diff options
Diffstat (limited to 'vendor/compiler_builtins/src/int/specialized_div_rem/binary_long.rs')
-rw-r--r-- | vendor/compiler_builtins/src/int/specialized_div_rem/binary_long.rs | 548 |
1 files changed, 548 insertions, 0 deletions
diff --git a/vendor/compiler_builtins/src/int/specialized_div_rem/binary_long.rs b/vendor/compiler_builtins/src/int/specialized_div_rem/binary_long.rs new file mode 100644 index 000000000..0d7822882 --- /dev/null +++ b/vendor/compiler_builtins/src/int/specialized_div_rem/binary_long.rs @@ -0,0 +1,548 @@ +/// Creates an unsigned division function that uses binary long division, designed for +/// computer architectures without division instructions. These functions have good performance for +/// microarchitectures with large branch miss penalties and architectures without the ability to +/// predicate instructions. For architectures with predicated instructions, one of the algorithms +/// described in the documentation of these functions probably has higher performance, and a custom +/// assembly routine should be used instead. +#[allow(unused_macros)] +macro_rules! impl_binary_long { + ( + $fn:ident, // name of the unsigned division function + $zero_div_fn:ident, // function called when division by zero is attempted + $normalization_shift:ident, // function for finding the normalization shift + $n:tt, // the number of bits in a $iX or $uX + $uX:ident, // unsigned integer type for the inputs and outputs of `$fn` + $iX:ident // signed integer type with same bitwidth as `$uX` + ) => { + /// Computes the quotient and remainder of `duo` divided by `div` and returns them as a + /// tuple. + pub fn $fn(duo: $uX, div: $uX) -> ($uX, $uX) { + let mut duo = duo; + // handle edge cases before calling `$normalization_shift` + if div == 0 { + $zero_div_fn() + } + if duo < div { + return (0, duo); + } + + // There are many variations of binary division algorithm that could be used. This + // documentation gives a tour of different methods so that future readers wanting to + // optimize further do not have to painstakingly derive them. The SWAR variation is + // especially hard to understand without reading the less convoluted methods first. + + // You may notice that a `duo < div_original` check is included in many these + // algorithms. A critical optimization that many algorithms miss is handling of + // quotients that will turn out to have many trailing zeros or many leading zeros. This + // happens in cases of exact or close-to-exact divisions, divisions by power of two, and + // in cases where the quotient is small. The `duo < div_original` check handles these + // cases of early returns and ends up replacing other kinds of mundane checks that + // normally terminate a binary division algorithm. + // + // Something you may see in other algorithms that is not special-cased here is checks + // for division by powers of two. The `duo < div_original` check handles this case and + // more, however it can be checked up front before the bisection using the + // `((div > 0) && ((div & (div - 1)) == 0))` trick. This is not special-cased because + // compilers should handle most cases where divisions by power of two occur, and we do + // not want to add on a few cycles for every division operation just to save a few + // cycles rarely. + + // The following example is the most straightforward translation from the way binary + // long division is typically visualized: + // Dividing 178u8 (0b10110010) by 6u8 (0b110). `div` is shifted left by 5, according to + // the result from `$normalization_shift(duo, div, false)`. + // + // Step 0: `sub` is negative, so there is not full normalization, so no `quo` bit is set + // and `duo` is kept unchanged. + // duo:10110010, div_shifted:11000000, sub:11110010, quo:00000000, shl:5 + // + // Step 1: `sub` is positive, set a `quo` bit and update `duo` for next step. + // duo:10110010, div_shifted:01100000, sub:01010010, quo:00010000, shl:4 + // + // Step 2: Continue based on `sub`. The `quo` bits start accumulating. + // duo:01010010, div_shifted:00110000, sub:00100010, quo:00011000, shl:3 + // duo:00100010, div_shifted:00011000, sub:00001010, quo:00011100, shl:2 + // duo:00001010, div_shifted:00001100, sub:11111110, quo:00011100, shl:1 + // duo:00001010, div_shifted:00000110, sub:00000100, quo:00011100, shl:0 + // The `duo < div_original` check terminates the algorithm with the correct quotient of + // 29u8 and remainder of 4u8 + /* + let div_original = div; + let mut shl = $normalization_shift(duo, div, false); + let mut quo = 0; + loop { + let div_shifted = div << shl; + let sub = duo.wrapping_sub(div_shifted); + // it is recommended to use `println!