Skip to content

Instantly share code, notes, and snippets.

@kripken

kripken/fast-math-wasm.diff

Created October 8, 2025 20:08
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save kripken/407496f6bf1040618262c96c583d52f6 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save kripken/407496f6bf1040618262c96c583d52f6 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
fast-math-wasm.diff
--- a 2025-10-08 13:06:30.544955211 -0700
+++ b 2025-10-08 13:06:48.673083467 -0700
@@ -1290924,24 +1290924,21 @@
(i32.const 36)
)
)
)
)
)
(return
(local.get $0)
)
)
- (f64.mul
(local.get $0)
- (f64.const 1)
- )
)
(func $6309 (type $14) (param $0 i32) (result f32)
(local $1 i32)
(local $2 f32)
(local $scratch f32)
(local.set $0
(global.get $global$6)
)
(global.set $global$6
(i32.add
@@ -2477973,30 +2477970,27 @@
(local.tee $1
(i32.load
(i32.const 9469604)
)
)
)
)
)
(f64.store offset=40
(local.get $0)
- (f64.add
(call_indirect $timport$0 (type $39)
(local.get $1)
(i32.and
(local.get $2)
(i32.const 31)
)
)
- (f64.const -0)
- )
)
(global.set $global$6
(local.get $3)
)
)
(func $10040 (type $1) (param $0 i32) (param $1 i32)
(local $2 i32)
(local $3 i32)
(local $4 i32)
(local $5 i32)
@@ -2478960,30 +2478954,27 @@
(local.tee $1
(i32.load
(i32.const 9469604)
)
)
)
)
)
(f64.store offset=40
(local.get $0)
- (f64.add
(call_indirect $timport$0 (type $39)
(local.get $1)
(i32.and
(local.get $3)
(i32.const 31)
)
)
- (f64.const -0)
- )
)
(global.set $global$6
(local.get $2)
)
)
(func $10046 (type $2) (param $0 i32) (result i32)
(local $1 i32)
(local $2 i32)
(if
(i32.and
@@ -2993362,35 +2993353,29 @@
(i32.and
(local.tee $2
(i32.load offset=580
(local.get $1)
)
)
(i32.const 512)
)
(then
(local.set $4
- (f32.add
(f32.load offset=624
(local.get $1)
)
- (f32.const -0)
- )
)
(f32.store
(local.get $0)
- (f32.add
(f32.load offset=620
(local.get $1)
)
- (f32.const -0)
- )
)
(f32.store offset=4
(local.get $0)
(local.get $4)
)
(return)
)
)
(local.set $7
(f32.add
@@ -6908235,27 +6908220,24 @@
(f32.load offset=24
(local.get $0)
)
)
)
)
)
(local.tee $3
(f32.gt
(f32.abs
- (f32.sub
- (f32.const 0)
(f32.load offset=12
(local.get $0)
)
)
- )
(f32.add
(f32.load offset=28
(local.get $0)
)
(f32.const 0)
)
)
)
)
)
@@ -8022950,27 +8022932,24 @@
(f32.add
(local.get $8)
(f32.load offset=24
(local.get $0)
)
)
)
)
(f32.gt
(f32.abs
- (f32.sub
- (f32.const 0)
(f32.load offset=12
(local.get $0)
)
)
- )
(f32.add
(f32.load offset=28
(local.get $0)
)
(f32.const 0)
)
)
)
)
)
@@ -8023089,27 +8023068,24 @@
(f32.add
(local.get $8)
(f32.load offset=24
(local.get $0)
)
)
)
)
(f32.gt
(f32.abs
- (f32.sub
- (f32.const 0)
(f32.load offset=12
(local.get $0)
)
)
- )
(f32.add
(f32.load offset=28
(local.get $0)
)
(f32.const 0)
)
)
)
)
)
@@ -9150875,39 +9150851,36 @@
(local.get $1)
(i32.const 28)
)
)
(local.tee $2
(if (result i32)
