From b1680f7b2053fa2a1dcbdd2b1445390a2f534d49 Mon Sep 17 00:00:00 2001
From: MaxGraey <maxgraey@gmail.com>
Date: Mon, 26 Oct 2020 18:19:48 +0200
Subject: [PATCH 1/4] slightly optimize Math.mod

---
 std/assembly/math.ts                        |  16 +-
 tests/compiler/binary.untouched.wat         |  22 +-
 tests/compiler/std/math.optimized.wat       | 480 ++++++++++----------
 tests/compiler/std/math.untouched.wat       |  22 +-
 tests/compiler/std/mod.optimized.wat        | 480 ++++++++++----------
 tests/compiler/std/mod.untouched.wat        |  22 +-
 tests/compiler/std/typedarray.optimized.wat | 394 ++++++++--------
 tests/compiler/std/typedarray.untouched.wat |  22 +-
 8 files changed, 733 insertions(+), 725 deletions(-)

diff --git a/std/assembly/math.ts b/std/assembly/math.ts
index ffc20132b8..4b37306fe9 100644
--- a/std/assembly/math.ts
+++ b/std/assembly/math.ts
@@ -1575,7 +1575,7 @@ export namespace NativeMath {
     var uy = reinterpret<u64>(y);
     var ex = <i64>(ux >> 52 & 0x7FF);
     var ey = <i64>(uy >> 52 & 0x7FF);
-    var sx = ux >> 63;
+    var sm = ux & 0x8000000000000000;
     var uy1 = uy << 1;
     if (uy1 == 0 || ex == 0x7FF || isNaN<f64>(y)) {
       let m = x * y;
@@ -1583,8 +1583,7 @@ export namespace NativeMath {
     }
     var ux1 = ux << 1;
     if (ux1 <= uy1) {
-      if (ux1 == uy1) return 0 * x;
-      return x;
+      return ux1 == uy1 ? 0 * x : x;
     }
     if (!ex) {
       ex -= builtin_clz<i64>(ux << 12);
@@ -1622,8 +1621,7 @@ export namespace NativeMath {
     } else {
       ux >>= -ex + 1;
     }
-    ux |= sx << 63;
-    return reinterpret<f64>(ux);
+    return reinterpret<f64>(ux | sm);
   }
 
   export function rem(x: f64, y: f64): f64 { // see: musl/src/math/remquo.c
@@ -2884,7 +2882,7 @@ export namespace NativeMathf {
     var uy = reinterpret<u32>(y);
     var ex = <i32>(ux >> 23 & 0xFF);
     var ey = <i32>(uy >> 23 & 0xFF);
-    var sx = ux & 0x80000000;
+    var sm = ux & 0x80000000;
     var uy1 = uy << 1;
     if (uy1 == 0 || ex == 0xFF || isNaN<f32>(y)) {
       let m = x * y;
@@ -2892,8 +2890,7 @@ export namespace NativeMathf {
     }
     var ux1 = ux << 1;
     if (ux1 <= uy1) {
-      if (ux1 == uy1) return 0 * x;
-      return x;
+      return ux1 == uy1 ? 0 * x : x;
     }
     if (!ex) {
       ex -= builtin_clz<u32>(ux << 9);
@@ -2931,8 +2928,7 @@ export namespace NativeMathf {
     } else {
       ux >>= -ex + 1;
     }
-    ux |= sx;
-    return reinterpret<f32>(ux);
+    return reinterpret<f32>(ux | sm);
   }
 
   export function rem(x: f32, y: f32): f32 { // see: musl/src/math/remquof.c
diff --git a/tests/compiler/binary.untouched.wat b/tests/compiler/binary.untouched.wat
index a7a522d0ef..7a32ac7e7d 100644
--- a/tests/compiler/binary.untouched.wat
+++ b/tests/compiler/binary.untouched.wat
@@ -1291,13 +1291,13 @@
    local.get $9
    local.get $7
    i32.eq
-   if
+   if (result f32)
     f32.const 0
     local.get $0
     f32.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -1460,8 +1460,6 @@
   local.get $2
   local.get $6
   i32.or
-  local.set $2
-  local.get $2
   f32.reinterpret_i32
  )
  (func $~lib/math/NativeMathf.pow (param $0 f32) (param $1 f32) (result f32)
@@ -2121,8 +2119,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $6
   local.get $3
   i64.const 1
@@ -2166,13 +2164,13 @@
    local.get $9
    local.get $7
    i64.eq
-   if
+   if (result f64)
     f64.const 0
     local.get $0
     f64.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -2338,11 +2336,7 @@
   end
   local.get $2
   local.get $6
-  i64.const 63
-  i64.shl
   i64.or
-  local.set $2
-  local.get $2
   f64.reinterpret_i64
  )
  (func $start:binary
diff --git a/tests/compiler/std/math.optimized.wat b/tests/compiler/std/math.optimized.wat
index 0be7bd82f2..8449ad1b56 100644
--- a/tests/compiler/std/math.optimized.wat
+++ b/tests/compiler/std/math.optimized.wat
@@ -6487,8 +6487,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $7
   i32.const 1
   local.get $1
@@ -6519,146 +6519,151 @@
    f64.div
    return
   end
-  block $folding-inner0
+  local.get $2
+  i64.const 1
+  i64.shl
+  local.tee $8
+  local.get $6
+  i64.le_u
+  if
+   local.get $0
+   f64.const 0
+   f64.mul
+   local.get $0
+   local.get $6
+   local.get $8
+   i64.eq
+   select
+   return
+  end
+  local.get $4
+  i64.eqz
+  if (result i64)
    local.get $2
    i64.const 1
+   local.get $4
+   local.get $2
+   i64.const 12
    i64.shl
-   local.tee $8
-   local.get $6
-   i64.le_u
-   if
-    local.get $6
-    local.get $8
-    i64.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i64.clz
+   i64.sub
+   local.tee $4
+   i64.sub
+   i64.shl
+  else
+   local.get $2
+   i64.const 4503599627370495
+   i64.and
+   i64.const 4503599627370496
+   i64.or
+  end
+  local.set $2
+  local.get $5
+  i64.eqz
+  if (result i64)
+   local.get $3
+   i64.const 1
+   local.get $5
+   local.get $3
+   i64.const 12
+   i64.shl
+   i64.clz
+   i64.sub
+   local.tee $5
+   i64.sub
+   i64.shl
+  else
+   local.get $3
+   i64.const 4503599627370495
+   i64.and
+   i64.const 4503599627370496
+   i64.or
+  end
+  local.set $3
+  loop $while-continue|0
    local.get $4
-   i64.eqz
-   if (result i64)
-    local.get $2
-    i64.const 1
-    local.get $4
-    local.get $2
-    i64.const 12
-    i64.shl
-    i64.clz
-    i64.sub
-    local.tee $4
-    i64.sub
-    i64.shl
-   else
-    local.get $2
-    i64.const 4503599627370495
-    i64.and
-    i64.const 4503599627370496
-    i64.or
-   end
-   local.set $2
    local.get $5
-   i64.eqz
-   if (result i64)
-    local.get $3
-    i64.const 1
-    local.get $5
-    local.get $3
-    i64.const 12
-    i64.shl
-    i64.clz
-    i64.sub
-    local.tee $5
-    i64.sub
-    i64.shl
-   else
+   i64.gt_s
+   if
+    local.get $2
     local.get $3
-    i64.const 4503599627370495
-    i64.and
-    i64.const 4503599627370496
-    i64.or
-   end
-   local.set $3
-   loop $while-continue|0
-    local.get $4
-    local.get $5
-    i64.gt_s
-    if
+    i64.ge_u
+    if (result i64)
      local.get $2
      local.get $3
-     i64.ge_u
-     if (result i64)
-      local.get $2
-      local.get $3
-      i64.eq
-      br_if $folding-inner0
-      local.get $2
-      local.get $3
-      i64.sub
-     else
-      local.get $2
+     i64.eq
+     if
+      local.get $0
