openssl/crypto/poly1305/poly1305_ieee754.c
Andy Polyakov 41013cd63c PPC assembly pack: correct POWER9 results.
As it turns out originally published results were skewed by "turbo"
mode. VM apparently remains oblivious to dynamic frequency scaling,
and reports that processor operates at "base" frequency at all times.
While actual frequency gets increased under load.

Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
2018-06-03 21:20:06 +02:00

488 lines
14 KiB
C

/*
* Copyright 2016-20018 The OpenSSL Project Authors. All Rights Reserved.
*
* Licensed under the OpenSSL license (the "License"). You may not use
* this file except in compliance with the License. You can obtain a copy
* in the file LICENSE in the source distribution or at
* https://www.openssl.org/source/license.html
*/
/*
* This module is meant to be used as template for non-x87 floating-
* point assembly modules. The template itself is x86_64-specific
* though, as it was debugged on x86_64. So that implementor would
* have to recognize platform-specific parts, UxTOy and inline asm,
* and act accordingly.
*
* Huh? x86_64-specific code as template for non-x87? Note seven, which
* is not a typo, but reference to 80-bit precision. This module on the
* other hand relies on 64-bit precision operations, which are default
* for x86_64 code. And since we are at it, just for sense of it,
* large-block performance in cycles per processed byte for *this* code
* is:
* gcc-4.8 icc-15.0 clang-3.4(*)
*
* Westmere 4.96 5.09 4.37
* Sandy Bridge 4.95 4.90 4.17
* Haswell 4.92 4.87 3.78
* Bulldozer 4.67 4.49 4.68
* VIA Nano 7.07 7.05 5.98
* Silvermont 10.6 9.61 12.6
*
* (*) clang managed to discover parallelism and deployed SIMD;
*
* And for range of other platforms with unspecified gcc versions:
*
* Freescale e300 12.5
* PPC74x0 10.8
* POWER6 4.92
* POWER7 4.50
* POWER8 4.10
*
* z10 11.2
* z196+ 7.30
*
* UltraSPARC III 16.0
* SPARC T4 16.1
*/
#if !(defined(__GNUC__) && __GNUC__>=2)
# error "this is gcc-specific template"
#endif
#include <stdlib.h>
typedef unsigned char u8;
typedef unsigned int u32;
typedef unsigned long long u64;
typedef union { double d; u64 u; } elem64;
#define TWO(p) ((double)(1ULL<<(p)))
#define TWO0 TWO(0)
#define TWO32 TWO(32)
#define TWO64 (TWO32*TWO(32))
#define TWO96 (TWO64*TWO(32))
#define TWO130 (TWO96*TWO(34))
#define EXP(p) ((1023ULL+(p))<<52)
#if defined(__x86_64__) || (defined(__PPC__) && defined(__LITTLE_ENDIAN__))
# define U8TOU32(p) (*(const u32 *)(p))
# define U32TO8(p,v) (*(u32 *)(p) = (v))
#elif defined(__PPC__)
# define U8TOU32(p) ({u32 ret; asm ("lwbrx %0,0,%1":"=r"(ret):"b"(p)); ret; })
# define U32TO8(p,v) asm ("stwbrx %0,0,%1"::"r"(v),"b"(p):"memory")
#elif defined(__s390x__)
# define U8TOU32(p) ({u32 ret; asm ("lrv %0,%1":"=d"(ret):"m"(*(u32 *)(p))); ret; })
# define U32TO8(p,v) asm ("strv %1,%0":"=m"(*(u32 *)(p)):"d"(v))
#endif
