diff --git a/apps/codecs/demac/libdemac/udiv32_arm.S b/apps/codecs/demac/libdemac/udiv32_arm.S
index 939fce17d4..10c0731db1 100644
--- a/apps/codecs/demac/libdemac/udiv32_arm.S
+++ b/apps/codecs/demac/libdemac/udiv32_arm.S
@@ -234,10 +234,12 @@ udiv32_arm:
     mul     \inv, \divisor, \neg
     smlawt  \divisor, \divisor, \inv, \divisor
     mul     \inv, \divisor, \neg
-    /* This will save a cycle on ARMv6, but does not produce a correct result
-       if numerator sign bit is set. This case accounts for about 1 in 10^7 of
-       divisions, done by the APE decoder, so we specialize for the more common
-       case and handle the uncommon large-numerator separately */
+    /* This will save a cycle on ARMv6, but requires that the numerator sign
+       bit is not set (that of inv is guaranteed unset). The branch should
+       predict very well, making it typically 1 cycle, and thus both the branch
+       and test fill delay cycles for the multiplies. Based on logging of
+       numerator sizes in the APE codec, the branch is taken about 1/10^7 of
+       the time. */
 #if ARM_ARCH >= 6
     tst     \numerator, \numerator
     smmla   \divisor, \divisor, \inv, \divisor
diff --git a/firmware/target/arm/support-arm.S b/firmware/target/arm/support-arm.S
index dd17e0f5f7..b4577f6e9b 100644
--- a/firmware/target/arm/support-arm.S
+++ b/firmware/target/arm/support-arm.S
@@ -244,7 +244,9 @@
     tst     \numerator, \numerator
     smmla   \divisor, \divisor, \inv, \divisor
     /* Branch to large-numerator handler, or else use smmul if sign bit is not
-       set. */
+       set. This wins on average with random numerators, and should be no
+       slower than using umull for small numerator, even if prediction fails.
+    */
     bmi     40f
     smmul   \inv, \numerator, \divisor
 #else