+      f64.const 0
+      f64.mul
+      return
      end
-     i64.const 1
-     i64.shl
-     local.set $2
-     local.get $4
-     i64.const 1
+     local.get $2
+     local.get $3
      i64.sub
-     local.set $4
-     br $while-continue|0
+    else
+     local.get $2
     end
+    i64.const 1
+    i64.shl
+    local.set $2
+    local.get $4
+    i64.const 1
+    i64.sub
+    local.set $4
+    br $while-continue|0
    end
+  end
+  local.get $2
+  local.get $3
+  i64.ge_u
+  if
    local.get $2
    local.get $3
-   i64.ge_u
+   i64.eq
    if
-    local.get $2
-    local.get $3
-    i64.eq
-    br_if $folding-inner0
-    local.get $2
-    local.get $3
-    i64.sub
-    local.set $2
+    local.get $0
+    f64.const 0
+    f64.mul
+    return
    end
    local.get $2
-   local.get $2
-   i64.const 11
-   i64.shl
-   i64.clz
-   local.tee $3
-   i64.shl
-   local.set $2
-   local.get $4
    local.get $3
    i64.sub
-   local.tee $4
-   i64.const 0
-   i64.gt_s
-   if (result i64)
-    local.get $2
-    i64.const 4503599627370496
-    i64.sub
-    local.get $4
-    i64.const 52
-    i64.shl
-    i64.or
-   else
-    local.get $2
-    i64.const 1
-    local.get $4
-    i64.sub
-    i64.shr_u
-   end
-   local.get $7
-   i64.const 63
+   local.set $2
+  end
+  local.get $2
+  local.get $2
+  i64.const 11
+  i64.shl
+  i64.clz
+  local.tee $3
+  i64.shl
+  local.set $2
+  local.get $4
+  local.get $3
+  i64.sub
+  local.tee $4
+  i64.const 0
+  i64.gt_s
+  if (result i64)
+   local.get $2
+   i64.const 4503599627370496
+   i64.sub
+   local.get $4
+   i64.const 52
    i64.shl
    i64.or
-   f64.reinterpret_i64
-   return
+  else
+   local.get $2
+   i64.const 1
+   local.get $4
+   i64.sub
+   i64.shr_u
   end
-  local.get $0
-  f64.const 0
-  f64.mul
+  local.get $7
+  i64.or
+  f64.reinterpret_i64
  )
  (func $std/math/test_mod (param $0 f64) (param $1 f64) (param $2 f64) (result i32)
   local.get $0
@@ -6741,142 +6746,149 @@
    f32.div
    return
   end
-  block $folding-inner0
+  local.get $2
+  i32.const 1
+  i32.shl
+  local.tee $8
+  local.get $6
+  i32.le_u
+  if
+   local.get $0
+   f32.const 0
+   f32.mul
+   local.get $0
+   local.get $6
+   local.get $8
+   i32.eq
+   select
+   return
+  end
+  local.get $4
+  if (result i32)
+   local.get $2
+   i32.const 8388607
+   i32.and
+   i32.const 8388608
+   i32.or
+  else
    local.get $2
    i32.const 1
+   local.get $4
+   local.get $2
+   i32.const 9
    i32.shl
-   local.tee $8
-   local.get $6
-   i32.le_u
-   if
-    local.get $6
-    local.get $8
-    i32.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i32.clz
+   i32.sub
+   local.tee $4
+   i32.sub
+   i32.shl
+  end
+  local.set $2
+  local.get $5
+  if (result i32)
+   local.get $3
+   i32.const 8388607
+   i32.and
+   i32.const 8388608
+   i32.or
+  else
+   local.get $3
+   i32.const 1
+   local.get $5
+   local.get $3
+   i32.const 9
+   i32.shl
+   i32.clz
+   i32.sub
+   local.tee $5
+   i32.sub
+   i32.shl
+  end
+  local.set $3
+  loop $while-continue|0
    local.get $4
-   if (result i32)
-    local.get $2
-    i32.const 8388607
-    i32.and
-    i32.const 8388608
-    i32.or
-   else
-    local.get $2
-    i32.const 1
-    local.get $4
-    local.get $2
-    i32.const 9
-    i32.shl
-    i32.clz
-    i32.sub
-    local.tee $4
-    i32.sub
-    i32.shl
-   end
-   local.set $2
    local.get $5
-   if (result i32)
-    local.get $3
-    i32.const 8388607
-    i32.and
-    i32.const 8388608
-    i32.or
-   else
-    local.get $3
-    i32.const 1
-    local.get $5
+   i32.gt_s
+   if
+    local.get $2
     local.get $3
-    i32.const 9
-    i32.shl
-    i32.clz
-    i32.sub
-    local.tee $5
-    i32.sub
-    i32.shl
-   end
-   local.set $3
-   loop $while-continue|0
-    local.get $4
-    local.get $5
-    i32.gt_s
-    if
+    i32.ge_u
+    if (result i32)
      local.get $2
      local.get $3
-     i32.ge_u
-     if (result i32)
-      local.get $2
-      local.get $3
-      i32.eq
-      br_if $folding-inner0
-      local.get $2
-      local.get $3
-      i32.sub
-     else
-      local.get $2
+     i32.eq
+     if
+      local.get $0
+      f32.const 0
+      f32.mul
+      return
      end
-     i32.const 1
-     i32.shl
-     local.set $2
-     local.get $4
-     i32.const 1
+     local.get $2
+     local.get $3
      i32.sub
-     local.set $4
-     br $while-continue|0
+    else
+     local.get $2
     end
+    i32.const 1
+    i32.shl
+    local.set $2
+    local.get $4
+    i32.const 1
+    i32.sub
+    local.set $4
+    br $while-continue|0
    end
+  end
+  local.get $2
+  local.get $3
+  i32.ge_u
+  if
    local.get $2
    local.get $3
-   i32.ge_u
+   i32.eq
    if
-    local.get $2
-    local.get $3
-    i32.eq
-    br_if $folding-inner0
-    local.get $2
-    local.get $3
-    i32.sub
-    local.set $2
+    local.get $0
+    f32.const 0
+    f32.mul
+    return
    end
    local.get $2
+   local.get $3
+   i32.sub
+   local.set $2
+  end
+  local.get $2
+  local.get $2
+  i32.const 8
+  i32.shl
+  i32.clz
+  local.tee $3
+  i32.shl
+  local.set $2
+  local.get $4
+  local.get $3
+  i32.sub
+  local.tee $4
+  i32.const 0
+  i32.gt_s
+  if (result i32)
    local.get $2
-   i32.const 8
-   i32.shl
-   i32.clz
-   local.tee $3
+   i32.const 8388608
+   i32.sub
+   local.get $4
+   i32.const 23
    i32.shl
-   local.set $2
+   i32.or
+  else
+   local.get $2
+   i32.const 1
    local.get $4
-   local.get $3
    i32.sub
-   local.tee $4
-   i32.const 0
-   i32.gt_s
-   if (result i32)
-    local.get $2
-    i32.const 8388608
-    i32.sub
-    local.get $4
-    i32.const 23
-    i32.shl
-    i32.or
-   else
-    local.get $2
-    i32.const 1
-    local.get $4
-    i32.sub
-    i32.shr_u
-   end
-   local.get $7
-   i32.or
-   f32.reinterpret_i32
-   return
+   i32.shr_u
   end
-  local.get $0
-  f32.const 0
-  f32.mul
+  local.get $7
+  i32.or
+  f32.reinterpret_i32
  )
  (func $~lib/math/NativeMath.pow (param $0 f64) (param $1 f64) (result f64)
   (local $2 f64)
diff --git a/tests/compiler/std/math.untouched.wat b/tests/compiler/std/math.untouched.wat
index 22a4c8bbc2..5a4750e43c 100644
--- a/tests/compiler/std/math.untouched.wat
+++ b/tests/compiler/std/math.untouched.wat
@@ -9185,8 +9185,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $6
   local.get $3