(local.get $7)
(then
(i32.reinterpret_f32
(f32.div
- (f32.mul
(f32.add
(f32.sub
(local.tee $9
(f32.load
(local.get $2)
)
)
(local.get $10)
)
(f32.sub
(f32.load
(local.get $4)
)
(local.get $9)
)
)
- (f32.const 1)
- )
(local.get $12)
)
)
)
(else
(block $block12 (result i32)
(block $block11
(block $block10
(br_if $block10
(i32.eqz
@@ -9150957,35 +9150930,32 @@
)
)
(local.get $9)
)
)
)
)
)
(local.set $9
(f32.div
- (f32.mul
(f32.add
(local.tee $13
(f32.sub
(local.get $9)
(local.get $10)
)
)
(f32.sub
(local.get $11)
(local.get $9)
)
)
- (f32.const 1)
- )
(local.get $12)
)
)
(if
(f32.gt
(local.tee $10
(f32.abs
(f32.div
(local.get $13)
(f32.sub
@@ -10153288,43 +10153258,37 @@
(local.set $11
(f32.add
(f32.add
(f32.add
(local.tee $8
(f32.mul
(local.get $11)
(f32.const 0)
)
)
- (f32.mul
(local.get $12)
- (f32.const 1)
- )
)
(local.get $2)
)
(local.get $3)
)
)
(local.set $3
(f32.add
(f32.add
(local.tee $23
(f32.add
(local.get $8)
(local.get $5)
)
)
- (f32.mul
(local.get $13)
- (f32.const 1)
- )
)
(local.get $3)
)
)
(local.set $13
(f32.add
(f32.add
(f32.add
(local.tee $8
(f32.load offset=16
@@ -10153368,43 +10153332,37 @@
(local.set $8
(f32.add
(f32.add
(f32.add
(local.tee $9
(f32.mul
(local.get $8)
(f32.const 0)
)
)
- (f32.mul
(local.get $15)
- (f32.const 1)
- )
)
(local.get $12)
)
(local.get $5)
)
)
(local.set $5
(f32.add
(f32.add
(local.tee $25
(f32.add
(local.get $9)
(local.get $6)
)
)
- (f32.mul
(local.get $16)
- (f32.const 1)
- )
)
(local.get $5)
)
)
(local.set $16
(f32.add
(f32.add
(f32.add
(local.tee $9
(f32.load offset=32
@@ -10153448,43 +10153406,37 @@
(local.set $9
(f32.add
(f32.add
(f32.add
(local.tee $10
(f32.mul
(local.get $9)
(f32.const 0)
)
)
- (f32.mul
(local.get $17)
- (f32.const 1)
- )
)
(local.get $15)
)
(local.get $6)
)
)
(local.set $6
(f32.add
(f32.add
(local.tee $27
(f32.add
(local.get $10)
(local.get $7)
)
)
- (f32.mul
(local.get $18)
- (f32.const 1)
- )
)
(local.get $6)
)
)
(local.set $18
(f32.add
(f32.add
(f32.add
(local.tee $10
(f32.load offset=48
@@ -10153528,43 +10153480,37 @@
(local.set $10
(f32.add
(f32.add
(f32.add
(local.tee $22
(f32.mul
(local.get $10)
(f32.const 0)
)
)
- (f32.mul
(local.get $19)
- (f32.const 1)
- )
)
(local.get $17)
)
(local.get $7)
)
)
(local.set $7
(f32.add
(f32.add
(local.tee $29
(f32.add
(local.get $22)
(local.get $20)
)
)
- (f32.mul
(local.get $21)
- (f32.const 1)
- )
)
(local.get $7)
)
)
(local.set $19
(f32.add
(local.tee $30
(f32.mul
(local.get $4)
(f32.const 0)
@@ -10798451,41 +10798397,35 @@
)
(f32.const 0)
)
(f32.store offset=4
(local.get $4)
(f32.const 0)
)
(f32.store offset=8
(local.get $4)
(local.tee $13
- (f32.mul
(local.get $13)
- (f32.const 1)
- )
)
)
(f32.store offset=12
(local.get $4)
(f32.const 0)
)
(f32.store offset=16
(local.get $4)
(f32.const 0)
)
(f32.store offset=20
(local.get $4)
