diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0e3e3ab..8f91a73 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -29,6 +29,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
   x87 stack.
 
 ### Fixed
+- The x87 native filter unit was denormalizing and eating up a lot of CPU ([#68][i68])
 - Modulating delaytime in wasm could crash, because delay time was converted to
   int with i32.trunc_f32_u. Using i32.trunc_f32_s fixed this.
 - When recording notes from VSTI, no track was created for instruments that had
@@ -153,6 +154,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 [0.2.0]: https://github.com/vsariola/sointu/compare/v0.1.0...v0.2.0
 [0.1.0]: https://github.com/vsariola/sointu/compare/4klang-3.11...v0.1.0
 [i65]: https://github.com/vsariola/sointu/issues/65
+[i68]: https://github.com/vsariola/sointu/issues/68
 [i112]: https://github.com/vsariola/sointu/issues/112
 [i116]: https://github.com/vsariola/sointu/issues/116
 [i120]: https://github.com/vsariola/sointu/issues/120
diff --git a/vm/compiler/templates/amd64-386/effects.asm b/vm/compiler/templates/amd64-386/effects.asm
index 6d7d371..747331f 100644
--- a/vm/compiler/templates/amd64-386/effects.asm
+++ b/vm/compiler/templates/amd64-386/effects.asm
@@ -156,6 +156,9 @@ su_op_dbgain_mono:
     fsubp   st2, st0                        ; r x-l'
     fmul    dword [{{.WRK}}+8]  ; r*b x-l'
     fsubp   st1, st0                        ; x-l'-r*b
+    {{- .Float 0.5 | .Prepare | indent 4}}
+    fadd    dword [{{.Float 0.5 | .Use}}]           ; add and sub small offset to prevent denormalization
+    fsub    dword [{{.Float 0.5 | .Use}}]           ; See for example: https://stackoverflow.com/questions/36781881/why-denormalized-floats-are-so-much-slower-than-other-floats-from-hardware-arch
     fst     dword [{{.WRK}}+4]  ; h'=x-l'-r*b
     fmul    dword [{{.WRK}}+12]                   ; f2*h'
     fadd    dword [{{.WRK}}+8]  ; f2*h'+b