`s like this if functionality is unclear + /* + println!("duo:{:08b}, div_shifted:{:08b}, sub:{:08b}, quo:{:08b}, shl:{}", + duo, + div_shifted, + sub, + quo, + shl + ); + */ + if 0 <= (sub as $iX) { + duo = sub; + quo += 1 << shl; + if duo < div_original { + // this branch is optional + return (quo, duo) + } + } + if shl == 0 { + return (quo, duo) + } + shl -= 1; + } + */ + + // This restoring binary long division algorithm reduces the number of operations + // overall via: + // - `pow` can be shifted right instead of recalculating from `shl` + // - starting `div` shifted left and shifting it right for each step instead of + // recalculating from `shl` + // - The `duo < div_original` branch is used to terminate the algorithm instead of the + // `shl == 0` branch. This check is strong enough to prevent set bits of `pow` and + // `div` from being shifted off the end. This check also only occurs on half of steps + // on average, since it is behind the `(sub as $iX) >= 0` branch. + // - `shl` is now not needed by any aspect of of the loop and thus only 3 variables are + // being updated between steps + // + // There are many variations of this algorithm, but this encompases the largest number + // of architectures and does not rely on carry flags, add-with-carry, or SWAR + // complications to be decently fast. + /* + let div_original = div; + let shl = $normalization_shift(duo, div, false); + let mut div: $uX = div << shl; + let mut pow: $uX = 1 << shl; + let mut quo: $uX = 0; + loop { + let sub = duo.wrapping_sub(div); + if 0 <= (sub as $iX) { + duo = sub; + quo |= pow; + if duo < div_original { + return (quo, duo) + } + } + div >>= 1; + pow >>= 1; + } + */ + + // If the architecture has flags and predicated arithmetic instructions, it is possible + // to do binary long division without branching and in only 3 or 4 instructions. This is + // a variation of a 3 instruction central loop from + // http://www.chiark.greenend.org.uk/~theom/riscos/docs/ultimate/a252div.txt. + // + // What allows doing division in only 3 instructions is realizing that instead of + // keeping `duo` in place and shifting `div` right to align bits, `div` can be kept in + // place and `duo` can be shifted left. This means `div` does not have to be updated, + // but causes edge case problems and makes `duo < div_original` tests harder. Some + // architectures have an option to shift an argument in an arithmetic operation, which + // means `duo` can be shifted left and subtracted from in one instruction. The other two + // instructions are updating `quo` and undoing the subtraction if it turns out things + // were not normalized. + + /* + // Perform one binary long division step on the already normalized arguments, because + // the main. Note that this does a full normalization since the central loop needs + // `duo.leading_zeros()` to be at least 1 more than `div.leading_zeros()`. The original + // variation only did normalization to the nearest 4 steps, but this makes handling edge + // cases much harder. We do a full normalization and perform a binary long division + // step. In the edge case where the msbs of `duo` and `div` are set, it clears the msb + // of `duo`, then the edge case handler shifts `div` right and does another long + // division step to always insure `duo.leading_zeros() + 1 >= div.leading_zeros()`. + let div_original = div; + let mut shl = $normalization_shift(duo, div, true); + let mut div: $uX = (div << shl); + let mut quo: $uX = 1; + duo = duo.wrapping_sub(div); + if duo < div_original { + return (1 << shl, duo); + } + let div_neg: $uX; + if (div as $iX) < 0 { + // A very ugly edge case where the most significant bit of `div` is set (after + // shifting