $2
+     local.get $4
+     local.get $5
+     i32.ne
+     br_if $__inlined_func$~lib/util/string/compareImpl
+     local.get $3
+     i32.const 2
+     i32.add
+     local.set $3
+     local.get $1
+     i32.const 2
+     i32.add
+     local.set $1
+     br $while-continue|1
+    end
+   end
+   i32.const 0
+   local.set $2
+  end
+  local.get $2
+  i32.eqz
+ )
  (func $start:simd
   (local $0 i32)
   (local $1 i32)
@@ -1855,11 +1972,11 @@
   i32.const 4
   i32.sub
   global.set $~lib/memory/__stack_pointer
-  block $folding-inner0
+  block $folding-inner1
    global.get $~lib/memory/__stack_pointer
-   i32.const 1628
+   i32.const 1692
    i32.lt_s
-   br_if $folding-inner0
+   br_if $folding-inner1
    global.get $~lib/memory/__stack_pointer
    local.tee $0
    i32.const 0
@@ -1867,7 +1984,7 @@
    memory.size
    i32.const 16
    i32.shl
-   i32.const 18012
+   i32.const 18076
    i32.sub
    i32.const 1
    i32.shr_u
@@ -1901,9 +2018,9 @@
    i32.sub
    global.set $~lib/memory/__stack_pointer
    global.get $~lib/memory/__stack_pointer
-   i32.const 1628
+   i32.const 1692
    i32.lt_s
-   br_if $folding-inner0
+   br_if $folding-inner1
    global.get $~lib/memory/__stack_pointer
    local.tee $1
    i32.const 0
@@ -2179,13 +2296,162 @@
    local.get $0
    call $~lib/rt/tlsf/__free
    global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.sub
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1692
+   i32.lt_s
+   br_if $folding-inner1
+   global.get $~lib/memory/__stack_pointer
+   local.tee $0
+   i64.const 0
+   i64.store
+   local.get $0
+   i32.const 1600
+   i32.store
+   local.get $0
+   i32.const 1600
+   i32.store offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1632
+   i32.store offset=4
+   i32.const 1632
+   call $~lib/string/String.__eq
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1137
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.eqz
+   if
+    call $~lib/rt/tlsf/initialize
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.const 16
+   call $~lib/rt/tlsf/allocateBlock
+   i32.const 4
+   i32.add
+   local.tee $0
+   i32.const 42
+   i32.store
+   local.get $0
+   v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000
+   v128.load32_lane 0
+   i32x4.extract_lane 0
+   i32.const 42
+   i32.ne
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1153
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   local.get $0
+   call $~lib/rt/tlsf/__free
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.add
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.sub
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1692
+   i32.lt_s
+   br_if $folding-inner1
+   global.get $~lib/memory/__stack_pointer
+   local.tee $0
+   i64.const 0
+   i64.store
+   local.get $0
+   i32.const 1600
+   i32.store
+   local.get $0
+   i32.const 1600
+   i32.store offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1632
+   i32.store offset=4
+   i32.const 1632
+   call $~lib/string/String.__eq
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1197
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.eqz
+   if
+    call $~lib/rt/tlsf/initialize
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.const 16
+   call $~lib/rt/tlsf/allocateBlock
+   i32.const 4
+   i32.add
+   local.tee $0
+   i32.const 42
+   i32.store
+   local.get $0
+   v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000
+   v128.load32_lane 0
+   i32x4.extract_lane 0
+   i32.const 42
+   i32.ne
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1213
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   local.get $0
+   call $~lib/rt/tlsf/__free
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.add
