diff --git a/std/assembly/internal/itoa.ts b/std/assembly/internal/itoa.ts
index f3be8b7b..0f5d36f3 100644
--- a/std/assembly/internal/itoa.ts
+++ b/std/assembly/internal/itoa.ts
@@ -67,23 +67,27 @@ function DIGITS(): u32[] {
   return table; // inlines to a constant memory offset
 }
 
-// Count number of decimals in value
-export function decimalCount<T>(value: T): i32 {
-  var v = abs<T>(value); // NOP if value is unsigned anyway
-  var l: usize = 8 * sizeof<T>() - <usize>clz<T>(v | 10); // log2
-  var t = l * 1233 >>> 12;                                // log10
+// Count number of decimals for u32 values
+// In our case input value always non-zero so we can simplify some parts
+function decimalCountU32(value: u32): i32 {
+  var l: usize = 32 - <usize>clz<u32>(value); // log2
+  var t = l * 1233 >>> 12; // log10
 
   var lutbuf = <ArrayBuffer>POWERS10().buffer_;
-  if (sizeof<T>() <= 4) {
-    let power  = loadUnsafe<u32,T>(lutbuf, t);
-    t -= <usize>(v < power);
-  } else { // sizeof<T>() == 8
-    let le10   = t <= 10;
-    let offset = select<usize>(0,        10, le10); // offset = t <= 10 ? 0 : 10
-    let factor = select< T >(1, 10000000000, le10); // factor = t <= 10 ? 1 : 10 ^ 10
-    let power  = loadUnsafe<u32,T>(lutbuf, t - offset);
-    t -= <usize>(v < factor * power);
-  }
+  var power  = loadUnsafe<u32,u32>(lutbuf, t);
+  t -= <usize>(value < power);
+  return t + 1;
+}
+
+// Count number of decimals for u64 values
+// In our case input value always greater than 2^32-1 so we can skip some parts
+function decimalCountU64(value: u64): i32 {
+  var l: usize = 64 - <usize>clz<u64>(value); // log2
+  var t = l * 1233 >>> 12; // log10
+
+  var lutbuf = <ArrayBuffer>POWERS10().buffer_;
+  var power  = loadUnsafe<u32,u64>(lutbuf, t - 10);
+  t -= <usize>(value < 10000000000 * power);
   return t + 1;
 }
 
@@ -189,7 +193,7 @@ export function utoa64_core(buffer: usize, num: u64, offset: u32): void {
 export function utoa32(value: u32): string {
   if (!value) return "0";
 
-  var decimals = decimalCount<u32>(value);
+  var decimals = decimalCountU32(value);
   var buffer   = allocateString(decimals);
 
   utoa32_core(changetype<usize>(buffer), value, decimals);
@@ -202,7 +206,7 @@ export function itoa32(value: i32): string {
   var isneg = value < 0;
   if (isneg) value = -value;
 
-  var decimals = decimalCount<u32>(value) + <i32>isneg;
+  var decimals = decimalCountU32(value) + <i32>isneg;
   var buffer   = allocateString(decimals);
 
