Add MSVC compiler support for 128-bit div/mul (untested)

This commit is contained in:
Mamy André-Ratsimbazafy 2020-03-01 16:02:14 +01:00
parent dfb007e6bc
commit 3fdd457b52
No known key found for this signature in database
GPG Key ID: 7B88AD1FE79492E1
1 changed files with 79 additions and 2 deletions

View File

@ -159,7 +159,84 @@ when sizeof(int) == 8 and GccCompatible:
# Don't forget to dereference the var param # Don't forget to dereference the var param
{.emit:["*",hi, " = (NU64)(", dblPrec," >> ", 63'u64, ");"].} {.emit:["*",hi, " = (NU64)(", dblPrec," >> ", 63'u64, ");"].}
elif sizeof(int) == 8 and defined(vcc):
func udiv128(highDividend, lowDividend, divisor: uint64, remainder: var uint64): uint64 {.importc:"_udiv128", header: "<immintrin.h>", nodecl.}
## Division 128 by 64, Microsoft only, 64-bit only,
## returns quotient as return value remainder as var parameter
## Warning ⚠️ :
## - if n_hi == d, quotient does not fit in an uint64 and will throw SIGFPE
## - if n_hi > d result is undefined
func unsafeDiv2n1n*(q, r: var Ct[uint64], n_hi, n_lo, d: Ct[uint64]) {.inline.}=
## Division uint128 by uint64
## Warning ⚠️ :
## - if n_hi == d, quotient does not fit in an uint64 and will throw SIGFPE
## - if n_hi > d result is undefined
{.warning: "unsafeDiv2n1n is not constant-time at the moment on most hardware".}
# TODO !!! - Replace by constant-time, portable, non-assembly version
# -> use uint128? Compiler might add unwanted branches
q = udiv128(n_hi, n_lo, d, r)
func addcarry_u64(carryIn: cuchar, a, b: uint64, sum: var uint64): cuchar {.importc:"_addcarry_u64", header:"<intrin.h>", nodecl.}
## (CarryOut, Sum) <-- a + b
## Available on MSVC and ICC (Clang and GCC have very bad codegen, use uint128 instead)
## Return value is the carry-out
func umul128(a, b: uint64, hi: var uint64): uint64 {.importc:"_umul128", header:"<intrin.h>", nodecl.}
## (hi, lo) <-- a * b
## Return value is the low word
func unsafeFMA*(hi, lo: var Ct[uint64], a, b, c: Ct[uint64]) {.inline.}=
## Extended precision multiplication + addition
## This is constant-time on most hardware except some specific one like Cortex M0
## (hi, lo) <- a*b + c
var carry: cuchar
var hi, lo: uint64
lo = umul128(uint64(a), uint64(b), hi)
carry = addcarry_u64(cuchar(0), lo, uint64(c), lo)
discard addcarry_u64(carry, hi, 0, hi)
func unsafeFMA2*(hi, lo: var Ct[uint64], a1, b1, a2, b2, c1, c2: Ct[uint64]) {.inline.}=
## (hi, lo) <- a1 * b1 + a2 * b2 + c1 + c2
var f1_lo, f1_hi, f2_lo, f2_hi: uint64
var carry: cuchar
f1_lo = umul128(uint64(a1), uint64(b1), f1_hi)
f2_lo = umul128(uint64(a2), uint64(b2), f2_hi)
# On CPU with ADX: we can use addcarryx_u64 (adcx/adox) to have
# separate carry chains that can be processed in parallel by CPU
# Carry chain 1
carry = addcarry_u64(cuchar(0), f1_lo, uint64(c1), f1_lo)
discard addcarry_u64(carry, f1_hi, 0, f1_hi)
# Carry chain 2
carry = addcarry_u64(cuchar(0), f2_lo, uint64(c2), f2_lo)
discard addcarry_u64(carry, f2_hi, 0, f2_hi)
# Merge
carry = addcarry_u64(cuchar(0), f1_lo, f2_lo, lo)
discard addcarry_u64(carry, f1_hi, f2_hi, hi)
func unsafeFMA2_hi*(hi: var Ct[uint64], a1, b1, a2, b2, c: Ct[uint64]) {.inline.}=
## Returns the high word of the sum of extended precision multiply-adds
## (hi, _) <- a1 * b1 + a2 * b2 + c
var f1_lo, f1_hi, f2_lo, f2_hi: uint64
var carry: cuchar
f1_lo = umul128(uint64(a1), uint64(b1), f1_hi)
f2_lo = umul128(uint64(a2), uint64(b2), f2_hi)
carry = addcarry_u64(cuchar(0), f1_lo, uint64(c), f1_lo)
discard addcarry_u64(carry, f1_hi, 0, f1_hi)
# Merge
var lo: uint64
carry = addcarry_u64(cuchar(0), f1_lo, f2_lo, lo)
discard addcarry_u64(carry, f1_hi, f2_hi, hi)
else: else:
{.error: "Compiler not implemented".} {.error: "Compiler not implemented".}
# For VCC and ICC use add_carry_u64, _add_carryx_u64 intrinsics
# and _umul128