diff options
| -rw-r--r-- | apps/codecs/demac/libdemac/udiv32_arm.S | 10 | ||||
| -rw-r--r-- | firmware/target/arm/support-arm.S | 4 |
2 files changed, 9 insertions, 5 deletions
diff --git a/apps/codecs/demac/libdemac/udiv32_arm.S b/apps/codecs/demac/libdemac/udiv32_arm.S index 939fce1..10c0731 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 dd17e0f..b4577f6 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 |