(local.tee $14
- (f32.mul
(local.get $14)
- (f32.const 1)
- )
)
)
(call $10212
(i32.add
(local.get $6)
(i32.const 136)
)
(local.get $4)
)
(i64.store offset=56 align=4
@@ -12443160,24 +12443100,21 @@
)
)
(else
(br_if $block
(i32.eq
(local.get $1)
(i32.const -1)
)
)
(return
- (f32.add
(local.get $5)
- (f32.const -0)
- )
)
)
)
)
)
(local.set $3
(i32.load offset=260
(local.get $0)
)
)
@@ -12510840,25 +12510777,22 @@
)
)
(f32.const 9.999999747378752e-05)
)
)
)
(br_if $block
(i32.eqz
(f32.le
(f32.abs
- (f32.sub
- (f32.const 0)
(local.get $36)
)
- )
(f32.const 9.999999747378752e-05)
)
)
)
(br $block1
(f32.const 1)
)
)
(block $block4
(block $block3
@@ -13598421,21 +13598355,20 @@
(i32.const 28)
)
)
)
)
)
)
(local.set $9
(local.tee $10
(f32.div
- (f32.mul
(f32.add
(f32.sub
(local.tee $9
(f32.load offset=8
(i32.add
(local.get $1)
(i32.mul
(local.get $2)
(i32.const 28)
)
@@ -13598458,22 +13598391,20 @@
(local.get $1)
(i32.mul
(local.get $0)
(i32.const 28)
)
)
)
(local.get $9)
)
)
- (f32.const 1)
- )
(f32.demote_f64
(select
(f64.const 9.999999747378752e-05)
(local.get $11)
(f64.le
(local.get $11)
(f64.const 9.999999747378752e-05)
)
)
)
@@ -13871073,55 +13871004,49 @@
)
)
(local.set $12
(if (result f32)
(i32.eq
(local.get $11)
(i32.const 5)
)
(then
(local.set $15
- (f32.mul
(call $83675
(f32.load offset=4
(local.get $3)
)
(local.get $12)
(f32.load offset=4
(local.get $4)
)
(local.get $13)
(f32.load offset=4
(local.get $6)
)
(local.get $14)
)
- (f32.const 1)
- )
)
- (f32.mul
(call $83675
(f32.load offset=8
(local.get $3)
)
(local.get $12)
(f32.load offset=8
(local.get $4)
)
(local.get $13)
(f32.load offset=8
(local.get $6)
)
(local.get $14)
)
- (f32.const 1)
- )
)
(else
(local.set $15
(f32.mul
(local.tee $12
(f32.div
(f32.const 1)
(f32.demote_f64
(select
(f64.const 9.999999747378752e-05)
@@ -13871134,46 +13871059,42 @@
)
)
(f64.le
(local.get $16)
(f64.const 9.999999747378752e-05)
)
)
)
)
)
- (f32.mul
(f32.add
(f32.sub
(local.tee $13
(f32.load offset=4
(local.get $4)
)
)
(f32.load offset=4
(local.get $3)
)
)
(f32.sub
(f32.load offset=4
(local.get $6)
)
(local.get $13)
)
)
- (f32.const 1)
- )
)
)
(f32.mul
(local.get $12)
- (f32.mul
(f32.add
(f32.sub
(local.tee $12
(f32.load offset=8
(i32.add
(local.get $1)
(i32.shl
(local.get $2)
(i32.const 5)
)
@@ -13871196,22 +13871117,20 @@
(local.get $1)
(i32.shl
(local.get $7)
(i32.const 5)
)
)
)
(local.get $12)
)
)
- (f32.const 1)
- )
)
)
)
)
(f32.store offset=12
(i32.add
(local.get $1)
(i32.shl
(local.get $2)
(i32.const 5)
@@ -13885715,254 +13885634,221 @@
(local $8 f32)
(local $9 f32)
(local $10 f32)
(local $11 f32)
(local $12 f64)
(if
(local.get $6)
(then
(f32.store
(local.get $7)
- (f32.mul
(call $83675
(f32.load
(local.get $1)
)
(local.get $0)
(f32.load
(local.get $3)
)
(local.get $2)
(f32.load
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