   i64.const 1
@@ -9230,13 +9230,13 @@
    local.get $9
    local.get $7
    i64.eq
-   if
+   if (result f64)
     f64.const 0
     local.get $0
     f64.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -9402,11 +9402,7 @@
   end
   local.get $2
   local.get $6
-  i64.const 63
-  i64.shl
   i64.or
-  local.set $2
-  local.get $2
   f64.reinterpret_i64
  )
  (func $std/math/test_mod (param $0 f64) (param $1 f64) (param $2 f64) (param $3 f64) (param $4 i32) (result i32)
@@ -9523,13 +9519,13 @@
    local.get $9
    local.get $7
    i32.eq
-   if
+   if (result f32)
     f32.const 0
     local.get $0
     f32.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -9692,8 +9688,6 @@
   local.get $2
   local.get $6
   i32.or
-  local.set $2
-  local.get $2
   f32.reinterpret_i32
  )
  (func $std/math/test_modf (param $0 f32) (param $1 f32) (param $2 f32) (param $3 f32) (param $4 i32) (result i32)
diff --git a/tests/compiler/std/mod.optimized.wat b/tests/compiler/std/mod.optimized.wat
index bb2ae1418a..c5df7a91f2 100644
--- a/tests/compiler/std/mod.optimized.wat
+++ b/tests/compiler/std/mod.optimized.wat
@@ -45,8 +45,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $7
   i32.const 1
   local.get $1
@@ -77,146 +77,151 @@
    f64.div
    return
   end
-  block $folding-inner0
+  local.get $2
+  i64.const 1
+  i64.shl
+  local.tee $8
+  local.get $6
+  i64.le_u
+  if
+   local.get $0
+   f64.const 0
+   f64.mul
+   local.get $0
+   local.get $6
+   local.get $8
+   i64.eq
+   select
+   return
+  end
+  local.get $4
+  i64.eqz
+  if (result i64)
    local.get $2
    i64.const 1
+   local.get $4
+   local.get $2
+   i64.const 12
    i64.shl
-   local.tee $8
-   local.get $6
-   i64.le_u
-   if
-    local.get $6
-    local.get $8
-    i64.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i64.clz
+   i64.sub
+   local.tee $4
+   i64.sub
+   i64.shl
+  else
+   local.get $2
+   i64.const 4503599627370495
+   i64.and
+   i64.const 4503599627370496
+   i64.or
+  end
+  local.set $2
+  local.get $5
+  i64.eqz
+  if (result i64)
+   local.get $3
+   i64.const 1
+   local.get $5
+   local.get $3
+   i64.const 12
+   i64.shl
+   i64.clz
+   i64.sub
+   local.tee $5
+   i64.sub
+   i64.shl
+  else
+   local.get $3
+   i64.const 4503599627370495
+   i64.and
+   i64.const 4503599627370496
+   i64.or
+  end
+  local.set $3
+  loop $while-continue|0
    local.get $4
-   i64.eqz
-   if (result i64)
-    local.get $2
-    i64.const 1
-    local.get $4
-    local.get $2
-    i64.const 12
-    i64.shl
-    i64.clz
-    i64.sub
-    local.tee $4
-    i64.sub
-    i64.shl
-   else
-    local.get $2
-    i64.const 4503599627370495
-    i64.and
-    i64.const 4503599627370496
-    i64.or
-   end
-   local.set $2
    local.get $5
-   i64.eqz
-   if (result i64)
-    local.get $3
-    i64.const 1
-    local.get $5
-    local.get $3
-    i64.const 12
-    i64.shl
-    i64.clz
-    i64.sub
-    local.tee $5
-    i64.sub
-    i64.shl
-   else
+   i64.gt_s
+   if
+    local.get $2
     local.get $3
-    i64.const 4503599627370495
-    i64.and
-    i64.const 4503599627370496
-    i64.or
-   end
-   local.set $3
-   loop $while-continue|0
-    local.get $4
-    local.get $5
-    i64.gt_s
-    if
+    i64.ge_u
+    if (result i64)
      local.get $2
      local.get $3
-     i64.ge_u
-     if (result i64)
-      local.get $2
-      local.get $3
-      i64.eq
-      br_if $folding-inner0
-      local.get $2
-      local.get $3
-      i64.sub
-     else
-      local.get $2
+     i64.eq
+     if
+      local.get $0
+      f64.const 0
+      f64.mul
+      return
      end
-     i64.const 1
-     i64.shl
-     local.set $2
-     local.get $4
-     i64.const 1
+     local.get $2
+     local.get $3
      i64.sub
-     local.set $4
-     br $while-continue|0
+    else
+     local.get $2
     end
+    i64.const 1
+    i64.shl
+    local.set $2
+    local.get $4
+    i64.const 1
+    i64.sub
+    local.set $4
+    br $while-continue|0
    end
+  end
+  local.get $2
+  local.get $3
+  i64.ge_u
+  if
    local.get $2
    local.get $3
-   i64.ge_u
+   i64.eq
    if
-    local.get $2
-    local.get $3
-    i64.eq
-    br_if $folding-inner0
-    local.get $2
-    local.get $3
-    i64.sub
-    local.set $2
+    local.get $0
+    f64.const 0
+    f64.mul
+    return
    end
    local.get $2
-   local.get $2
-   i64.const 11
-   i64.shl
-   i64.clz
-   local.tee $3
-   i64.shl
-   local.set $2
-   local.get $4
    local.get $3
    i64.sub
-   local.tee $4
-   i64.const 0
-   i64.gt_s
-   if (result i64)
-    local.get $2
-    i64.const 4503599627370496
-    i64.sub
-    local.get $4
-    i64.const 52
-    i64.shl
-    i64.or
-   else
-    local.get $2
-    i64.const 1
-    local.get $4
-    i64.sub
-    i64.shr_u
-   end
-   local.get $7
-   i64.const 63
+   local.set $2
+  end
+  local.get $2
+  local.get $2
+  i64.const 11
+  i64.shl
+  i64.clz
+  local.tee $3
+  i64.shl
+  local.set $2
+  local.get $4
+  local.get $3
+  i64.sub
+  local.tee $4
+  i64.const 0
+  i64.gt_s
+  if (result i64)
+   local.get $2
+   i64.const 4503599627370496
+   i64.sub
+   local.get $4
+   i64.const 52
    i64.shl
    i64.or
-   f64.reinterpret_i64
-   return
+  else
+   local.get $2
+   i64.const 1
+   local.get $4
+   i64.sub
+   i64.shr_u
   end
-  local.get $0
-  f64.const 0
-  f64.mul
+  local.get $7
+  i64.or
+  f64.reinterpret_i64
  )
  (func $std/mod/check<f64> (param $0 f64) (param $1 f64) (result i32)
   local.get $1
@@ -324,142 +329,149 @@
    f32.div
    return
   end
-  block $folding-inner0
+  local.get $2
+  i32.const 1
+  i32.shl
+  local.tee $8
+  local.get $6
+  i32.le_u
+  if
+   local.get $0
+   f32.const 0
+   f32.mul
+   local.get $0
+   local.get $6
+   local.get $8
+   i32.eq
+   select
+   return
+  end
+  local.get $4
+  if (result i32)
+   local.get $2
+   i32.const 8388607
+   i32.and
+   i32.const 8388608
+   i32.or
+  else
    local.get $2
    i32.const 1
+   local.get $4
+   local.get $2
+   i32.const 9
    i32.shl
-   local.tee $8
-   local.get $6
-   i32.le_u
-   if
-    local.get $6
-    local.get $8
-    i32.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i32.clz
+   i32.sub
+   local.tee $4
+   i32.sub
+   i32.shl
+  end
+  local.set $2
+  local.get $5
+  if (result i32)
+   local.get $3
+   i32.const 8388607
+   i32.and