#ifndef U8TOU32
# define U8TOU32(p) ((u32)(p)[0] | (u32)(p)[1]<<8 | \
(u32)(p)[2]<<16 | (u32)(p)[3]<<24 )
#endif
#ifndef U32TO8
# define U32TO8(p,v) ((p)[0] = (u8)(v), (p)[1] = (u8)((v)>>8), \
(p)[2] = (u8)((v)>>16), (p)[3] = (u8)((v)>>24) )
#endif
typedef struct {
elem64 h[4];
double r[8];
double s[6];
} poly1305_internal;
/* "round toward zero (truncate), mask all exceptions" */
#if defined(__x86_64__)
static const u32 mxcsr = 0x7f80;
#elif defined(__PPC__)
static const u64 one = 1;
#elif defined(__s390x__)
static const u32 fpc = 1;
#elif defined(__sparc__)
static const u64 fsr = 1ULL<<30;
#elif defined(__mips__)
static const u32 fcsr = 1;
#else
#error "unrecognized platform"
#endif
int poly1305_init(void *ctx, const unsigned char key[16])
{
poly1305_internal *st = (poly1305_internal *) ctx;
elem64 r0, r1, r2, r3;
/* h = 0, biased */
#if 0
st->h[0].d = TWO(52)*TWO0;
st->h[1].d = TWO(52)*TWO32;
st->h[2].d = TWO(52)*TWO64;
st->h[3].d = TWO(52)*TWO96;
#else
st->h[0].u = EXP(52+0);
st->h[1].u = EXP(52+32);
st->h[2].u = EXP(52+64);
st->h[3].u = EXP(52+96);
#endif
if (key) {
/*
* set "truncate" rounding mode
*/
#if defined(__x86_64__)
u32 mxcsr_orig;
asm volatile ("stmxcsr %0":"=m"(mxcsr_orig));
asm volatile ("ldmxcsr %0"::"m"(mxcsr));
#elif defined(__PPC__)
double fpscr_orig, fpscr = *(double *)&one;
asm volatile ("mffs %0":"=f"(fpscr_orig));
asm volatile ("mtfsf 255,%0"::"f"(fpscr));
#elif defined(__s390x__)
u32 fpc_orig;
asm volatile ("stfpc %0":"=m"(fpc_orig));
asm volatile ("lfpc %0"::"m"(fpc));
#elif defined(__sparc__)
u64 fsr_orig;
asm volatile ("stx %%fsr,%0":"=m"(fsr_orig));
asm volatile ("ldx %0,%%fsr"::"m"(fsr));
#elif defined(__mips__)
u32 fcsr_orig;
asm volatile ("cfc1 %0,$31":"=r"(fcsr_orig));
asm volatile ("ctc1 %0,$31"::"r"(fcsr));
#endif
/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
r0.u = EXP(52+0) | (U8TOU32(&key[0]) & 0x0fffffff);
r1.u = EXP(52+32) | (U8TOU32(&key[4]) & 0x0ffffffc);
r2.u = EXP(52+64) | (U8TOU32(&key[8]) & 0x0ffffffc);
r3.u = EXP(52+96) | (U8TOU32(&key[12]) & 0x0ffffffc);
st->r[0] = r0.d - TWO(52)*TWO0;
st->r[2] = r1.d - TWO(52)*TWO32;
st->r[4] = r2.d - TWO(52)*TWO64;
st->r[6] = r3.d - TWO(52)*TWO96;
st->s[0] = st->r[2] * (5.0/TWO130);
st->s[2] = st->r[4] * (5.0/TWO130);
st->s[4] = st->r[6] * (5.0/TWO130);
/*
* base 2^32 -> base 2^16
*/
st->r[1] = (st->r[0] + TWO(52)*TWO(16)*TWO0) -
TWO(52)*TWO(16)*TWO0;
st->r[0] -= st->r[1];
st->r[3] = (st->r[2] + TWO(52)*TWO(16)*TWO32) -
TWO(52)*TWO(16)*TWO32;
st->r[2] -= st->r[3];
st->r[5] = (st->r[4] + TWO(52)*TWO(16)*TWO64) -
TWO(52)*TWO(16)*TWO64;
st->r[4] -= st->r[5];
st->r[7] = (st->r[6] + TWO(52)*TWO(16)*TWO96) -
TWO(52)*TWO(16)*TWO96;
st->r[6] -= st->r[7];
st->s[1] = (st->s[0] + TWO(52)*TWO(16)*TWO0/TWO96) -
TWO(52)*TWO(16)*TWO0/TWO96;
st->s[0] -= st->s[1];
st->s[3] = (st->s[2] + TWO(52)*TWO(16)*TWO32/TWO96) -
TWO(52)*TWO(16)*TWO32/TWO96;
st->s[2] -= st->s[3];