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
    i32.const 4
    i32.add
    global.set $~lib/memory/__stack_pointer
    return
   end
-  i32.const 18032
-  i32.const 18080
+  i32.const 18096
+  i32.const 18144
   i32.const 1
   i32.const 1
   call $~lib/builtins/abort
diff --git a/tests/compiler/simd.ts b/tests/compiler/simd.ts
index 2a68d5cb8b..2255149c42 100644
--- a/tests/compiler/simd.ts
+++ b/tests/compiler/simd.ts
@@ -1080,6 +1080,140 @@ function test_f64x2(): void {
     f64x2(1.0, 4.0)
   );
 }
+function test_usize(): void {
+  assert(
+    v128.mul<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) 
+    == 
+    v128.splat<usize>(42)
+  );
+  assert(
+    v128.mul<usize>(v128.splat<usize>(42), v128.splat<usize>(2)) 
+    == 
+    v128.splat<usize>(84)
+  );
+  assert(
+    v128.add<usize>(v128.splat<usize>(42), v128.splat<usize>(24))
+    == 
+    v128.splat<usize>(66)
+  );
+  assert(v128.extract_lane<usize>(v128.splat<usize>(42), 0) == 42);
+  assert(v128.extract_lane<usize>(v128.replace_lane<usize>(v128.splat<usize>(42), 0, 24), 0) == 24);
+  assert(v128.add<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<usize>(43));
+  assert(v128.sub<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<usize>(41));
+  assert(v128.eq<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<isize>(0));
+  assert(v128.eq<usize>(v128.splat<usize>(42), v128.splat<usize>(42)) == v128.splat<isize>(-1));
+  assert(v128.ne<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<isize>(-1));
+  assert(v128.ne<usize>(v128.splat<usize>(42), v128.splat<usize>(42)) == v128.splat<isize>(0));
+  if (process.arch == "wasm32") { // binary does not have opcode: LtU64x2, LeU64x2, GtU64x2, GeU64x2
+    assert(v128.le<usize>(v128.splat<usize>(1), v128.splat<usize>(42)) == v128.splat<isize>(-1));
+    assert(v128.le<usize>(v128.splat<usize>(1), v128.splat<usize>(1)) == v128.splat<isize>(-1));
+    assert(v128.lt<usize>(v128.splat<usize>(1), v128.splat<usize>(42)) == v128.splat<isize>(-1));
+    assert(v128.lt<usize>(v128.splat<usize>(1), v128.splat<usize>(1)) == v128.splat<isize>(0));
+    assert(v128.ge<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<isize>(-1));
+    assert(v128.ge<usize>(v128.splat<usize>(1), v128.splat<usize>(1)) == v128.splat<isize>(-1));
+    assert(v128.gt<usize>(v128.splat<usize>(42), v128.splat<usize>(1)) == v128.splat<isize>(-1));
+    assert(v128.gt<usize>(v128.splat<usize>(1), v128.splat<usize>(1)) == v128.splat<isize>(0));
+  }
+  assert(v128.neg<usize>(v128.splat<usize>(1)) == v128.splat<isize>(-1));
+  assert(v128.neg<usize>(v128.splat<usize>(usize.MAX_VALUE)) == v128.splat<isize>(1));
+  assert(v128.eq<usize>(v128.splat<usize>(42), v128.splat<usize>(42)) == v128.splat<isize>(-1));
+  assert(v128.abs<usize>(v128.splat<usize>(42)) == v128.splat<usize>(42));
+  assert(v128.shl<usize>(v128.splat<usize>(1), 1) == v128.splat<usize>(2));
+  assert(v128.shr<usize>(v128.splat<usize>(2), 1) == v128.splat<usize>(1));
+  assert(v128.all_true<usize>(v128.splat<usize>(1)));
+  if (process.arch == "wasm32") { //bitmask will return different value depends on compile target
+    assert(
+      v128.bitmask<usize>(v128.splat<usize>(usize.MAX_VALUE))
+      ==
+      0xf
+    );
+    assert(
+      v128.bitmask<usize>(v128.splat<usize>(0))
+      ==
+      0
+    );
+  }
+  if (process.arch == "wasm64") {
+    assert(
+      v128.bitmask<usize>(v128.splat<usize>(usize.MAX_VALUE))
+      ==
+      0x3
+    );
+    assert(
+      v128.bitmask<usize>(v128.splat<usize>(0))
+      ==
+      0
+    );
+  }
+  {
+    let ptr = __alloc(16);
+    store<usize>(ptr, 42);
+    let v: v128 = v128.splat<usize>(0);