+   i32.const 8388608
+   i32.or
+  else
+   local.get $3
+   i32.const 1
+   local.get $5
+   local.get $3
+   i32.const 9
+   i32.shl
+   i32.clz
+   i32.sub
+   local.tee $5
+   i32.sub
+   i32.shl
+  end
+  local.set $3
+  loop $while-continue|0
    local.get $4
-   if (result i32)
-    local.get $2
-    i32.const 8388607
-    i32.and
-    i32.const 8388608
-    i32.or
-   else
-    local.get $2
-    i32.const 1
-    local.get $4
-    local.get $2
-    i32.const 9
-    i32.shl
-    i32.clz
-    i32.sub
-    local.tee $4
-    i32.sub
-    i32.shl
-   end
-   local.set $2
    local.get $5
-   if (result i32)
-    local.get $3
-    i32.const 8388607
-    i32.and
-    i32.const 8388608
-    i32.or
-   else
-    local.get $3
-    i32.const 1
-    local.get $5
+   i32.gt_s
+   if
+    local.get $2
     local.get $3
-    i32.const 9
-    i32.shl
-    i32.clz
-    i32.sub
-    local.tee $5
-    i32.sub
-    i32.shl
-   end
-   local.set $3
-   loop $while-continue|0
-    local.get $4
-    local.get $5
-    i32.gt_s
-    if
+    i32.ge_u
+    if (result i32)
      local.get $2
      local.get $3
-     i32.ge_u
-     if (result i32)
-      local.get $2
-      local.get $3
-      i32.eq
-      br_if $folding-inner0
-      local.get $2
-      local.get $3
-      i32.sub
-     else
-      local.get $2
+     i32.eq
+     if
+      local.get $0
+      f32.const 0
+      f32.mul
+      return
      end
-     i32.const 1
-     i32.shl
-     local.set $2
-     local.get $4
-     i32.const 1
+     local.get $2
+     local.get $3
      i32.sub
-     local.set $4
-     br $while-continue|0
+    else
+     local.get $2
     end
+    i32.const 1
+    i32.shl
+    local.set $2
+    local.get $4
+    i32.const 1
+    i32.sub
+    local.set $4
+    br $while-continue|0
    end
+  end
+  local.get $2
+  local.get $3
+  i32.ge_u
+  if
    local.get $2
    local.get $3
-   i32.ge_u
+   i32.eq
    if
-    local.get $2
-    local.get $3
-    i32.eq
-    br_if $folding-inner0
-    local.get $2
-    local.get $3
-    i32.sub
-    local.set $2
+    local.get $0
+    f32.const 0
+    f32.mul
+    return
    end
    local.get $2
+   local.get $3
+   i32.sub
+   local.set $2
+  end
+  local.get $2
+  local.get $2
+  i32.const 8
+  i32.shl
+  i32.clz
+  local.tee $3
+  i32.shl
+  local.set $2
+  local.get $4
+  local.get $3
+  i32.sub
+  local.tee $4
+  i32.const 0
+  i32.gt_s
+  if (result i32)
    local.get $2
-   i32.const 8
-   i32.shl
-   i32.clz
-   local.tee $3
+   i32.const 8388608
+   i32.sub
+   local.get $4
+   i32.const 23
    i32.shl
-   local.set $2
+   i32.or
+  else
+   local.get $2
+   i32.const 1
    local.get $4
-   local.get $3
    i32.sub
-   local.tee $4
-   i32.const 0
-   i32.gt_s
-   if (result i32)
-    local.get $2
-    i32.const 8388608
-    i32.sub
-    local.get $4
-    i32.const 23
-    i32.shl
-    i32.or
-   else
-    local.get $2
-    i32.const 1
-    local.get $4
-    i32.sub
-    i32.shr_u
-   end
-   local.get $7
-   i32.or
-   f32.reinterpret_i32
-   return
+   i32.shr_u
   end
-  local.get $0
-  f32.const 0
-  f32.mul
+  local.get $7
+  i32.or
+  f32.reinterpret_i32
  )
  (func $std/mod/test_fmodf (param $0 f32) (param $1 f32) (param $2 f32) (result i32)
   block $__inlined_func$std/mod/check<f32> (result i32)
diff --git a/tests/compiler/std/mod.untouched.wat b/tests/compiler/std/mod.untouched.wat
index 3edcfeadf3..5790978826 100644
--- a/tests/compiler/std/mod.untouched.wat
+++ b/tests/compiler/std/mod.untouched.wat
@@ -59,8 +59,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $6
   local.get $3
   i64.const 1
@@ -104,13 +104,13 @@
    local.get $9
    local.get $7
    i64.eq
-   if
+   if (result f64)
     f64.const 0
     local.get $0
     f64.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -276,11 +276,7 @@
   end
   local.get $2
   local.get $6
-  i64.const 63
-  i64.shl
   i64.or
-  local.set $2
-  local.get $2
   f64.reinterpret_i64
  )
  (func $std/mod/check<f64> (param $0 f64) (param $1 f64) (result i32)
@@ -421,13 +417,13 @@
    local.get $9
    local.get $7
    i32.eq
-   if
+   if (result f32)
     f32.const 0
     local.get $0
     f32.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -590,8 +586,6 @@
   local.get $2
   local.get $6
   i32.or
-  local.set $2
-  local.get $2
   f32.reinterpret_i32
  )
  (func $std/mod/check<f32> (param $0 f32) (param $1 f32) (result i32)
diff --git a/tests/compiler/std/typedarray.optimized.wat b/tests/compiler/std/typedarray.optimized.wat
index 2b7d584dfd..899b03046d 100644
--- a/tests/compiler/std/typedarray.optimized.wat
+++ b/tests/compiler/std/typedarray.optimized.wat
@@ -9616,121 +9616,128 @@
    f32.div
    return
   end
-  block $folding-inner0
+  local.get $1
+  i32.const 1
+  i32.shl
+  local.tee $3
+  i32.const -2147483648
+  i32.le_u
+  if
+   local.get $0
+   f32.const 0
+   f32.mul
+   local.get $0
+   local.get $3
+   i32.const -2147483648
+   i32.eq
+   select
+   return
+  end
+  local.get $2
+  if (result i32)
+   local.get $1
+   i32.const 8388607
+   i32.and
+   i32.const 8388608
+   i32.or
+  else
    local.get $1
    i32.const 1
+   local.get $2
+   local.get $1
+   i32.const 9
    i32.shl
-   local.tee $3
-   i32.const -2147483648
-   i32.le_u
-   if
-    local.get $3
-    i32.const -2147483648
-    i32.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i32.clz
+   i32.sub
+   local.tee $2
+   i32.sub
+   i32.shl
+  end
+  local.set $1
+  loop $while-continue|0
    local.get $2
-   if (result i32)
+   i32.const 128
+   i32.gt_s
+   if
     local.get $1
-    i32.const 8388607
-    i32.and
     i32.const 8388608
-    i32.or
-   else
-    local.get $1
-    i32.const 1
-    local.get $2
-    local.get $1
-    i32.const 9
-    i32.shl
-    i32.clz
-    i32.sub
-    local.tee $2
-    i32.sub
-    i32.shl
-   end
-   local.set $1
-   loop $while-continue|0
-    local.get $2
-    i32.const 128
-    i32.gt_s
-    if
+    i32.ge_u
+    if (result i32)
      local.get $1
      i32.const 8388608
-     i32.ge_u
-     if (result i32)
-      local.get $1
-      i32.const 8388608
-      i32.eq
-      br_if $folding-inner0
-      local.get $1
-      i32.const 8388608
-      i32.sub
-     else
-      local.get $1
+     i32.eq
+     if
+      local.get $0
+      f32.const 0
+      f32.mul
+      return
      end
-     i32.const 1