st->s[5] = (st->s[4] + TWO(52)*TWO(16)*TWO64/TWO96) -
TWO(52)*TWO(16)*TWO64/TWO96;
st->s[4] -= st->s[5];
/*
* restore original FPU control register
*/
#if defined(__x86_64__)
asm volatile ("ldmxcsr %0"::"m"(mxcsr_orig));
#elif defined(__PPC__)
asm volatile ("mtfsf 255,%0"::"f"(fpscr_orig));
#elif defined(__s390x__)
asm volatile ("lfpc %0"::"m"(fpc_orig));
#elif defined(__sparc__)
asm volatile ("ldx %0,%%fsr"::"m"(fsr_orig));
#elif defined(__mips__)
asm volatile ("ctc1 %0,$31"::"r"(fcsr_orig));
#endif
}
return 0;
}
void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
int padbit)
{
poly1305_internal *st = (poly1305_internal *)ctx;
elem64 in0, in1, in2, in3;
u64 pad = (u64)padbit<<32;
double x0, x1, x2, x3;
double h0lo, h0hi, h1lo, h1hi, h2lo, h2hi, h3lo, h3hi;
double c0lo, c0hi, c1lo, c1hi, c2lo, c2hi, c3lo, c3hi;
const double r0lo = st->r[0];
const double r0hi = st->r[1];
const double r1lo = st->r[2];
const double r1hi = st->r[3];
const double r2lo = st->r[4];
const double r2hi = st->r[5];
const double r3lo = st->r[6];
const double r3hi = st->r[7];
const double s1lo = st->s[0];
const double s1hi = st->s[1];
const double s2lo = st->s[2];
const double s2hi = st->s[3];
const double s3lo = st->s[4];
const double s3hi = st->s[5];
/*
* set "truncate" rounding mode
*/
#if defined(__x86_64__)
u32 mxcsr_orig;
asm volatile ("stmxcsr %0":"=m"(mxcsr_orig));
asm volatile ("ldmxcsr %0"::"m"(mxcsr));
#elif defined(__PPC__)
double fpscr_orig, fpscr = *(double *)&one;
asm volatile ("mffs %0":"=f"(fpscr_orig));
asm volatile ("mtfsf 255,%0"::"f"(fpscr));
#elif defined(__s390x__)
u32 fpc_orig;
asm volatile ("stfpc %0":"=m"(fpc_orig));
asm volatile ("lfpc %0"::"m"(fpc));
#elif defined(__sparc__)
u64 fsr_orig;
asm volatile ("stx %%fsr,%0":"=m"(fsr_orig));
asm volatile ("ldx %0,%%fsr"::"m"(fsr));
#elif defined(__mips__)
u32 fcsr_orig;
asm volatile ("cfc1 %0,$31":"=r"(fcsr_orig));
asm volatile ("ctc1 %0,$31"::"r"(fcsr));
#endif
/*
* load base 2^32 and de-bias
*/
h0lo = st->h[0].d - TWO(52)*TWO0;
h1lo = st->h[1].d - TWO(52)*TWO32;
h2lo = st->h[2].d - TWO(52)*TWO64;
h3lo = st->h[3].d - TWO(52)*TWO96;
#ifdef __clang__
h0hi = 0;
h1hi = 0;
h2hi = 0;
h3hi = 0;
#else
in0.u = EXP(52+0) | U8TOU32(&inp[0]);
in1.u = EXP(52+32) | U8TOU32(&inp[4]);
in2.u = EXP(52+64) | U8TOU32(&inp[8]);
in3.u = EXP(52+96) | U8TOU32(&inp[12]) | pad;
x0 = in0.d - TWO(52)*TWO0;
x1 = in1.d - TWO(52)*TWO32;
x2 = in2.d - TWO(52)*TWO64;
x3 = in3.d - TWO(52)*TWO96;
x0 += h0lo;
x1 += h1lo;
x2 += h2lo;
x3 += h3lo;
goto fast_entry;
#endif
do {
in0.u = EXP(52+0) | U8TOU32(&inp[0]);
in1.u = EXP(52+32) | U8TOU32(&inp[4]);
in2.u = EXP(52+64) | U8TOU32(&inp[8]);
in3.u = EXP(52+96) | U8TOU32(&inp[12]) | pad;
x0 = in0.d - TWO(52)*TWO0;
x1 = in1.d - TWO(52)*TWO32;
x2 = in2.d - TWO(52)*TWO64;
x3 = in3.d - TWO(52)*TWO96;
/*
* note that there are multiple ways to accumulate input, e.g.