+    v = v128.load_lane<usize>(ptr, v, 0);
+    assert(v128.extract_lane<usize>(v, 0) == 42);
+    __free(ptr);
+  }
+}
+
+function test_isize(): void {
+  assert(v128.extract_lane<isize>(v128.splat<usize>(42), 0) == 42);
+  assert(v128.extract_lane<isize>(v128.replace_lane<usize>(v128.splat<usize>(42), 0, 24), 0) == 24);
+  assert(v128.add<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(43));
+  assert(v128.mul<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(42));
+  assert(v128.sub<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(41));
+  assert(v128.eq<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(0));
+  assert(v128.eq<isize>(v128.splat<isize>(42), v128.splat<isize>(42)) == v128.splat<isize>(-1));
+  assert(v128.ne<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.ne<isize>(v128.splat<isize>(42), v128.splat<isize>(42)) == v128.splat<isize>(0));
+  assert(v128.le<isize>(v128.splat<isize>(1), v128.splat<isize>(42)) == v128.splat<isize>(-1));
+  assert(v128.le<isize>(v128.splat<isize>(1), v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.lt<isize>(v128.splat<isize>(1), v128.splat<isize>(42)) == v128.splat<isize>(-1));
+  assert(v128.lt<isize>(v128.splat<isize>(1), v128.splat<isize>(1)) == v128.splat<isize>(0));
+  assert(v128.ge<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.ge<isize>(v128.splat<isize>(1), v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.gt<isize>(v128.splat<isize>(42), v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.gt<isize>(v128.splat<isize>(1), v128.splat<isize>(1)) == v128.splat<isize>(0));
+  assert(v128.eq<isize>(v128.splat<isize>(42), v128.splat<isize>(42)) == v128.splat<isize>(-1));
+  assert(v128.neg<isize>(v128.splat<isize>(1)) == v128.splat<isize>(-1));
+  assert(v128.neg<isize>(v128.splat<isize>(-1)) == v128.splat<isize>(1));
+  assert(v128.neg<isize>(v128.splat<isize>(usize.MAX_VALUE)) == v128.splat<isize>(1));
+  assert(v128.abs<isize>(v128.splat<isize>(42)) == v128.splat<isize>(42));
+  assert(v128.shl<isize>(v128.splat<isize>(1), 1) == v128.splat<isize>(2));
+  assert(v128.shr<isize>(v128.splat<isize>(2), 1) == v128.splat<isize>(1));
+  assert(v128.all_true<isize>(v128.splat<isize>(1)));
+  if (process.arch == "wasm32") { //bitmask will return different value depends on compile target
+    assert(
+      v128.bitmask<isize>(v128.splat<isize>(usize.MAX_VALUE))
+      ==
+      0xf
+    );
+    assert(
+      v128.bitmask<isize>(v128.splat<isize>(0))
+      ==
+      0
+    );
+  }
+  if (process.arch == "wasm64") {
+    assert(
+      v128.bitmask<isize>(v128.splat<isize>(usize.MAX_VALUE))
+      ==
+      0x3
+    );
+    assert(
+      v128.bitmask<isize>(v128.splat<isize>(0))
+      ==
+      0
+    );
+  }
+  {
+    let ptr = __alloc(16);
+    store<isize>(ptr, 42);
+    let v: v128 = v128.splat<isize>(0);
+    v = v128.load_lane<isize>(ptr, v, 0);
+    assert(v128.extract_lane<isize>(v, 0) == 42);
+    __free(ptr);
+  }
+}
 
 function test_const(): v128 {
   const one = i32x4.splat(1); // should precompute
@@ -1163,4 +1297,6 @@ test_i32x4();
 test_i64x2();
 test_f32x4();
 test_f64x2();
+test_usize();
+test_isize();
 test_const();

From e05ed663caa474aff7a255cd98b7da8d2b04f9be Mon Sep 17 00:00:00 2001
From: chen ruixiang <ruixiang.chen@partner.bmw.com>
Date: Thu, 25 Aug 2022 17:43:52 +0800
Subject: [PATCH 2/3] fix format issue

---
 tests/compiler/simd.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tests/compiler/simd.ts b/tests/compiler/simd.ts
index 2255149c42..81ebfe79b4 100644
--- a/tests/compiler/simd.ts
+++ b/tests/compiler/simd.ts
@@ -1121,7 +1121,7 @@ function test_usize(): void {
   assert(v128.shl<usize>(v128.splat<usize>(1), 1) == v128.splat<usize>(2));