   utoa32_core(changetype<usize>(buffer), value, decimals);
@@ -217,11 +221,11 @@ export function utoa64(value: u64): string {
   var buffer: String;
   if (value <= u32.MAX_VALUE) {
     let value32  = <u32>value;
-    let decimals = decimalCount<u32>(value32);
+    let decimals = decimalCountU32(value32);
     buffer = allocateString(decimals);
     utoa32_core(changetype<usize>(buffer), value32, decimals);
   } else {
-    let decimals = decimalCount<u64>(value);
+    let decimals = decimalCountU64(value);
     buffer = allocateString(decimals);
     utoa64_core(changetype<usize>(buffer), value, decimals);
   }
@@ -238,11 +242,11 @@ export function itoa64(value: i64): string {
   var buffer: String;
   if (<u64>value <= <u64>u32.MAX_VALUE) {
     let value32  = <u32>value;
-    let decimals = decimalCount<u32>(value32) + <i32>isneg;
+    let decimals = decimalCountU32(value32) + <i32>isneg;
     buffer = allocateString(decimals);
     utoa32_core(changetype<usize>(buffer), value32, decimals);
   } else {
-    let decimals = decimalCount<u64>(value) + <i32>isneg;
+    let decimals = decimalCountU64(value) + <i32>isneg;
     buffer = allocateString(decimals);
     utoa64_core(changetype<usize>(buffer), value, decimals);
   }
diff --git a/tests/compiler/std/string.optimized.wat b/tests/compiler/std/string.optimized.wat
index eadfbcbb..db2a6fe7 100644
--- a/tests/compiler/std/string.optimized.wat
+++ b/tests/compiler/std/string.optimized.wat
@@ -3734,7 +3734,7 @@
   )
   (get_local $4)
  )
- (func $~lib/internal/itoa/decimalCount<u32> (; 27 ;) (type $ii) (param $0 i32) (result i32)
+ (func $~lib/internal/itoa/decimalCountU32 (; 27 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (set_local $2
@@ -3750,10 +3750,7 @@
          (i32.sub
           (i32.const 32)
           (i32.clz
-           (i32.or
-            (get_local $0)
-            (i32.const 10)
-           )
+           (get_local $0)
           )
          )
          (i32.const 1233)
@@ -3969,7 +3966,7 @@
     (call $~lib/internal/string/allocate
      (tee_local $3
       (i32.add
-       (call $~lib/internal/itoa/decimalCount<u32>
+       (call $~lib/internal/itoa/decimalCountU32
         (get_local $0)
        )
        (get_local $1)
@@ -4004,7 +4001,7 @@
    (tee_local $2
     (call $~lib/internal/string/allocate
      (tee_local $1
-      (call $~lib/internal/itoa/decimalCount<u32>
+      (call $~lib/internal/itoa/decimalCountU32
        (get_local $0)
       )
      )
@@ -4015,31 +4012,38 @@
   )
   (get_local $2)
  )
- (func $~lib/internal/itoa/decimalCount<u64> (; 31 ;) (type $Ii) (param $0 i64) (result i32)
+ (func $~lib/internal/itoa/decimalCountU64 (; 31 ;) (type $Ii) (param $0 i64) (result i32)
   (local $1 i32)
-  (local $2 i32)
+  (local $2 i64)
   (set_local $2
-   (i32.le_u
-    (tee_local $1
-     (i32.shr_u
-      (i32.mul
-       (i32.sub
-        (i32.const 64)
-        (i32.wrap/i64
-         (i64.clz
-          (i64.or
-           (get_local $0)
-           (i64.const 10)
+   (i64.load32_u offset=8
+    (i32.add
+     (i32.load
+      (i32.const 1400)
+     )
+     (i32.shl
+      (i32.sub
+       (tee_local $1
+        (i32.shr_u
+         (i32.mul
+          (i32.sub
+           (i32.const 64)
+           (i32.wrap/i64
+            (i64.clz
+             (get_local $0)
+            )
+           )
           )
+          (i32.const 1233)
          )
+         (i32.const 12)