* one can as well accumulate to h0lo-h1lo-h1hi-h2hi...
*/
h0lo += x0;
h0hi += x1;
h2lo += x2;
h2hi += x3;
/*
* carries that cross 32n-bit (and 130-bit) boundaries
*/
c0lo = (h0lo + TWO(52)*TWO32) - TWO(52)*TWO32;
c1lo = (h1lo + TWO(52)*TWO64) - TWO(52)*TWO64;
c2lo = (h2lo + TWO(52)*TWO96) - TWO(52)*TWO96;
c3lo = (h3lo + TWO(52)*TWO130) - TWO(52)*TWO130;
c0hi = (h0hi + TWO(52)*TWO32) - TWO(52)*TWO32;
c1hi = (h1hi + TWO(52)*TWO64) - TWO(52)*TWO64;
c2hi = (h2hi + TWO(52)*TWO96) - TWO(52)*TWO96;
c3hi = (h3hi + TWO(52)*TWO130) - TWO(52)*TWO130;
/*
* base 2^48 -> base 2^32 with last reduction step
*/
x1 = (h1lo - c1lo) + c0lo;
x2 = (h2lo - c2lo) + c1lo;
x3 = (h3lo - c3lo) + c2lo;
x0 = (h0lo - c0lo) + c3lo * (5.0/TWO130);
x1 += (h1hi - c1hi) + c0hi;
x2 += (h2hi - c2hi) + c1hi;
x3 += (h3hi - c3hi) + c2hi;
x0 += (h0hi - c0hi) + c3hi * (5.0/TWO130);
#ifndef __clang__
fast_entry:
#endif
/*
* base 2^32 * base 2^16 = base 2^48
*/
h0lo = s3lo * x1 + s2lo * x2 + s1lo * x3 + r0lo * x0;
h1lo = r0lo * x1 + s3lo * x2 + s2lo * x3 + r1lo * x0;
h2lo = r1lo * x1 + r0lo * x2 + s3lo * x3 + r2lo * x0;
h3lo = r2lo * x1 + r1lo * x2 + r0lo * x3 + r3lo * x0;
h0hi = s3hi * x1 + s2hi * x2 + s1hi * x3 + r0hi * x0;
h1hi = r0hi * x1 + s3hi * x2 + s2hi * x3 + r1hi * x0;
h2hi = r1hi * x1 + r0hi * x2 + s3hi * x3 + r2hi * x0;
h3hi = r2hi * x1 + r1hi * x2 + r0hi * x3 + r3hi * x0;
inp += 16;
len -= 16;
} while (len >= 16);
/*
* carries that cross 32n-bit (and 130-bit) boundaries
*/
c0lo = (h0lo + TWO(52)*TWO32) - TWO(52)*TWO32;
c1lo = (h1lo + TWO(52)*TWO64) - TWO(52)*TWO64;
c2lo = (h2lo + TWO(52)*TWO96) - TWO(52)*TWO96;
c3lo = (h3lo + TWO(52)*TWO130) - TWO(52)*TWO130;
c0hi = (h0hi + TWO(52)*TWO32) - TWO(52)*TWO32;
c1hi = (h1hi + TWO(52)*TWO64) - TWO(52)*TWO64;
c2hi = (h2hi + TWO(52)*TWO96) - TWO(52)*TWO96;
c3hi = (h3hi + TWO(52)*TWO130) - TWO(52)*TWO130;
/*
* base 2^48 -> base 2^32 with last reduction step
*/
x1 = (h1lo - c1lo) + c0lo;
x2 = (h2lo - c2lo) + c1lo;
x3 = (h3lo - c3lo) + c2lo;
x0 = (h0lo - c0lo) + c3lo * (5.0/TWO130);
