The new method of doing comparisons did not work all of the time: if

the low bytes were equal, then the saturating subtraction result in
that byte would be 0 with the high byte containing a non-zero value.
To deal with this, shift and or together the two values into the low
byte.

Add some new tests which check out the SIMD implementation more
carefully, including values that trigger the problem in the earlier
version.

1 files changed, 3 insertions, 1 deletions

diff --git a/src/block/idea_sse2/idea_sse2.cpp b/src/block/idea_sse2/idea_sse2.cpp
index 81b0fd9c1..70698560d 100644
--- a/src/block/idea_sse2/idea_sse2.cpp
+++ b/src/block/idea_sse2/idea_sse2.cpp
@@ -28,7 +28,9 @@ inline __m128i mul(__m128i X, u16bit K_16)
    __m128i T = _mm_sub_epi16(mul_lo, mul_hi);
 
    // Unsigned compare; cmp = 1 if mul_lo < mul_hi else 0
-   const __m128i cmp = _mm_min_epu8(ones, _mm_subs_epu16(mul_hi, mul_lo));
+   const __m128i subs = _mm_subs_epu16(mul_hi, mul_lo);
+   const __m128i cmp = _mm_min_epu8(
+     _mm_or_si128(subs, _mm_srli_epi16(subs, 8)), ones);
 
    T = _mm_add_epi16(T, cmp);