(f32.store offset=4
(local.get $7)
- (f32.mul
(call $83675
(f32.load offset=4
(local.get $1)
)
(local.get $0)
(f32.load offset=4
(local.get $3)
)
(local.get $2)
(f32.load offset=4
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
(f32.store offset=8
(local.get $7)
- (f32.mul
(call $83675
(f32.load offset=8
(local.get $1)
)
(local.get $0)
(f32.load offset=8
(local.get $3)
)
(local.get $2)
(f32.load offset=8
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
(f32.store offset=12
(local.get $7)
- (f32.mul
(call $83675
(f32.load offset=12
(local.get $1)
)
(local.get $0)
(f32.load offset=12
(local.get $3)
)
(local.get $2)
(f32.load offset=12
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
(f32.store offset=16
(local.get $7)
- (f32.mul
(call $83675
(f32.load offset=16
(local.get $1)
)
(local.get $0)
(f32.load offset=16
(local.get $3)
)
(local.get $2)
(f32.load offset=16
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
(f32.store offset=20
(local.get $7)
- (f32.mul
(call $83675
(f32.load offset=20
(local.get $1)
)
(local.get $0)
(f32.load offset=20
(local.get $3)
)
(local.get $2)
(f32.load offset=20
(local.get $5)
)
(local.get $4)
)
- (f32.const 1)
- )
)
)
(else
(local.set $2
- (f32.mul
(f32.add
(f32.sub
(local.tee $2
(f32.load offset=4
(local.get $3)
)
)
(f32.load offset=4
(local.get $1)
)
)
(f32.sub
(f32.load offset=4
(local.get $5)
)
(local.get $2)
)
)
- (f32.const 1)
- )
)
(local.set $8
- (f32.mul
(f32.add
(f32.sub
(local.tee $8
(f32.load offset=8
(local.get $3)
)
)
(f32.load offset=8
(local.get $1)
)
)
(f32.sub
(f32.load offset=8
(local.get $5)
)
(local.get $8)
)
)
- (f32.const 1)
- )
)
(local.set $9
- (f32.mul
(f32.add
(f32.sub
(local.tee $9
(f32.load offset=12
(local.get $3)
)
)
(f32.load offset=12
(local.get $1)
)
)
(f32.sub
(f32.load offset=12
(local.get $5)
)
(local.get $9)
)
)
- (f32.const 1)
- )
)
(local.set $10
- (f32.mul
(f32.add
(f32.sub
(local.tee $10
(f32.load offset=16
(local.get $3)
)
)
(f32.load offset=16
(local.get $1)
)
)
(f32.sub
(f32.load offset=16
(local.get $5)
)
(local.get $10)
)
)
- (f32.const 1)
- )
)
(local.set $11
- (f32.mul
(f32.add
(f32.sub
(local.tee $11
(f32.load offset=20
(local.get $3)
)
)
(f32.load offset=20
(local.get $1)
)
)
(f32.sub
(f32.load offset=20
(local.get $5)
)
(local.get $11)
)
)
- (f32.const 1)
- )
)
(f32.store
(local.get $7)
(f32.mul
(local.tee $0
(f32.div
(f32.const 1)
(f32.demote_f64
(select
(f64.const 9.999999747378752e-05)
@@ -13885975,41 +13885861,38 @@
)
)
(f64.le
(local.get $12)
(f64.const 9.999999747378752e-05)
)
)
)
)
)
- (f32.mul
(f32.add
(f32.sub
(local.tee $4
(f32.load
(local.get $3)
)
)
(f32.load
(local.get $1)
)
)
(f32.sub
(f32.load
(local.get $5)
)
(local.get $4)
)
)
- (f32.const 1)
- )
)
)
(f32.store offset=4
(local.get $7)
(f32.mul
(local.get $0)
(local.get $2)