x1 += (h1hi - c1hi) + c0hi;
x2 += (h2hi - c2hi) + c1hi;
x3 += (h3hi - c3hi) + c2hi;
x0 += (h0hi - c0hi) + c3hi * (5.0/TWO130);
/*
* store base 2^32, with bias
*/
st->h[1].d = x1 + TWO(52)*TWO32;
st->h[2].d = x2 + TWO(52)*TWO64;
st->h[3].d = x3 + TWO(52)*TWO96;
st->h[0].d = x0 + TWO(52)*TWO0;
/*
* restore original FPU control register
*/
#if defined(__x86_64__)
asm volatile ("ldmxcsr %0"::"m"(mxcsr_orig));
#elif defined(__PPC__)
asm volatile ("mtfsf 255,%0"::"f"(fpscr_orig));
#elif defined(__s390x__)
asm volatile ("lfpc %0"::"m"(fpc_orig));
#elif defined(__sparc__)
asm volatile ("ldx %0,%%fsr"::"m"(fsr_orig));
#elif defined(__mips__)
asm volatile ("ctc1 %0,$31"::"r"(fcsr_orig));
#endif
}
void poly1305_emit(void *ctx, unsigned char mac[16], const u32 nonce[4])
{
poly1305_internal *st = (poly1305_internal *) ctx;
u64 h0, h1, h2, h3, h4;
u32 g0, g1, g2, g3, g4;
u64 t;
u32 mask;
/*
* thanks to bias masking exponent gives integer result
*/
h0 = st->h[0].u & 0x000fffffffffffffULL;
h1 = st->h[1].u & 0x000fffffffffffffULL;
h2 = st->h[2].u & 0x000fffffffffffffULL;
h3 = st->h[3].u & 0x000fffffffffffffULL;
/*
* can be partially reduced, so reduce...
*/
h4 = h3>>32; h3 &= 0xffffffffU;
g4 = h4&-4;
h4 &= 3;
g4 += g4>>2;
h0 += g4;
h1 += h0>>32; h0 &= 0xffffffffU;
h2 += h1>>32; h1 &= 0xffffffffU;
h3 += h2>>32; h2 &= 0xffffffffU;
/* compute h + -p */
g0 = (u32)(t = h0 + 5);
g1 = (u32)(t = h1 + (t >> 32));
g2 = (u32)(t = h2 + (t >> 32));
g3 = (u32)(t = h3 + (t >> 32));
g4 = h4 + (u32)(t >> 32);
/* if there was carry, select g0-g3 */
mask = 0 - (g4 >> 2);
g0 &= mask;
g1 &= mask;
g2 &= mask;
g3 &= mask;
mask = ~mask;
g0 |= (h0 & mask);
g1 |= (h1 & mask);
g2 |= (h2 & mask);
g3 |= (h3 & mask);
/* mac = (h + nonce) % (2^128) */
g0 = (u32)(t = (u64)g0 + nonce[0]);
g1 = (u32)(t = (u64)g1 + (t >> 32) + nonce[1]);
g2 = (u32)(t = (u64)g2 + (t >> 32) + nonce[2]);
g3 = (u32)(t = (u64)g3 + (t >> 32) + nonce[3]);
U32TO8(mac + 0, g0);
U32TO8(mac + 4, g1);
U32TO8(mac + 8, g2);
U32TO8(mac + 12, g3);
}