-     i32.shl
-     local.set $1
-     local.get $2
-     i32.const 1
+     local.get $1
+     i32.const 8388608
      i32.sub
-     local.set $2
-     br $while-continue|0
+    else
+     local.get $1
     end
+    i32.const 1
+    i32.shl
+    local.set $1
+    local.get $2
+    i32.const 1
+    i32.sub
+    local.set $2
+    br $while-continue|0
    end
+  end
+  local.get $1
+  i32.const 8388608
+  i32.ge_u
+  if
    local.get $1
    i32.const 8388608
-   i32.ge_u
+   i32.eq
    if
-    local.get $1
-    i32.const 8388608
-    i32.eq
-    br_if $folding-inner0
-    local.get $1
-    i32.const 8388608
-    i32.sub
-    local.set $1
+    local.get $0
+    f32.const 0
+    f32.mul
+    return
    end
    local.get $1
+   i32.const 8388608
+   i32.sub
+   local.set $1
+  end
+  local.get $1
+  local.get $1
+  i32.const 8
+  i32.shl
+  i32.clz
+  local.tee $3
+  i32.shl
+  local.set $1
+  local.get $2
+  local.get $3
+  i32.sub
+  local.tee $2
+  i32.const 0
+  i32.gt_s
+  if (result i32)
    local.get $1
-   i32.const 8
-   i32.shl
-   i32.clz
-   local.tee $3
+   i32.const 8388608
+   i32.sub
+   local.get $2
+   i32.const 23
    i32.shl
-   local.set $1
+   i32.or
+  else
+   local.get $1
+   i32.const 1
    local.get $2
-   local.get $3
    i32.sub
-   local.tee $2
-   i32.const 0
-   i32.gt_s
-   if (result i32)
-    local.get $1
-    i32.const 8388608
-    i32.sub
-    local.get $2
-    i32.const 23
-    i32.shl
-    i32.or
-   else
-    local.get $1
-    i32.const 1
-    local.get $2
-    i32.sub
-    i32.shr_u
-   end
-   local.get $4
-   i32.or
-   f32.reinterpret_i32
-   return
+   i32.shr_u
   end
-  local.get $0
-  f32.const 0
-  f32.mul
+  local.get $4
+  i32.or
+  f32.reinterpret_i32
  )
  (func $std/typedarray/testArrayEvery<~lib/typedarray/Float32Array,f32>~anonymous|0 (param $0 f32) (param $1 i32) (param $2 i32) (result i32)
   local.get $2
@@ -9818,8 +9825,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $1
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $4
   local.get $1
   i64.const 52
@@ -9838,124 +9845,129 @@
    f64.div
    return
   end
-  block $folding-inner0
+  local.get $1
+  i64.const 1
+  i64.shl
+  local.tee $3
+  i64.const -9223372036854775808
+  i64.le_u
+  if
+   local.get $0
+   f64.const 0
+   f64.mul
+   local.get $0
+   local.get $3
+   i64.const -9223372036854775808
+   i64.eq
+   select
+   return
+  end
+  local.get $2
+  i64.eqz
+  if (result i64)
    local.get $1
    i64.const 1
+   local.get $2
+   local.get $1
+   i64.const 12
    i64.shl
-   local.tee $3
-   i64.const -9223372036854775808
-   i64.le_u
-   if
-    local.get $3
-    i64.const -9223372036854775808
-    i64.eq
-    br_if $folding-inner0
-    local.get $0
-    return
-   end
+   i64.clz
+   i64.sub
+   local.tee $2
+   i64.sub
+   i64.shl
+  else
+   local.get $1
+   i64.const 4503599627370495
+   i64.and
+   i64.const 4503599627370496
+   i64.or
+  end
+  local.set $1
+  loop $while-continue|0
    local.get $2
-   i64.eqz
-   if (result i64)
-    local.get $1
-    i64.const 1
-    local.get $2
-    local.get $1
-    i64.const 12
-    i64.shl
-    i64.clz
-    i64.sub
-    local.tee $2
-    i64.sub
-    i64.shl
-   else
+   i64.const 1024
+   i64.gt_s
+   if
     local.get $1
-    i64.const 4503599627370495
-    i64.and
     i64.const 4503599627370496
-    i64.or
-   end
-   local.set $1
-   loop $while-continue|0
-    local.get $2
-    i64.const 1024
-    i64.gt_s
-    if
+    i64.ge_u
+    if (result i64)
      local.get $1
      i64.const 4503599627370496
-     i64.ge_u
-     if (result i64)
-      local.get $1
-      i64.const 4503599627370496
-      i64.eq
-      br_if $folding-inner0
-      local.get $1
-      i64.const 4503599627370496
-      i64.sub
-     else
-      local.get $1
+     i64.eq
+     if
+      local.get $0
+      f64.const 0
+      f64.mul
+      return
      end
-     i64.const 1
-     i64.shl
-     local.set $1
-     local.get $2
-     i64.const 1
+     local.get $1
+     i64.const 4503599627370496
      i64.sub
-     local.set $2
-     br $while-continue|0
+    else
+     local.get $1
     end
+    i64.const 1
+    i64.shl
+    local.set $1
+    local.get $2
+    i64.const 1
+    i64.sub
+    local.set $2
+    br $while-continue|0
    end
+  end
+  local.get $1
+  i64.const 4503599627370496
+  i64.ge_u
+  if
    local.get $1
    i64.const 4503599627370496
-   i64.ge_u
+   i64.eq
    if
-    local.get $1
-    i64.const 4503599627370496
-    i64.eq
-    br_if $folding-inner0
-    local.get $1
-    i64.const 4503599627370496
-    i64.sub
-    local.set $1
+    local.get $0
+    f64.const 0
+    f64.mul
+    return
    end
    local.get $1
-   local.get $1
-   i64.const 11
-   i64.shl
-   i64.clz
-   local.tee $3
-   i64.shl
+   i64.const 4503599627370496
+   i64.sub
    local.set $1
-   local.get $2
-   local.get $3
+  end
+  local.get $1
+  local.get $1
+  i64.const 11
+  i64.shl
+  i64.clz
+  local.tee $3
+  i64.shl
+  local.set $1
+  local.get $2
+  local.get $3
+  i64.sub
+  local.tee $2
+  i64.const 0
+  i64.gt_s
+  if (result i64)
+   local.get $1
+   i64.const 4503599627370496
    i64.sub
-   local.tee $2
-   i64.const 0
-   i64.gt_s
-   if (result i64)
-    local.get $1
-    i64.const 4503599627370496
-    i64.sub
-    local.get $2
-    i64.const 52
-    i64.shl
-    i64.or
-   else
-    local.get $1
-    i64.const 1
-    local.get $2
-    i64.sub
-    i64.shr_u
-   end
-   local.get $4
-   i64.const 63
+   local.get $2
+   i64.const 52
    i64.shl
    i64.or
-   f64.reinterpret_i64
-   return
+  else
+   local.get $1
+   i64.const 1
+   local.get $2
+   i64.sub
+   i64.shr_u
   end
-  local.get $0
-  f64.const 0
-  f64.mul
+  local.get $4
+  i64.or
+  f64.reinterpret_i64
  )
  (func $std/typedarray/testArrayEvery<~lib/typedarray/Float64Array,f64>~anonymous|0 (param $0 f64) (param $1 i32) (param $2 i32) (result i32)
   local.get $2
diff --git a/tests/compiler/std/typedarray.untouched.wat b/tests/compiler/std/typedarray.untouched.wat
index ddc0bedaf9..477af1e666 100644
--- a/tests/compiler/std/typedarray.untouched.wat
+++ b/tests/compiler/std/typedarray.untouched.wat
@@ -19104,13 +19104,13 @@
    local.get $9
    local.get $7
    i32.eq
-   if
+   if (result f32)
     f32.const 0
     local.get $0
     f32.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -19273,8 +19273,6 @@
   local.get $2
   local.get $6
   i32.or
-  local.set $2
-  local.get $2
   f32.reinterpret_i32
  )
  (func $std/typedarray/testArrayEvery<~lib/typedarray/Float32Array,f32>~anonymous|0 (param $0 f32) (param $1 i32) (param $2 i32) (result i32)
@@ -19486,8 +19484,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const 63
-  i64.shr_u
+  i64.const -9223372036854775808
+  i64.and
   local.set $6
   local.get $3
   i64.const 1
@@ -19531,13 +19529,13 @@
    local.get $9
    local.get $7
    i64.eq
-   if
+   if (result f64)
     f64.const 0
     local.get $0
     f64.mul
-    return
+   else
+    local.get $0
    end
-   local.get $0
    return
   end
   local.get $4
@@ -19703,11 +19701,7 @@
   end
   local.get $2
   local.get $6
-  i64.const 63
-  i64.shl
   i64.or
-  local.set $2
-  local.get $2
   f64.reinterpret_i64
  )
  (func $std/typedarray/testArrayEvery<~lib/typedarray/Float64Array,f64>~anonymous|0 (param $0 f64) (param $1 i32) (param $2 i32) (result i32)

From 7787cf3ff298a9cfdd71479d660ae42faef4dd72 Mon Sep 17 00:00:00 2001
From: MaxGraey <maxgraey@gmail.com>
Date: Mon, 26 Oct 2020 18:30:13 +0200
Subject: [PATCH 2/4] more

---
 std/assembly/math.ts                        |  4 ++--
 tests/compiler/binary.untouched.wat         | 24 +++++++--------------
 tests/compiler/std/math.optimized.wat       | 16 ++++++--------
 tests/compiler/std/math.untouched.wat       | 24 +++++++--------------
 tests/compiler/std/mod.optimized.wat        | 16 ++++++--------
 tests/compiler/std/mod.untouched.wat        | 24 +++++++--------------
 tests/compiler/std/typedarray.optimized.wat | 16 ++++++--------
 tests/compiler/std/typedarray.untouched.wat | 24 +++++++--------------
 8 files changed, 52 insertions(+), 96 deletions(-)

diff --git a/std/assembly/math.ts b/std/assembly/math.ts
index 4b37306fe9..2a1533070f 100644
--- a/std/assembly/math.ts
+++ b/std/assembly/math.ts
@@ -1583,7 +1583,7 @@ export namespace NativeMath {
     }
     var ux1 = ux << 1;
     if (ux1 <= uy1) {
-      return ux1 == uy1 ? 0 * x : x;
+      return x * f64(ux1 != uy1);
     }
     if (!ex) {
       ex -= builtin_clz<i64>(ux << 12);
@@ -2890,7 +2890,7 @@ export namespace NativeMathf {
     }
     var ux1 = ux << 1;
     if (ux1 <= uy1) {
-      return ux1 == uy1 ? 0 * x : x;
+      return x * f32(ux1 != uy1);
     }
     if (!ex) {
       ex -= builtin_clz<u32>(ux << 9);
diff --git a/tests/compiler/binary.untouched.wat b/tests/compiler/binary.untouched.wat
index 7a32ac7e7d..bf31c6078e 100644
--- a/tests/compiler/binary.untouched.wat
+++ b/tests/compiler/binary.untouched.wat
@@ -1288,16 +1288,12 @@
   local.get $7
   i32.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i32.eq
-   if (result f32)
-    f32.const 0
-    local.get $0
-    f32.mul
-   else
-    local.get $0
-   end
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
@@ -2161,16 +2157,12 @@
   local.get $7
   i64.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i64.eq
-   if (result f64)
-    f64.const 0
-    local.get $0
-    f64.mul
-   else
-    local.get $0
-   end
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4
diff --git a/tests/compiler/std/math.optimized.wat b/tests/compiler/std/math.optimized.wat
index 8449ad1b56..f929d6b24c 100644
--- a/tests/compiler/std/math.optimized.wat
+++ b/tests/compiler/std/math.optimized.wat
@@ -6526,14 +6526,12 @@
   local.get $6
   i64.le_u
   if
-   local.get $0
-   f64.const 0
-   f64.mul
    local.get $0
    local.get $6
    local.get $8
-   i64.eq
-   select
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4
@@ -6753,14 +6751,12 @@
   local.get $6
   i32.le_u
   if
-   local.get $0
-   f32.const 0
-   f32.mul
    local.get $0
    local.get $6
    local.get $8
-   i32.eq
-   select
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
diff --git a/tests/compiler/std/math.untouched.wat b/tests/compiler/std/math.untouched.wat
index 5a4750e43c..57505df8c0 100644
--- a/tests/compiler/std/math.untouched.wat
+++ b/tests/compiler/std/math.untouched.wat
@@ -9227,16 +9227,12 @@
   local.get $7
   i64.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i64.eq
-   if (result f64)
-    f64.const 0
-    local.get $0
-    f64.mul
-   else
-    local.get $0
-   end
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4
@@ -9516,16 +9512,12 @@
   local.get $7
   i32.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i32.eq
-   if (result f32)
-    f32.const 0
-    local.get $0
-    f32.mul
-   else
-    local.get $0
-   end
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
diff --git a/tests/compiler/std/mod.optimized.wat b/tests/compiler/std/mod.optimized.wat
index c5df7a91f2..6c399febf5 100644
--- a/tests/compiler/std/mod.optimized.wat
+++ b/tests/compiler/std/mod.optimized.wat
@@ -84,14 +84,12 @@
   local.get $6
   i64.le_u
   if
-   local.get $0
-   f64.const 0
-   f64.mul
    local.get $0
    local.get $6
    local.get $8
-   i64.eq
-   select
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4
@@ -336,14 +334,12 @@
   local.get $6
   i32.le_u
   if
-   local.get $0
-   f32.const 0
-   f32.mul
    local.get $0
    local.get $6
    local.get $8
-   i32.eq
-   select
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
diff --git a/tests/compiler/std/mod.untouched.wat b/tests/compiler/std/mod.untouched.wat
index 5790978826..3b4e02c1b6 100644
--- a/tests/compiler/std/mod.untouched.wat
+++ b/tests/compiler/std/mod.untouched.wat
@@ -101,16 +101,12 @@
   local.get $7
   i64.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i64.eq
-   if (result f64)
-    f64.const 0
-    local.get $0
-    f64.mul
-   else
-    local.get $0
-   end
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4
@@ -414,16 +410,12 @@
   local.get $7
   i32.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i32.eq
-   if (result f32)
-    f32.const 0
-    local.get $0
-    f32.mul
-   else
-    local.get $0
-   end
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
diff --git a/tests/compiler/std/typedarray.optimized.wat b/tests/compiler/std/typedarray.optimized.wat
index 899b03046d..499de64f02 100644
--- a/tests/compiler/std/typedarray.optimized.wat
+++ b/tests/compiler/std/typedarray.optimized.wat
@@ -9623,14 +9623,12 @@
   i32.const -2147483648
   i32.le_u
   if
-   local.get $0
-   f32.const 0
-   f32.mul
    local.get $0
    local.get $3
    i32.const -2147483648
-   i32.eq
-   select
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $2
@@ -9852,14 +9850,12 @@
   i64.const -9223372036854775808
   i64.le_u
   if
-   local.get $0
-   f64.const 0
-   f64.mul
    local.get $0
    local.get $3
    i64.const -9223372036854775808
-   i64.eq
-   select
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $2
diff --git a/tests/compiler/std/typedarray.untouched.wat b/tests/compiler/std/typedarray.untouched.wat
index 477af1e666..973b491a2b 100644
--- a/tests/compiler/std/typedarray.untouched.wat
+++ b/tests/compiler/std/typedarray.untouched.wat
@@ -19101,16 +19101,12 @@
   local.get $7
   i32.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i32.eq
-   if (result f32)
-    f32.const 0
-    local.get $0
-    f32.mul
-   else
-    local.get $0
-   end
+   i32.ne
+   f32.convert_i32_u
+   f32.mul
    return
   end
   local.get $4
@@ -19526,16 +19522,12 @@
   local.get $7
   i64.le_u
   if
+   local.get $0
    local.get $9
    local.get $7
-   i64.eq
-   if (result f64)
-    f64.const 0
-    local.get $0
-    f64.mul
-   else
-    local.get $0
-   end
+   i64.ne
+   f64.convert_i32_u
+   f64.mul
    return
   end
   local.get $4

From fe5a9d40cf97ef742cc3d45212568295c3ab6b3d Mon Sep 17 00:00:00 2001
From: MaxGraey <maxgraey@gmail.com>
Date: Tue, 27 Oct 2020 11:26:06 +0200
Subject: [PATCH 3/4] revert sign extraction for f64.mod

---
 std/assembly/math.ts                        | 4 ++--
 tests/compiler/binary.untouched.wat         | 6 ++++--
 tests/compiler/std/math.optimized.wat       | 6 ++++--
 tests/compiler/std/math.untouched.wat       | 6 ++++--
 tests/compiler/std/mod.optimized.wat        | 6 ++++--
 tests/compiler/std/mod.untouched.wat        | 6 ++++--
 tests/compiler/std/typedarray.optimized.wat | 6 ++++--
 tests/compiler/std/typedarray.untouched.wat | 6 ++++--
 8 files changed, 30 insertions(+), 16 deletions(-)

diff --git a/std/assembly/math.ts b/std/assembly/math.ts
index 2a1533070f..47241f079b 100644
--- a/std/assembly/math.ts
+++ b/std/assembly/math.ts
@@ -1575,7 +1575,7 @@ export namespace NativeMath {
     var uy = reinterpret<u64>(y);
     var ex = <i64>(ux >> 52 & 0x7FF);
     var ey = <i64>(uy >> 52 & 0x7FF);
-    var sm = ux & 0x8000000000000000;
+    var sx = ux >> 63;
     var uy1 = uy << 1;
     if (uy1 == 0 || ex == 0x7FF || isNaN<f64>(y)) {
       let m = x * y;
@@ -1621,7 +1621,7 @@ export namespace NativeMath {
     } else {
       ux >>= -ex + 1;
     }
-    return reinterpret<f64>(ux | sm);
+    return reinterpret<f64>(ux | (sx << 63));
   }
 
   export function rem(x: f64, y: f64): f64 { // see: musl/src/math/remquo.c
diff --git a/tests/compiler/binary.untouched.wat b/tests/compiler/binary.untouched.wat
index bf31c6078e..87a57a2b41 100644
--- a/tests/compiler/binary.untouched.wat
+++ b/tests/compiler/binary.untouched.wat
@@ -2115,8 +2115,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $6
   local.get $3
   i64.const 1
@@ -2328,6 +2328,8 @@
   end
   local.get $2
   local.get $6
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/math.optimized.wat b/tests/compiler/std/math.optimized.wat
index f929d6b24c..b3418d80b0 100644
--- a/tests/compiler/std/math.optimized.wat
+++ b/tests/compiler/std/math.optimized.wat
@@ -6487,8 +6487,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $7
   i32.const 1
   local.get $1
@@ -6660,6 +6660,8 @@
    i64.shr_u
   end
   local.get $7
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/math.untouched.wat b/tests/compiler/std/math.untouched.wat
index 57505df8c0..3ceaaf94f0 100644
--- a/tests/compiler/std/math.untouched.wat
+++ b/tests/compiler/std/math.untouched.wat
@@ -9185,8 +9185,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $6
   local.get $3
   i64.const 1
@@ -9398,6 +9398,8 @@
   end
   local.get $2
   local.get $6
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/mod.optimized.wat b/tests/compiler/std/mod.optimized.wat
index 6c399febf5..e399f928d3 100644
--- a/tests/compiler/std/mod.optimized.wat
+++ b/tests/compiler/std/mod.optimized.wat
@@ -45,8 +45,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $7
   i32.const 1
   local.get $1
@@ -218,6 +218,8 @@
    i64.shr_u
   end
   local.get $7
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/mod.untouched.wat b/tests/compiler/std/mod.untouched.wat
index 3b4e02c1b6..d97129d1d2 100644
--- a/tests/compiler/std/mod.untouched.wat
+++ b/tests/compiler/std/mod.untouched.wat
@@ -59,8 +59,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $6
   local.get $3
   i64.const 1
@@ -272,6 +272,8 @@
   end
   local.get $2
   local.get $6
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/typedarray.optimized.wat b/tests/compiler/std/typedarray.optimized.wat
index 499de64f02..c34566198c 100644
--- a/tests/compiler/std/typedarray.optimized.wat
+++ b/tests/compiler/std/typedarray.optimized.wat
@@ -9823,8 +9823,8 @@
   local.get $0
   i64.reinterpret_f64
   local.tee $1
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $4
   local.get $1
   i64.const 52
@@ -9962,6 +9962,8 @@
    i64.shr_u
   end
   local.get $4
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )
diff --git a/tests/compiler/std/typedarray.untouched.wat b/tests/compiler/std/typedarray.untouched.wat
index 973b491a2b..08a7fbbe07 100644
--- a/tests/compiler/std/typedarray.untouched.wat
+++ b/tests/compiler/std/typedarray.untouched.wat
@@ -19480,8 +19480,8 @@
   i64.and
   local.set $5
   local.get $2
-  i64.const -9223372036854775808
-  i64.and
+  i64.const 63
+  i64.shr_u
   local.set $6
   local.get $3
   i64.const 1
@@ -19693,6 +19693,8 @@
   end
   local.get $2
   local.get $6
+  i64.const 63
+  i64.shl
   i64.or
   f64.reinterpret_i64
  )

From 145d9703fe0f6dcb3d7b6efe3b87fb28130618fe Mon Sep 17 00:00:00 2001
From: MaxGraey <maxgraey@gmail.com>
Date: Tue, 27 Oct 2020 12:02:08 +0200
Subject: [PATCH 4/4] refactoring

---
 std/assembly/math.ts                        | 16 +++++------
 tests/compiler/binary.untouched.wat         | 16 +++--------
 tests/compiler/std/math.untouched.wat       | 32 ++++++---------------
 tests/compiler/std/mod.untouched.wat        | 16 +++--------
 tests/compiler/std/typedarray.untouched.wat | 16 +++--------
 5 files changed, 28 insertions(+), 68 deletions(-)

diff --git a/std/assembly/math.ts b/std/assembly/math.ts
index 47241f079b..0b4313a61d 100644
--- a/std/assembly/math.ts
+++ b/std/assembly/math.ts
@@ -1587,14 +1587,14 @@ export namespace NativeMath {
     }
     if (!ex) {
       ex -= builtin_clz<i64>(ux << 12);
-      ux <<= -ex + 1;
+      ux <<= 1 - ex;
     } else {
       ux &= <u64>-1 >> 12;
       ux |= 1 << 52;
     }
     if (!ey) {
       ey -= builtin_clz<i64>(uy << 12);
-      uy <<= -ey + 1;
+      uy <<= 1 - ey;
     } else {
       uy &= <u64>-1 >> 12;
       uy |= 1 << 52;
@@ -1638,14 +1638,14 @@ export namespace NativeMath {
     var uxi = ux;
     if (!ex) {
       ex -= builtin_clz<i64>(uxi << 12);
-      uxi <<= -ex + 1;
+      uxi <<= 1 - ex;
     } else {
       uxi &= <u64>-1 >> 12;
       uxi |= 1 << 52;
     }
     if (!ey) {
       ey -= builtin_clz<i64>(uy << 12);
-      uy <<= -ey + 1;
+      uy <<= 1 - ey;
     } else {
       uy &= <u64>-1 >> 12;
       uy |= 1 << 52;
@@ -2894,14 +2894,14 @@ export namespace NativeMathf {
     }
     if (!ex) {
       ex -= builtin_clz<u32>(ux << 9);
-      ux <<= -ex + 1;
+      ux <<= 1 - ex;
     } else {
       ux &= <u32>-1 >> 9;
       ux |= 1 << 23;
     }
     if (!ey) {
       ey -= builtin_clz<u32>(uy << 9);
-      uy <<= -ey + 1;
+      uy <<= 1 - ey;
     } else {
       uy &= <u32>-1 >> 9;
       uy |= 1 << 23;
@@ -2942,14 +2942,14 @@ export namespace NativeMathf {
     if (ux << 1 == 0) return x;
     if (!ex) {
       ex -= builtin_clz<u32>(uxi << 9);
-      uxi <<= -ex + 1;
+      uxi <<= 1 - ex;
     } else {
       uxi &= <u32>-1 >> 9;
       uxi |= 1 << 23;
     }
     if (!ey) {
       ey -= builtin_clz<u32>(uy << 9);
-      uy <<= -ey + 1;
+      uy <<= 1 - ey;
     } else {
       uy &= <u32>-1 >> 9;
       uy |= 1 << 23;
diff --git a/tests/compiler/binary.untouched.wat b/tests/compiler/binary.untouched.wat
index 87a57a2b41..fe63b928b4 100644
--- a/tests/compiler/binary.untouched.wat
+++ b/tests/compiler/binary.untouched.wat
@@ -1307,11 +1307,9 @@
    i32.sub
    local.set $4
    local.get $2
-   i32.const 0
+   i32.const 1
    local.get $4
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $2
   else
@@ -1339,11 +1337,9 @@
    i32.sub
    local.set $5
    local.get $3
-   i32.const 0
+   i32.const 1
    local.get $5
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $3
   else
@@ -2178,11 +2174,9 @@
    i64.sub
    local.set $4
    local.get $2
-   i64.const 0
+   i64.const 1
    local.get $4
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $2
   else
@@ -2212,11 +2206,9 @@
    i64.sub
    local.set $5
    local.get $3
-   i64.const 0
+   i64.const 1
    local.get $5
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $3
   else
diff --git a/tests/compiler/std/math.untouched.wat b/tests/compiler/std/math.untouched.wat
index 3ceaaf94f0..271d353465 100644
--- a/tests/compiler/std/math.untouched.wat
+++ b/tests/compiler/std/math.untouched.wat
@@ -9248,11 +9248,9 @@
    i64.sub
    local.set $4
    local.get $2
-   i64.const 0
+   i64.const 1
    local.get $4
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $2
   else
@@ -9282,11 +9280,9 @@
    i64.sub
    local.set $5
    local.get $3
-   i64.const 0
+   i64.const 1
    local.get $5
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $3
   else
@@ -9533,11 +9529,9 @@
    i32.sub
    local.set $4
    local.get $2
-   i32.const 0
+   i32.const 1
    local.get $4
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $2
   else
@@ -9565,11 +9559,9 @@
    i32.sub
    local.set $5
    local.get $3
-   i32.const 0
+   i32.const 1
    local.get $5
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $3
   else
@@ -11741,11 +11733,9 @@
    i64.sub
    local.set $4
    local.get $8
-   i64.const 0
+   i64.const 1
    local.get $4
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $8
   else
@@ -11775,11 +11765,9 @@
    i64.sub
    local.set $5
    local.get $3
-   i64.const 0
+   i64.const 1
    local.get $5
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $3
   else
@@ -12063,11 +12051,9 @@
    i32.sub
    local.set $4
    local.get $7
-   i32.const 0
+   i32.const 1
    local.get $4
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $7
   else
@@ -12095,11 +12081,9 @@
    i32.sub
    local.set $5
    local.get $3
-   i32.const 0
+   i32.const 1
    local.get $5
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $3
   else
diff --git a/tests/compiler/std/mod.untouched.wat b/tests/compiler/std/mod.untouched.wat
index d97129d1d2..5ef26eb943 100644
--- a/tests/compiler/std/mod.untouched.wat
+++ b/tests/compiler/std/mod.untouched.wat
@@ -122,11 +122,9 @@
    i64.sub
    local.set $4
    local.get $2
-   i64.const 0
+   i64.const 1
    local.get $4
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $2
   else
@@ -156,11 +154,9 @@
    i64.sub
    local.set $5
    local.get $3
-   i64.const 0
+   i64.const 1
    local.get $5
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $3
   else
@@ -431,11 +427,9 @@
    i32.sub
    local.set $4
    local.get $2
-   i32.const 0
+   i32.const 1
    local.get $4
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $2
   else
@@ -463,11 +457,9 @@
    i32.sub
    local.set $5
    local.get $3
-   i32.const 0
+   i32.const 1
    local.get $5
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $3
   else
diff --git a/tests/compiler/std/typedarray.untouched.wat b/tests/compiler/std/typedarray.untouched.wat
index 08a7fbbe07..329b76adb2 100644
--- a/tests/compiler/std/typedarray.untouched.wat
+++ b/tests/compiler/std/typedarray.untouched.wat
@@ -19120,11 +19120,9 @@
    i32.sub
    local.set $4
    local.get $2
-   i32.const 0
+   i32.const 1
    local.get $4
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $2
   else
@@ -19152,11 +19150,9 @@
    i32.sub
    local.set $5
    local.get $3
-   i32.const 0
+   i32.const 1
    local.get $5
    i32.sub
-   i32.const 1
-   i32.add
    i32.shl
    local.set $3
   else
@@ -19543,11 +19539,9 @@
    i64.sub
    local.set $4
    local.get $2
-   i64.const 0
+   i64.const 1
    local.get $4
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $2
   else
@@ -19577,11 +19571,9 @@
    i64.sub
    local.set $5
    local.get $3
-   i64.const 0
+   i64.const 1
    local.get $5
    i64.sub
-   i64.const 1
-   i64.add
    i64.shl
    local.set $3
   else