         )
        )
-       (i32.const 1233)
+       (i32.const 10)
       )
-      (i32.const 12)
+      (i32.const 2)
      )
     )
-    (i32.const 10)
    )
   )
   (i32.add
@@ -4048,29 +4052,8 @@
     (i64.lt_u
      (get_local $0)
      (i64.mul
-      (select
-       (i64.const 1)
-       (i64.const 10000000000)
-       (get_local $2)
-      )
-      (i64.load32_u offset=8
-       (i32.add
-        (i32.load
-         (i32.const 1400)
-        )
-        (i32.shl
-         (i32.sub
-          (get_local $1)
-          (select
-           (i32.const 0)
-           (i32.const 10)
-           (get_local $2)
-          )
-         )
-         (i32.const 2)
-        )
-       )
-      )
+      (get_local $2)
+      (i64.const 10000000000)
      )
     )
    )
@@ -4244,7 +4227,7 @@
     (tee_local $2
      (call $~lib/internal/string/allocate
       (tee_local $1
-       (call $~lib/internal/itoa/decimalCount<u32>
+       (call $~lib/internal/itoa/decimalCountU32
         (tee_local $3
          (i32.wrap/i64
           (get_local $0)
@@ -4261,7 +4244,7 @@
     (tee_local $2
      (call $~lib/internal/string/allocate
       (tee_local $1
-       (call $~lib/internal/itoa/decimalCount<u64>
+       (call $~lib/internal/itoa/decimalCountU64
         (get_local $0)
        )
       )
@@ -4310,7 +4293,7 @@
      (call $~lib/internal/string/allocate
       (tee_local $2
        (i32.add
-        (call $~lib/internal/itoa/decimalCount<u32>
+        (call $~lib/internal/itoa/decimalCountU32
          (tee_local $4
           (i32.wrap/i64
            (get_local $0)
@@ -4330,7 +4313,7 @@
      (call $~lib/internal/string/allocate
       (tee_local $2
        (i32.add
-        (call $~lib/internal/itoa/decimalCount<u64>
+        (call $~lib/internal/itoa/decimalCountU64
          (get_local $0)
         )
         (get_local $1)
diff --git a/tests/compiler/std/string.untouched.wat b/tests/compiler/std/string.untouched.wat
index de70dfc8..b34550ed 100644
--- a/tests/compiler/std/string.untouched.wat
+++ b/tests/compiler/std/string.untouched.wat
@@ -4448,69 +4448,59 @@
   )
   (get_local $4)
  )
- (func $~lib/internal/itoa/decimalCount<u32> (; 27 ;) (type $ii) (param $0 i32) (result i32)
+ (func $~lib/internal/itoa/decimalCountU32 (; 27 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
   (set_local $1
-   (get_local $0)
-  )
-  (set_local $2
    (i32.sub
-    (i32.mul
-     (i32.const 8)
-     (i32.const 4)
-    )
+    (i32.const 32)
     (i32.clz
-     (i32.or
-      (get_local $1)
-      (i32.const 10)
-     )
+     (get_local $0)
     )
    )
   )
-  (set_local $3
+  (set_local $2
    (i32.shr_u
     (i32.mul
-     (get_local $2)
+     (get_local $1)
      (i32.const 1233)
     )
     (i32.const 12)
    )
   )
-  (set_local $4
+  (set_local $3
    (i32.load
     (block $~lib/internal/itoa/POWERS10|inlined.0 (result i32)
      (i32.const 552)
     )
    )
   )
-  (set_local $5
+  (set_local $4
    (block $~lib/internal/arraybuffer/loadUnsafe<u32,u32>|inlined.0 (result i32)
     (i32.load offset=8
      (i32.add
-      (get_local $4)
+      (get_local $3)
       (i32.shl
-       (get_local $3)
+       (get_local $2)
        (i32.const 2)
       )
      )
     )
    )
   )
-  (set_local $3
+  (set_local $2
    (i32.sub
-    (get_local $3)
+    (get_local $2)
     (i32.lt_u
-     (get_local $1)
-     (get_local $5)
+     (get_local $0)
+     (get_local $4)
     )
    )
   )