to match `duo` when its most significant bit is at the sign bit), which + // leads to the sign bit of `div_neg` being cut off and carries not happening when + // they should. This branch performs a long division step that keeps `duo` in place + // and shifts `div` down. + div >>= 1; + div_neg = div.wrapping_neg(); + let (sub, carry) = duo.overflowing_add(div_neg); + duo = sub; + quo = quo.wrapping_add(quo).wrapping_add(carry as $uX); + if !carry { + duo = duo.wrapping_add(div); + } + shl -= 1; + } else { + div_neg = div.wrapping_neg(); + } + // The add-with-carry that updates `quo` needs to have the carry set when a normalized + // subtract happens. Using `duo.wrapping_shl(1).overflowing_sub(div)` to do the + // subtraction generates a carry when an unnormalized subtract happens, which is the + // opposite of what we want. Instead, we use + // `duo.wrapping_shl(1).overflowing_add(div_neg)`, where `div_neg` is negative `div`. + let mut i = shl; + loop { + if i == 0 { + break; + } + i -= 1; + // `ADDS duo, div, duo, LSL #1` + // (add `div` to `duo << 1` and set flags) + let (sub, carry) = duo.wrapping_shl(1).overflowing_add(div_neg); + duo = sub; + // `ADC quo, quo, quo` + // (add with carry). Effectively shifts `quo` left by 1 and sets the least + // significant bit to the carry. + quo = quo.wrapping_add(quo).wrapping_add(carry as $uX); + // `ADDCC duo, duo, div` + // (add if carry clear). Undoes the subtraction if no carry was generated. + if !carry { + duo = duo.wrapping_add(div); + } + } + return (quo, duo >> shl); + */ + + // This is the SWAR (SIMD within in a register) restoring division algorithm. + // This combines several ideas of the above algorithms: + // - If `duo` is shifted left instead of shifting `div` right like in the 3 instruction + // restoring division algorithm, some architectures can do the shifting and + // subtraction step in one instruction. + // - `quo` can be constructed by adding powers-of-two to it or shifting it left by one + // and adding one. + // - Every time `duo` is shifted left, there is another unused 0 bit shifted into the + // LSB, so what if we use those bits to store `quo`? + // Through a complex setup, it is possible to manage `duo` and `quo` in the same + // register, and perform one step with 2 or 3 instructions. The only major downsides are + // that there is significant setup (it is only saves instructions if `shl` is + // approximately more than 4), `duo < div_original` checks are impractical once SWAR is + // initiated, and the number of division steps taken has to be exact (we cannot do more + // division steps than `shl`, because it introduces edge cases where quotient bits in + // `duo` start to collide with the real part of `div`. + /* + // first step. The quotient bit is stored in `quo` for now + let div_original = div; + let mut shl = $normalization_shift(duo, div, true); + let mut div: $uX = (div << shl); + duo = duo.wrapping_sub(div); + let mut quo: $uX = 1 << shl; + if duo < div_original { + return (quo, duo); + } + + let mask: $uX; + if (div as $iX) < 0 { + // deal with same edge case as the 3 instruction restoring division algorithm, but + // the quotient bit from this step also has to be stored in `quo` + div >>= 1; + shl -= 1; + let tmp = 1 << shl; + mask = tmp - 1; + let sub = duo.wrapping_sub(div); + if (sub as $iX) >= 0 { + // restore + duo = sub; + quo |= tmp; + } + if duo < div_original { + return (quo, duo); + } + } else { + mask = quo - 1; + } + // There is now room for quotient bits in `duo`. + + // Note that `div` is already shifted left and has `shl` unset bits. We subtract 1 from + // `div` and end up with the subset of `shl` bits being all being set. This subset acts + // just like a two's complement negative one. The subset of `div` containing the divisor + // had 1 subtracted from it, but a carry will always be generated from the `shl` subset + // as long as the quotient stays positive. + // + // When the modified `div` is subtracted from `duo.wrapping_shl(1)`, the `shl` subset + // adds a quotient bit to the least significant bit. + // For example, 89 (0b01011001) divided by 3 (0b11): + // + // shl:4, div:0b00110000 + // first step: + // duo:0b01011001 + // + div_neg:0b11010000 + // ____________________ + // 0b00101001 + // quo is set to 0b00010000 and mask is set to 0b00001111 for later + // + // 1 is subtracted from `div`. I will differentiate the `shl` part of `div` and the + // quotient part of `duo` with `^`s. + // chars. + // div:0b00110000 + // ^^^^ + // + 0b11111111 + // ________________ + // 0b00101111 + // ^^^^ + // div_neg:0b11010001 + // + // first SWAR step: + // duo_shl1:0b01010010 + // ^ + // + div_neg:0b11010001 + // ____________________ + // 0b00100011 + // ^ + // second: + // duo_shl1:0b01000110 + // ^^ + // + div_neg:0b11010001 + // ____________________ + // 0b00010111 + // ^^ + // third: + // duo_shl1:0b00101110 + // ^^^ + // + div_neg:0b11010001 + // ____________________ + // 0b11111111 + // ^^^ + // 3 steps resulted in the quotient with 3 set bits as expected, but currently the real + // part of `duo` is negative and the third step was an unnormalized step. The restore + // branch then restores `duo`. Note that the restore branch does not shift `duo` left. + // + // duo:0b11111111 + // ^^^ + // + div:0b00101111 + // ^^^^ + // ________________ + // 0b00101110 + // ^^^ + // `duo` is now back in the `duo_shl1` state it was at in the the third step, with an + // unset quotient bit. + // + // final step (`shl` was 4, so exactly 4 steps must be taken) + // duo_shl1:0b01011100 + // ^^^^ + // + div_neg:0b11010001 + // ____________________ + // 0b00101101 + // ^^^^ + // The quotient includes the `^` bits added with the `quo` bits from the beginning that + // contained the first step and potential edge case step, + // `quo:0b00010000 + (duo:0b00101101 & mask:0b00001111) == 0b00011101 == 29u8`. + // The remainder is the bits remaining in `duo` that are not part of the quotient bits, + // `duo:0b00101101 >> shl == 0b0010 == 2u8`. + let div: $uX = div.wrapping_sub(1); + let mut i = shl; + loop { + if i == 0 { + break; + } + i -= 1; + duo = duo.wrapping_shl(1).wrapping_sub(div); + if (duo as $iX) < 0 { + // restore + duo = duo.wrapping_add(div); + } + } + // unpack the results of SWAR + return ((duo & mask) | quo, duo >> shl); + */ + + // The problem with the conditional restoring SWAR algorithm above is that, in practice, + // it requires assembly code to bring out its full unrolled potential (It seems that + // LLVM can't use unrolled conditionals optimally and ends up erasing all the benefit + // that my algorithm intends. On architectures without predicated instructions, the code + // gen is especially bad. We need a default software division algorithm that is + // guaranteed to get decent code gen for the central loop. + + // For non-SWAR algorithms, there is a way to do binary long division without + // predication or even branching. This involves creating a mask from the sign bit and + // performing different kinds of steps using that. + /* + let shl = $normalization_shift(duo, div, true); + let mut div: $uX = div << shl; + let mut pow: $uX = 1 << shl; + let mut quo: $uX = 0; + loop { + let sub = duo.wrapping_sub(div); + let sign_mask = !((sub as $iX).wrapping_shr($n - 1) as $uX); + duo -= div & sign_mask; + quo |= pow & sign_mask; + div >>= 1; + pow >>= 1; + if pow == 0 { + break; + } + } + return (quo, duo); + */ + // However, it requires about 4 extra operations (smearing the sign bit, negating the + // mask, and applying the mask twice) on top of the operations done by the actual + // algorithm. With SWAR however, just 2 extra operations are needed, making it + // practical and even the most optimal algorithm for some architectures. + + // What we do is use custom assembly for predicated architectures that need software + // division, and for the default algorithm use a mask based restoring SWAR algorithm + // without conditionals or branches. On almost all architectures, this Rust code is + // guaranteed to compile down to 5 assembly instructions or less for each step, and LLVM + // will unroll it in a decent way. + + // standard opening for SWAR algorithm with first step and edge case handling + let div_original = div; + let mut shl = $normalization_shift(duo, div, true); + let mut div: $uX = (div << shl); + duo = duo.wrapping_sub(div); + let