   assert(v128.shr<usize>(v128.splat<usize>(2), 1) == v128.splat<usize>(1));
   assert(v128.all_true<usize>(v128.splat<usize>(1)));
-  if (process.arch == "wasm32") { //bitmask will return different value depends on compile target
+  if (process.arch == "wasm32") { // bitmask will return different value depends on compile target
     assert(
       v128.bitmask<usize>(v128.splat<usize>(usize.MAX_VALUE))
       ==
@@ -1181,7 +1181,7 @@ function test_isize(): void {
   assert(v128.shl<isize>(v128.splat<isize>(1), 1) == v128.splat<isize>(2));
   assert(v128.shr<isize>(v128.splat<isize>(2), 1) == v128.splat<isize>(1));
   assert(v128.all_true<isize>(v128.splat<isize>(1)));
-  if (process.arch == "wasm32") { //bitmask will return different value depends on compile target
+  if (process.arch == "wasm32") { // bitmask will return different value depends on compile target
     assert(
       v128.bitmask<isize>(v128.splat<isize>(usize.MAX_VALUE))
       ==

From 11a05957fc5f356533eb19521642443fe78b2a47 Mon Sep 17 00:00:00 2001
From: chen ruixiang <ruixiang.chen@partner.bmw.com>
Date: Thu, 1 Sep 2022 17:06:31 +0800
Subject: [PATCH 3/3] fix conflicts

---
 tests/compiler/simd.debug.wat   |  42 ++---
 tests/compiler/simd.release.wat | 316 +++++++++++++++++++++++++++++---
 2 files changed, 312 insertions(+), 46 deletions(-)

diff --git a/tests/compiler/simd.debug.wat b/tests/compiler/simd.debug.wat
index a7d52d5664..d5dcf270dc 100644
--- a/tests/compiler/simd.debug.wat
+++ b/tests/compiler/simd.debug.wat
@@ -6455,7 +6455,7 @@
   local.get $this
   i32.const 20
   i32.sub
-  i32.load offset=16
+  i32.load $0 offset=16
   i32.const 1
   i32.shr_u
  )
@@ -6500,9 +6500,9 @@
    block $do-break|0
     loop $do-loop|0
      local.get $ptr1
-     i64.load
+     i64.load $0
      local.get $ptr2
-     i64.load
+     i64.load $0
      i64.ne
      if
       br $do-break|0
@@ -6537,10 +6537,10 @@
    local.get $var$7
    if
     local.get $ptr1
-    i32.load16_u
+    i32.load16_u $0
     local.set $a
     local.get $ptr2
-    i32.load16_u
+    i32.load16_u $0
     local.set $b
     local.get $a
     local.get $b
@@ -6941,7 +6941,7 @@
   call $~stack_check
   global.get $~lib/memory/__stack_pointer
   i64.const 0
-  i64.store
+  i64.store $0
   i32.const 42
   i32x4.splat
   i32.const 1
@@ -7068,13 +7068,13 @@
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store
+  i32.store $0
   local.get $2
   i32.const 576
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store offset=4
+  i32.store $0 offset=4
   local.get $2
   call $~lib/string/String.__eq
   if
@@ -7248,13 +7248,13 @@
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store
+  i32.store $0
   local.get $2
   i32.const 576
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store offset=4
+  i32.store $0 offset=4
   local.get $2
   call $~lib/string/String.__eq
   if
@@ -7275,13 +7275,13 @@
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store
+  i32.store $0
   local.get $2
   i32.const 608
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store offset=4
+  i32.store $0 offset=4
   local.get $2
   call $~lib/string/String.__eq
   if
@@ -7311,13 +7311,13 @@
   local.set $0
   local.get $0
   i32.const 42
-  i32.store
+  i32.store $0
   i32.const 0
   i32x4.splat
   local.set $1
   local.get $0
   local.get $1
-  v128.load32_lane 0
+  v128.load32_lane $0 0
   local.set $1
   local.get $1
   i32x4.extract_lane 0
@@ -7350,7 +7350,7 @@
   call $~stack_check
   global.get $~lib/memory/__stack_pointer
   i64.const 0
-  i64.store
+  i64.store $0
   i32.const 42
   i32x4.splat
   i32x4.extract_lane 0
@@ -7629,13 +7629,13 @@
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store
+  i32.store $0
   local.get $2
   i32.const 576
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store offset=4
+  i32.store $0 offset=4
   local.get $2
   call $~lib/string/String.__eq
   if
@@ -7656,13 +7656,13 @@
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store
+  i32.store $0
   local.get $2
   i32.const 608
   local.set $2
   global.get $~lib/memory/__stack_pointer
   local.get $2
-  i32.store offset=4
+  i32.store $0 offset=4
   local.get $2
   call $~lib/string/String.__eq
   if
@@ -7692,13 +7692,13 @@
   local.set $0
   local.get $0
   i32.const 42
-  i32.store
+  i32.store $0
   i32.const 0
   i32x4.splat
   local.set $1
   local.get $0
   local.get $1
-  v128.load32_lane 0
+  v128.load32_lane $0 0
   local.set $1
   local.get $1
   i32x4.extract_lane 0
diff --git a/tests/compiler/simd.release.wat b/tests/compiler/simd.release.wat
index f81a93bcf5..85f27f8929 100644
--- a/tests/compiler/simd.release.wat
+++ b/tests/compiler/simd.release.wat
@@ -4,10 +4,10 @@
  (type $i32_i32_i32_=>_v128 (func (param i32 i32 i32) (result v128)))
  (type $i32_i32_=>_none (func (param i32 i32)))
  (type $i32_i32_=>_i32 (func (param i32 i32) (result i32)))
+ (type $i32_=>_i32 (func (param i32) (result i32)))
  (type $i32_i32_i32_i32_=>_none (func (param i32 i32 i32 i32)))
  (type $i32_i32_i32_=>_none (func (param i32 i32 i32)))
  (type $none_=>_i32 (func (result i32)))
- (type $i32_=>_i32 (func (param i32) (result i32)))
  (type $v128_=>_v128 (func (param v128) (result v128)))
  (type $i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_i32_=>_v128 (func (param i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32 i32) (result v128)))
  (type $i32_i32_i32_i32_i32_i32_i32_i32_=>_v128 (func (param i32 i32 i32 i32 i32 i32 i32 i32) (result v128)))
@@ -30,7 +30,7 @@
  (global $~lib/rt/itcms/fromSpace (mut i32) (i32.const 0))
  (global $~lib/rt/tlsf/ROOT (mut i32) (i32.const 0))
  (global $simd/vec (mut v128) (v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000))
- (global $~lib/memory/__stack_pointer (mut i32) (i32.const 18012))
+ (global $~lib/memory/__stack_pointer (mut i32) (i32.const 18076))
  (memory $0 1)
  (data (i32.const 1036) ",")
  (data (i32.const 1052) "\10\00\00\00\01\00\00\00\02\00\00\00\03\00\00\00\04")
@@ -48,8 +48,12 @@
  (data (i32.const 1496) "\01\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s")
  (data (i32.const 1532) ",")
  (data (i32.const 1544) "\01\00\00\00\0e\00\00\00s\00i\00m\00d\00.\00t\00s")
- (data (i32.const 1584) "\05\00\00\00 \00\00\00\00\00\00\00 ")
- (data (i32.const 1612) "\02\04\00\00\00\00\00\00\02\t")
+ (data (i32.const 1580) "\1c")
+ (data (i32.const 1592) "\01\00\00\00\0c\00\00\00w\00a\00s\00m\003\002")
+ (data (i32.const 1612) "\1c")
+ (data (i32.const 1624) "\01\00\00\00\0c\00\00\00w\00a\00s\00m\006\004")
+ (data (i32.const 1648) "\05\00\00\00 \00\00\00\00\00\00\00 ")
+ (data (i32.const 1676) "\02\04\00\00\00\00\00\00\02\t")
  (export "reexport" (func $simd/reexport))
  (export "test_vars_i8x16_partial" (func $simd/test_vars_i8x16_partial))
  (export "test_vars_i8x16_full" (func $simd/test_vars_i8x16_full))
@@ -69,6 +73,8 @@
  (func $~lib/rt/itcms/visitRoots
   (local $0 i32)
   (local $1 i32)
+  i32.const 1600
+  call $byn-split-outlined-A$~lib/rt/itcms/__visit
   i32.const 1296
   call $byn-split-outlined-A$~lib/rt/itcms/__visit
   i32.const 1104
@@ -146,7 +152,7 @@
     i32.load $0 offset=8
     i32.eqz
     local.get $0
-    i32.const 18012
+    i32.const 18076
     i32.lt_u
     i32.and
     i32.eqz
@@ -195,7 +201,7 @@
    i32.const 1
   else
    local.get $1
-   i32.const 1584
+   i32.const 1648
    i32.load $0
    i32.gt_u
    if
@@ -209,7 +215,7 @@
    local.get $1
    i32.const 3
    i32.shl
-   i32.const 1588
+   i32.const 1652
    i32.add
    i32.load $0
    i32.const 32
@@ -774,10 +780,10 @@
   if
    unreachable
   end
-  i32.const 18016
+  i32.const 18080
   i32.const 0
   i32.store $0
-  i32.const 19584
+  i32.const 19648
   i32.const 0
   i32.store $0
   loop $for-loop|0
@@ -788,7 +794,7 @@
     local.get $0
     i32.const 2
     i32.shl
-    i32.const 18016
+    i32.const 18080
     i32.add
     i32.const 0
     i32.store $0 offset=4
@@ -806,7 +812,7 @@
       i32.add
       i32.const 2
       i32.shl
-      i32.const 18016
+      i32.const 18080
       i32.add
       i32.const 0
       i32.store $0 offset=96
@@ -824,20 +830,20 @@
     br $for-loop|0
    end
   end
-  i32.const 18016
-  i32.const 19588
+  i32.const 18080
+  i32.const 19652
   memory.size $0
   i32.const 16
   i32.shl
   call $~lib/rt/tlsf/addMemory
-  i32.const 18016
+  i32.const 18080
   global.set $~lib/rt/tlsf/ROOT
  )
  (func $~lib/rt/tlsf/__free (param $0 i32)
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 18012
+  i32.const 18076
   i32.lt_u
   if
    return
@@ -967,7 +973,7 @@
      local.set $0
      loop $while-continue|0
       local.get $0
-      i32.const 18012
+      i32.const 18076
       i32.lt_u
       if
        local.get $0
@@ -1067,7 +1073,7 @@
      unreachable
     end
     local.get $0
-    i32.const 18012
+    i32.const 18076
     i32.lt_u
     if
      local.get $0
@@ -1845,6 +1851,117 @@
   local.get $0
   call $~lib/rt/tlsf/__free
  )
+ (func $~lib/string/String.__eq (param $0 i32) (result i32)
+  (local $1 i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  local.get $0
+  i32.const 1600
+  i32.eq
+  if
+   i32.const 1
+   return
+  end
+  local.get $0
+  i32.eqz
+  if
+   i32.const 0
+   return
+  end
+  i32.const 1596
+  i32.load $0
+  i32.const 1
+  i32.shr_u
+  local.tee $2
+  local.get $0
+  i32.const 20
+  i32.sub
+  i32.load $0 offset=16
+  i32.const 1
+  i32.shr_u
+  i32.ne
+  if
+   i32.const 0
+   return
+  end
+  i32.const 1600
+  local.set $3
+  local.get $0
+  local.tee $1
+  i32.const 7
+  i32.and
+  i32.eqz
+  local.get $2
+  local.tee $0
+  i32.const 4
+  i32.ge_u
+  i32.and
+  if
+   loop $do-loop|0
+    local.get $3
+    i64.load $0
+    local.get $1
+    i64.load $0
+    i64.eq
+    if
+     local.get $3
+     i32.const 8
+     i32.add
+     local.set $3
+     local.get $1
+     i32.const 8
+     i32.add
+     local.set $1
+     local.get $0
+     i32.const 4
+     i32.sub
+     local.tee $0
+     i32.const 4
+     i32.ge_u
+     br_if $do-loop|0
+    end
+   end
+  end
+  block $__inlined_func$~lib/util/string/compareImpl
+   loop $while-continue|1
+    local.get $0
+    local.tee $2
+    i32.const 1
+    i32.sub
+    local.set $0
+    local.get $2
+    if
+     local.get $3
+     i32.load16_u $0
+     local.tee $4
+     local.get $1
+     i32.load16_u $0
+     local.tee $5
+     i32.sub
+     local.set $2
+     local.get $4
+     local.get $5
+     i32.ne
+     br_if $__inlined_func$~lib/util/string/compareImpl
+     local.get $3
+     i32.const 2
+     i32.add
+     local.set $3
+     local.get $1
+     i32.const 2
+     i32.add
+     local.set $1
+     br $while-continue|1
+    end
+   end
+   i32.const 0
+   local.set $2
+  end
+  local.get $2
+  i32.eqz
+ )
  (func $start:simd
   (local $0 i32)
   (local $1 i32)
@@ -1855,11 +1972,11 @@
   i32.const 4
   i32.sub
   global.set $~lib/memory/__stack_pointer
-  block $folding-inner0
+  block $folding-inner1
    global.get $~lib/memory/__stack_pointer
-   i32.const 1628
+   i32.const 1692
    i32.lt_s
-   br_if $folding-inner0
+   br_if $folding-inner1
    global.get $~lib/memory/__stack_pointer
    local.tee $0
    i32.const 0
@@ -1867,7 +1984,7 @@
    memory.size $0
    i32.const 16
    i32.shl
-   i32.const 18012
+   i32.const 18076
    i32.sub
    i32.const 1
    i32.shr_u
@@ -1901,9 +2018,9 @@
    i32.sub
    global.set $~lib/memory/__stack_pointer
    global.get $~lib/memory/__stack_pointer
-   i32.const 1628
+   i32.const 1692
    i32.lt_s
-   br_if $folding-inner0
+   br_if $folding-inner1
    global.get $~lib/memory/__stack_pointer
    local.tee $1
    i32.const 0
@@ -2179,13 +2296,162 @@
    local.get $0
    call $~lib/rt/tlsf/__free
    global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.sub
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1692
+   i32.lt_s
+   br_if $folding-inner1
+   global.get $~lib/memory/__stack_pointer
+   local.tee $0
+   i64.const 0
+   i64.store $0
+   local.get $0
+   i32.const 1600
+   i32.store $0
+   local.get $0
+   i32.const 1600
+   i32.store $0 offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store $0
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store $0 offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store $0
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1632
+   i32.store $0 offset=4
+   i32.const 1632
+   call $~lib/string/String.__eq
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1137
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.eqz
+   if
+    call $~lib/rt/tlsf/initialize
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.const 16
+   call $~lib/rt/tlsf/allocateBlock
+   i32.const 4
+   i32.add
+   local.tee $0
+   i32.const 42
+   i32.store $0
+   local.get $0
+   v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000
+   v128.load32_lane $0 0
+   i32x4.extract_lane 0
+   i32.const 42
+   i32.ne
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1153
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   local.get $0
+   call $~lib/rt/tlsf/__free
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.add
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.sub
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1692
+   i32.lt_s
+   br_if $folding-inner1
+   global.get $~lib/memory/__stack_pointer
+   local.tee $0
+   i64.const 0
+   i64.store $0
+   local.get $0
+   i32.const 1600
+   i32.store $0
+   local.get $0
+   i32.const 1600
+   i32.store $0 offset=4
+   i32.const 1600
+   call $~lib/string/String.__eq
+   drop
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1600
+   i32.store $0
+   global.get $~lib/memory/__stack_pointer
+   i32.const 1632
+   i32.store $0 offset=4
+   i32.const 1632
+   call $~lib/string/String.__eq
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1197
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.eqz
+   if
+    call $~lib/rt/tlsf/initialize
+   end
+   global.get $~lib/rt/tlsf/ROOT
+   i32.const 16
+   call $~lib/rt/tlsf/allocateBlock
+   i32.const 4
+   i32.add
+   local.tee $0
+   i32.const 42
+   i32.store $0
+   local.get $0
+   v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000
+   v128.load32_lane $0 0
+   i32x4.extract_lane 0
+   i32.const 42
+   i32.ne
+   if
+    i32.const 0
+    i32.const 1552
+    i32.const 1213
+    i32.const 5
+    call $~lib/builtins/abort
+    unreachable
+   end
+   local.get $0
+   call $~lib/rt/tlsf/__free
+   global.get $~lib/memory/__stack_pointer
+   i32.const 8
+   i32.add
+   global.set $~lib/memory/__stack_pointer
+   global.get $~lib/memory/__stack_pointer
    i32.const 4
    i32.add
    global.set $~lib/memory/__stack_pointer
    return
   end
-  i32.const 18032
-  i32.const 18080
+  i32.const 18096
+  i32.const 18144
   i32.const 1
   i32.const 1
   call $~lib/builtins/abort