   (i32.add
-   (get_local $3)
+   (get_local $2)
    (i32.const 1)
   )
  )
@@ -4762,7 +4752,7 @@
   )
   (set_local $2
    (i32.add
-    (call $~lib/internal/itoa/decimalCount<u32>
+    (call $~lib/internal/itoa/decimalCountU32
      (get_local $0)
     )
     (get_local $1)
@@ -4801,7 +4791,7 @@
    )
   )
   (set_local $1
-   (call $~lib/internal/itoa/decimalCount<u32>
+   (call $~lib/internal/itoa/decimalCountU32
     (get_local $0)
    )
   )
@@ -4819,45 +4809,32 @@
   )
   (get_local $2)
  )
- (func $~lib/internal/itoa/decimalCount<u64> (; 31 ;) (type $Ii) (param $0 i64) (result i32)
-  (local $1 i64)
+ (func $~lib/internal/itoa/decimalCountU64 (; 31 ;) (type $Ii) (param $0 i64) (result i32)
+  (local $1 i32)
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (local $6 i32)
-  (local $7 i64)
-  (local $8 i32)
-  (local $9 i64)
+  (local $5 i64)
   (set_local $1
-   (get_local $0)
-  )
-  (set_local $2
    (i32.sub
-    (i32.mul
-     (i32.const 8)
-     (i32.const 8)
-    )
+    (i32.const 64)
     (i32.wrap/i64
      (i64.clz
-      (i64.or
-       (get_local $1)
-       (i64.const 10)
-      )
+      (get_local $0)
      )
     )
    )
   )
-  (set_local $3
+  (set_local $2
    (i32.shr_u
     (i32.mul
-     (get_local $2)
+     (get_local $1)
      (i32.const 1233)
     )
     (i32.const 12)
    )
   )
-  (set_local $4
+  (set_local $3
    (i32.load
     (block $~lib/internal/itoa/POWERS10|inlined.1 (result i32)
      (i32.const 1400)
@@ -4865,58 +4842,38 @@
    )
   )
   (set_local $5
-   (i32.le_u
-    (get_local $3)
-    (i32.const 10)
-   )
-  )
-  (set_local $6
-   (select
-    (i32.const 0)
-    (i32.const 10)
-    (get_local $5)
-   )
-  )
-  (set_local $7
-   (select
-    (i64.const 1)
-    (i64.const 10000000000)
-    (get_local $5)
-   )
-  )
-  (set_local $9
    (block $~lib/internal/arraybuffer/loadUnsafe<u32,u64>|inlined.2 (result i64)
-    (set_local $8
+    (set_local $4
      (i32.sub
-      (get_local $3)
-      (get_local $6)
+      (get_local $2)
+      (i32.const 10)
      )
     )
     (i64.load32_u offset=8
      (i32.add
-      (get_local $4)
+      (get_local $3)
       (i32.shl
-       (get_local $8)
+       (get_local $4)
        (i32.const 2)
       )
      )
     )
    )
   )
-  (set_local $3
+  (set_local $2
    (i32.sub
-    (get_local $3)
+    (get_local $2)
     (i64.lt_u
-     (get_local $1)
+     (get_local $0)
      (i64.mul
-      (get_local $7)
-      (get_local $9)
+      (i64.const 10000000000)
+      (get_local $5)
      )
     )
    )
   )
   (i32.add
-   (get_local $3)
+   (get_local $2)
    (i32.const 1)
   )
  )
@@ -5140,7 +5097,7 @@
      )
     )
     (set_local $3
-     (call $~lib/internal/itoa/decimalCount<u32>
+     (call $~lib/internal/itoa/decimalCountU32
       (get_local $2)
      )
     )
@@ -5159,7 +5116,7 @@
    )
    (block
     (set_local $3
-     (call $~lib/internal/itoa/decimalCount<u64>
+     (call $~lib/internal/itoa/decimalCountU64
       (get_local $0)
      )
     )
@@ -5220,7 +5177,7 @@
     )
     (set_local $4
      (i32.add
-      (call $~lib/internal/itoa/decimalCount<u32>
+      (call $~lib/internal/itoa/decimalCountU32
        (get_local $3)
       )
       (get_local $1)
@@ -5242,7 +5199,7 @@
    (block
     (set_local $4
      (i32.add
-      (call $~lib/internal/itoa/decimalCount<u64>
+      (call $~lib/internal/itoa/decimalCountU64
        (get_local $0)
       )
       (get_local $1)