mut quo: $uX = 1 << shl; + if duo < div_original { + return (quo, duo); + } + let mask: $uX; + if (div as $iX) < 0 { + div >>= 1; + shl -= 1; + let tmp = 1 << shl; + mask = tmp - 1; + let sub = duo.wrapping_sub(div); + if (sub as $iX) >= 0 { + duo = sub; + quo |= tmp; + } + if duo < div_original { + return (quo, duo); + } + } else { + mask = quo - 1; + } + + // central loop + div = div.wrapping_sub(1); + let mut i = shl; + loop { + if i == 0 { + break; + } + i -= 1; + // shift left 1 and subtract + duo = duo.wrapping_shl(1).wrapping_sub(div); + // create mask + let mask = (duo as $iX).wrapping_shr($n - 1) as $uX; + // restore + duo = duo.wrapping_add(div & mask); + } + // unpack + return ((duo & mask) | quo, duo >> shl); + + // miscellanious binary long division algorithms that might be better for specific + // architectures + + // Another kind of long division uses an interesting fact that `div` and `pow` can be + // negated when `duo` is negative to perform a "negated" division step that works in + // place of any normalization mechanism. This is a non-restoring division algorithm that + // is very similar to the non-restoring division algorithms that can be found on the + // internet, except there is only one test for `duo < 0`. The subtraction from `quo` can + // be viewed as shifting the least significant set bit right (e.x. if we enter a series + // of negated binary long division steps starting with `quo == 0b1011_0000` and + // `pow == 0b0000_1000`, `quo` will progress like this: 0b1010_1000, 0b1010_0100, + // 0b1010_0010, 0b1010_0001). + /* + let div_original = div; + let shl = $normalization_shift(duo, div, true); + let mut div: $uX = (div << shl); + let mut pow: $uX = 1 << shl; + let mut quo: $uX = pow; + duo = duo.wrapping_sub(div); + if duo < div_original { + return (quo, duo); + } + div >>= 1; + pow >>= 1; + loop { + if (duo as $iX) < 0 { + // Negated binary long division step. + duo = duo.wrapping_add(div); + quo = quo.wrapping_sub(pow); + } else { + // Normal long division step. + if duo < div_original { + return (quo, duo) + } + duo = duo.wrapping_sub(div); + quo = quo.wrapping_add(pow); + } + pow >>= 1; + div >>= 1; + } + */ + + // This is the Nonrestoring SWAR algorithm, combining the nonrestoring algorithm with + // SWAR techniques that makes the only difference between steps be negation of `div`. + // If there was an architecture with an instruction that negated inputs to an adder + // based on conditionals, and in place shifting (or a three input addition operation + // that can have `duo` as two of the inputs to effectively shift it left by 1), then a + // single instruction central loop is possible. Microarchitectures often have inputs to + // their ALU that can invert the arguments and carry in of adders, but the architectures + // unfortunately do not have an instruction to dynamically invert this input based on + // conditionals. + /* + // SWAR opening + let div_original = div; + let mut shl = $normalization_shift(duo, div, true); + let mut div: $uX = (div << shl); + duo = duo.wrapping_sub(div); + let mut quo: $uX = 1 << shl; + if duo < div_original { + return (quo, duo); + } + let mask: $uX; + if (div as $iX) < 0 { + div >>= 1; + shl -= 1; + let tmp = 1 << shl; + let sub = duo.wrapping_sub(div); + if (sub as $iX) >= 0 { + // restore + duo = sub; + quo |= tmp; + } + if duo < div_original { + return (quo, duo); + } + mask = tmp - 1; + } else { + mask = quo - 1; + } + + // central loop + let div: $uX = div.wrapping_sub(1); + let mut i = shl; + loop { + if i == 0 { + break; + } + i -= 1; + // note: the `wrapping_shl(1)` can be factored out, but would require another + // restoring division step to prevent `(duo as $iX)` from overflowing + if (duo as $iX) < 0 { + // Negated binary long division step. + duo = duo.wrapping_shl(1).wrapping_add(div); + } else { + // Normal long division step. + duo = duo.wrapping_shl(1).wrapping_sub(div); + } + } + if (duo as $iX) < 0 { + // Restore. This was not needed in the original nonrestoring algorithm because of + // the `duo < div_original` checks. + duo = duo.wrapping_add(div); + } + // unpack + return ((duo & mask) | quo, duo >> shl); + */ + } + }; +} |