diff --git a/silk/Inlines.h b/silk/Inlines.h index 50060e63f18ab5d5dc69fd57a6b25aefdd7937a9..765e77083f2aaaff07bf3c80ea959b4c47991bb0 100644 --- a/silk/Inlines.h +++ b/silk/Inlines.h @@ -117,7 +117,8 @@ static inline opus_int32 silk_DIV32_varQ( /* O returns a good approximatio result = silk_SMULWB(a32_nrm, b32_inv); /* Q: 29 + a_headrm - b_headrm */ /* Compute residual by subtracting product of denominator and first approximation */ - a32_nrm -= silk_LSHIFT( silk_SMMUL(b32_nrm, result), 3 ); /* Q: a_headrm */ + /* It's OK to overflow because the final value of a32_nrm should always be small */ + a32_nrm = silk_SUB32_ovflw(a32_nrm, silk_LSHIFT_ovflw( silk_SMMUL(b32_nrm, result), 3 )); /* Q: a_headrm */ /* Refinement */ result = silk_SMLAWB(result, a32_nrm, b32_inv); /* Q: 29 + a_headrm - b_headrm */ diff --git a/silk/SigProc_FIX.h b/silk/SigProc_FIX.h index a65ac13f39e5c25168205660866686a7af810c38..f9e2664aea9faa21d69be8151cb5345c1973dbd9 100644 --- a/silk/SigProc_FIX.h +++ b/silk/SigProc_FIX.h @@ -417,6 +417,9 @@ static inline opus_int32 silk_ROR32( opus_int32 a32, opus_int rot ) /* Adds two signed 32-bit values in a way that can overflow, while not relying on undefined behaviour (just standard two's complement implementation-specific behaviour) */ #define silk_ADD32_ovflw(a, b) ((opus_int32)((opus_uint32)(a) + (opus_uint32)(b))) +/* Subtractss two signed 32-bit values in a way that can overflow, while not relying on undefined behaviour + (just standard two's complement implementation-specific behaviour) */ +#define silk_SUB32_ovflw(a, b) ((opus_int32)((opus_uint32)(a) - (opus_uint32)(b))) /* Multiply-accumulate macros that allow overflow in the addition (ie, no asserts in debug mode)*/ #define silk_MLA_ovflw(a32, b32, c32) silk_ADD32_ovflw((a32), (opus_uint32)(b32) * (opus_uint32)(c32)) @@ -484,6 +487,7 @@ static inline opus_int32 silk_ROR32( opus_int32 a32, opus_int rot ) #define silk_LSHIFT_SAT32(a, shift) (silk_LSHIFT32( silk_LIMIT( (a), silk_RSHIFT32( silk_int32_MIN, (shift) ), \ silk_RSHIFT32( silk_int32_MAX, (shift) ) ), (shift) )) +#define silk_LSHIFT_ovflw(a, shift) ((opus_int32)((opus_uint32)(a)<<(shift))) /* shift >= 0, allowed to overflow */ #define silk_LSHIFT_uint(a, shift) ((a)<<(shift)) /* shift >= 0 */ #define silk_RSHIFT_uint(a, shift) ((a)>>(shift)) /* shift >= 0 */