)
)
(f32.store offset=8
@@ -16076952,27 +16076835,24 @@
(f32.load offset=24
(local.get $0)
)
)
)
)
)
(local.tee $1
(f32.gt
(f32.abs
- (f32.sub
- (f32.const 0)
(f32.load offset=12
(local.get $0)
)
)
- )
(f32.add
(f32.load offset=28
(local.get $0)
)
(f32.const 0)
)
)
)
)
)
@@ -20109854,27 +20109734,24 @@
(f32.load offset=24
(local.get $0)
)
)
)
)
)
(local.tee $3
(f32.gt
(f32.abs
- (f32.sub
- (f32.const 0)
(f32.load offset=12
(local.get $0)
)
)
- )
(f32.add
(f32.load offset=28
(local.get $0)
)
(f32.const 0)
)
)
)
)
)
@@ -22420788,24 +22420665,21 @@
(f32.mul
(local.get $7)
(f32.const 0)
)
(local.get $9)
)
)
(f32.store offset=24
(local.get $2)
(f32.add
- (f32.mul
(local.get $7)
- (f32.const 1)
- )
(local.get $11)
)
)
)
)
(global.set $global$6
(local.get $3)
)
(i32.const 1)
)
@@ -22514044,24 +22513918,21 @@
(f32.const 0.5)
)
)
(local.tee $19
(f32.mul
(f32.sub
(f32.add
(local.get $19)
(f32.const 0)
)
- (f32.add
(local.get $18)
- (f32.const -0)
- )
)
(local.tee $3
(f32.mul
(local.get $3)
(f32.const 0.5)
)
)
)
)
)
@@ -22514070,45 +22513941,39 @@
(local.get $0)
(f32.sub
(local.get $25)
(local.tee $17
(f32.mul
(f32.sub
(f32.add
(local.get $20)
(f32.const 0)
)
- (f32.add
(local.get $17)
- (f32.const -0)
- )
)
(local.get $3)
)
)
)
)
(f32.store offset=8
(local.get $0)
(f32.sub
(local.get $23)
(local.tee $3
(f32.mul
(f32.sub
(f32.add
(local.get $22)
(f32.const 0)
)
- (f32.add
(local.get $16)
- (f32.const -0)
- )
)
(local.get $3)
)
)
)
)
(f32.store offset=12
(local.get $0)
(f32.add
(local.get $21)
@@ -22514789,24 +22514654,21 @@
(i32.const -25165825)
)
(local.set $7
(i32.const 2122317823)
)
(local.set $8
(f32.const 8507058665963221495292604e13)
)
(i32.reinterpret_f32
(local.tee $9
- (f32.add
(local.get $10)
- (f32.const -0)
- )
)
)
)
)
)
(local.set $10
(f32.const 8507058665963221495292604e13)
)
(local.set $12
(f32.const -8507058665963221495292604e13)
@@ -22514881,24 +22514743,21 @@
)
(local.set $4
(i32.const 2122317823)
)
(local.set $5
(i32.const -25165825)
)
(local.set $6
(i32.reinterpret_f32
(local.tee $11
- (f32.add
(local.get $10)
- (f32.const -0)
- )
)
)
)
(local.set $7
(i32.const 2122317823)
)
(local.set $9
(f32.const -8507058665963221495292604e13)
)
(local.set $10
@@ -22514979,24 +22514838,21 @@
(else
(local.set $3
(i32.const 2122317823)
)
(local.set $4
(i32.const 2122317823)
)
(local.set $5
(i32.reinterpret_f32
(local.tee $12
- (f32.add
(local.get $10)
- (f32.const -0)
- )
)
)
)
(local.set $6
(i32.const -25165825)
)
(local.set $7
(i32.const 2122317823)
)
(local.set $9
@@ -23742882,26 +23742738,23 @@
(local.tee $9
(i32.sub
(local.get $5)
(i32.const 1)
)
)
)
(then
(f32.store
(local.get $7)
- (f32.mul
(f32.load
(local.get $0)
)
- (f32.const 1)
- )
)
(return)
)
)
)
)
(local.set $2
(i32.const 0)
)
(loop $label8
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment