diff options
Diffstat (limited to 'powerpc-cpu/sysdeps/powerpc/powerpc32/power4')
7 files changed, 977 insertions, 0 deletions
diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Implies new file mode 100644 index 0000000000..b2ac1558f5 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Implies @@ -0,0 +1 @@ +powerpc/powerpc32/powerpc64
\ No newline at end of file diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Makefile b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Makefile new file mode 100644 index 0000000000..60aa508ba4 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/Makefile @@ -0,0 +1,6 @@ +# Makefile fragment for POWER4/5/5+. + +ifeq ($(subdir),string) +CFLAGS-wordcopy.c += --param max-variable-expansions-in-unroller=2 --param max-unroll-times=2 -funroll-loops -fpeel-loops -ftree-loop-linear +CFLAGS-memmove.c += --param max-variable-expansions-in-unroller=2 --param max-unroll-times=2 -funroll-loops -fpeel-loops -ftree-loop-linear +endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/Makefile b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/Makefile new file mode 100644 index 0000000000..a6fa75ecbc --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/Makefile @@ -0,0 +1,5 @@ +# Makefile fragment for POWER4/5/5+ with FPU. + +ifeq ($(subdir),math) +CFLAGS-mpa.c += --param max-unroll-times=4 -funroll-loops -fpeel-loops -ftree-loop-linear +endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/mpa.c b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/mpa.c new file mode 100644 index 0000000000..4a232e27bf --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/mpa.c @@ -0,0 +1,549 @@ + +/* + * IBM Accurate Mathematical Library + * written by International Business Machines Corp. + * Copyright (C) 2001, 2006 Free Software Foundation + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU Lesser General Public License as published by + * the Free Software Foundation; either version 2.1 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + */ +/************************************************************************/ +/* MODULE_NAME: mpa.c */ +/* */ +/* FUNCTIONS: */ +/* mcr */ +/* acr */ +/* cr */ +/* cpy */ +/* cpymn */ +/* norm */ +/* denorm */ +/* mp_dbl */ +/* dbl_mp */ +/* add_magnitudes */ +/* sub_magnitudes */ +/* add */ +/* sub */ +/* mul */ +/* inv */ +/* dvd */ +/* */ +/* Arithmetic functions for multiple precision numbers. */ +/* Relative errors are bounded */ +/************************************************************************/ + + +#include "endian.h" +#include "mpa.h" +#include "mpa2.h" +#include <sys/param.h> /* For MIN() */ +/* mcr() compares the sizes of the mantissas of two multiple precision */ +/* numbers. Mantissas are compared regardless of the signs of the */ +/* numbers, even if x->d[0] or y->d[0] are zero. Exponents are also */ +/* disregarded. */ +static int mcr(const mp_no *x, const mp_no *y, int p) { + long i; + long p2 = p; + for (i=1; i<=p2; i++) { + if (X[i] == Y[i]) continue; + else if (X[i] > Y[i]) return 1; + else return -1; } + return 0; +} + + + +/* acr() compares the absolute values of two multiple precision numbers */ +int __acr(const mp_no *x, const mp_no *y, int p) { + long i; + + if (X[0] == ZERO) { + if (Y[0] == ZERO) i= 0; + else i=-1; + } + else if (Y[0] == ZERO) i= 1; + else { + if (EX > EY) i= 1; + else if (EX < EY) i=-1; + else i= mcr(x,y,p); + } + + return i; +} + + +/* cr90 compares the values of two multiple precision numbers */ +int __cr(const mp_no *x, const mp_no *y, int p) { + int i; + + if (X[0] > Y[0]) i= 1; + else if (X[0] < Y[0]) i=-1; + else if (X[0] < ZERO ) i= __acr(y,x,p); + else i= __acr(x,y,p); + + return i; +} + + +/* Copy a multiple precision number. Set *y=*x. x=y is permissible. */ +void __cpy(const mp_no *x, mp_no *y, int p) { + long i; + + EY = EX; + for (i=0; i <= p; i++) Y[i] = X[i]; + + return; +} + + +/* Copy a multiple precision number x of precision m into a */ +/* multiple precision number y of precision n. In case n>m, */ +/* the digits of y beyond the m'th are set to zero. In case */ +/* n<m, the digits of x beyond the n'th are ignored. */ +/* x=y is permissible. */ + +void __cpymn(const mp_no *x, int m, mp_no *y, int n) { + + long i,k; + long n2 = n; + long m2 = m; + + EY = EX; k=MIN(m2,n2); + for (i=0; i <= k; i++) Y[i] = X[i]; + for ( ; i <= n2; i++) Y[i] = ZERO; + + return; +} + +/* Convert a multiple precision number *x into a double precision */ +/* number *y, normalized case (|x| >= 2**(-1022))) */ +static void norm(const mp_no *x, double *y, int p) +{ + #define R radixi.d + long i; +#if 0 + int k; +#endif + double a,c,u,v,z[5]; + if (p<5) { + if (p==1) c = X[1]; + else if (p==2) c = X[1] + R* X[2]; + else if (p==3) c = X[1] + R*(X[2] + R* X[3]); + else if (p==4) c =(X[1] + R* X[2]) + R*R*(X[3] + R*X[4]); + } + else { + for (a=ONE, z[1]=X[1]; z[1] < TWO23; ) + {a *= TWO; z[1] *= TWO; } + + for (i=2; i<5; i++) { + z[i] = X[i]*a; + u = (z[i] + CUTTER)-CUTTER; + if (u > z[i]) u -= RADIX; + z[i] -= u; + z[i-1] += u*RADIXI; + } + + u = (z[3] + TWO71) - TWO71; + if (u > z[3]) u -= TWO19; + v = z[3]-u; + + if (v == TWO18) { + if (z[4] == ZERO) { + for (i=5; i <= p; i++) { + if (X[i] == ZERO) continue; + else {z[3] += ONE; break; } + } + } + else z[3] += ONE; + } + + c = (z[1] + R *(z[2] + R * z[3]))/a; + } + + c *= X[0]; + + for (i=1; i<EX; i++) c *= RADIX; + for (i=1; i>EX; i--) c *= RADIXI; + + *y = c; + return; +#undef R +} + +/* Convert a multiple precision number *x into a double precision */ +/* number *y, denormalized case (|x| < 2**(-1022))) */ +static void denorm(const mp_no *x, double *y, int p) +{ + long i,k; + long p2 = p; + double c,u,z[5]; +#if 0 + double a,v; +#endif + +#define R radixi.d + if (EX<-44 || (EX==-44 && X[1]<TWO5)) + { *y=ZERO; return; } + + if (p2==1) { + if (EX==-42) {z[1]=X[1]+TWO10; z[2]=ZERO; z[3]=ZERO; k=3;} + else if (EX==-43) {z[1]= TWO10; z[2]=X[1]; z[3]=ZERO; k=2;} + else {z[1]= TWO10; z[2]=ZERO; z[3]=X[1]; k=1;} + } + else if (p2==2) { + if (EX==-42) {z[1]=X[1]+TWO10; z[2]=X[2]; z[3]=ZERO; k=3;} + else if (EX==-43) {z[1]= TWO10; z[2]=X[1]; z[3]=X[2]; k=2;} + else {z[1]= TWO10; z[2]=ZERO; z[3]=X[1]; k=1;} + } + else { + if (EX==-42) {z[1]=X[1]+TWO10; z[2]=X[2]; k=3;} + else if (EX==-43) {z[1]= TWO10; z[2]=X[1]; k=2;} + else {z[1]= TWO10; z[2]=ZERO; k=1;} + z[3] = X[k]; + } + + u = (z[3] + TWO57) - TWO57; + if (u > z[3]) u -= TWO5; + + if (u==z[3]) { + for (i=k+1; i <= p2; i++) { + if (X[i] == ZERO) continue; + else {z[3] += ONE; break; } + } + } + + c = X[0]*((z[1] + R*(z[2] + R*z[3])) - TWO10); + + *y = c*TWOM1032; + return; + +#undef R +} + +/* Convert a multiple precision number *x into a double precision number *y. */ +/* The result is correctly rounded to the nearest/even. *x is left unchanged */ + +void __mp_dbl(const mp_no *x, double *y, int p) { +#if 0 + int i,k; + double a,c,u,v,z[5]; +#endif + + if (X[0] == ZERO) {*y = ZERO; return; } + + if (EX> -42) norm(x,y,p); + else if (EX==-42 && X[1]>=TWO10) norm(x,y,p); + else denorm(x,y,p); +} + + +/* dbl_mp() converts a double precision number x into a multiple precision */ +/* number *y. If the precision p is too small the result is truncated. x is */ +/* left unchanged. */ + +void __dbl_mp(double x, mp_no *y, int p) { + + long i,n; + long p2 = p; + double u; + + /* Sign */ + if (x == ZERO) {Y[0] = ZERO; return; } + else if (x > ZERO) Y[0] = ONE; + else {Y[0] = MONE; x=-x; } + + /* Exponent */ + for (EY=ONE; x >= RADIX; EY += ONE) x *= RADIXI; + for ( ; x < ONE; EY -= ONE) x *= RADIX; + + /* Digits */ + n=MIN(p2,4); + for (i=1; i<=n; i++) { + u = (x + TWO52) - TWO52; + if (u>x) u -= ONE; + Y[i] = u; x -= u; x *= RADIX; } + for ( ; i<=p2; i++) Y[i] = ZERO; + return; +} + + +/* add_magnitudes() adds the magnitudes of *x & *y assuming that */ +/* abs(*x) >= abs(*y) > 0. */ +/* The sign of the sum *z is undefined. x&y may overlap but not x&z or y&z. */ +/* No guard digit is used. The result equals the exact sum, truncated. */ +/* *x & *y are left unchanged. */ + +static void add_magnitudes(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + long i,j,k; + long p2 = p; + + EZ = EX; + + i=p2; j=p2+ EY - EX; k=p2+1; + + if (j<1) + {__cpy(x,z,p); return; } + else Z[k] = ZERO; + + for (; j>0; i--,j--) { + Z[k] += X[i] + Y[j]; + if (Z[k] >= RADIX) { + Z[k] -= RADIX; + Z[--k] = ONE; } + else + Z[--k] = ZERO; + } + + for (; i>0; i--) { + Z[k] += X[i]; + if (Z[k] >= RADIX) { + Z[k] -= RADIX; + Z[--k] = ONE; } + else + Z[--k] = ZERO; + } + + if (Z[1] == ZERO) { + for (i=1; i<=p2; i++) Z[i] = Z[i+1]; } + else EZ += ONE; +} + + +/* sub_magnitudes() subtracts the magnitudes of *x & *y assuming that */ +/* abs(*x) > abs(*y) > 0. */ +/* The sign of the difference *z is undefined. x&y may overlap but not x&z */ +/* or y&z. One guard digit is used. The error is less than one ulp. */ +/* *x & *y are left unchanged. */ + +static void sub_magnitudes(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + long i,j,k; + long p2 = p; + + EZ = EX; + + if (EX == EY) { + i=j=k=p2; + Z[k] = Z[k+1] = ZERO; } + else { + j= EX - EY; + if (j > p2) {__cpy(x,z,p); return; } + else { + i=p2; j=p2+1-j; k=p2; + if (Y[j] > ZERO) { + Z[k+1] = RADIX - Y[j--]; + Z[k] = MONE; } + else { + Z[k+1] = ZERO; + Z[k] = ZERO; j--;} + } + } + + for (; j>0; i--,j--) { + Z[k] += (X[i] - Y[j]); + if (Z[k] < ZERO) { + Z[k] += RADIX; + Z[--k] = MONE; } + else + Z[--k] = ZERO; + } + + for (; i>0; i--) { + Z[k] += X[i]; + if (Z[k] < ZERO) { + Z[k] += RADIX; + Z[--k] = MONE; } + else + Z[--k] = ZERO; + } + + for (i=1; Z[i] == ZERO; i++) ; + EZ = EZ - i + 1; + for (k=1; i <= p2+1; ) + Z[k++] = Z[i++]; + for (; k <= p2; ) + Z[k++] = ZERO; + + return; +} + + +/* Add two multiple precision numbers. Set *z = *x + *y. x&y may overlap */ +/* but not x&z or y&z. One guard digit is used. The error is less than */ +/* one ulp. *x & *y are left unchanged. */ + +void __add(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + int n; + + if (X[0] == ZERO) {__cpy(y,z,p); return; } + else if (Y[0] == ZERO) {__cpy(x,z,p); return; } + + if (X[0] == Y[0]) { + if (__acr(x,y,p) > 0) {add_magnitudes(x,y,z,p); Z[0] = X[0]; } + else {add_magnitudes(y,x,z,p); Z[0] = Y[0]; } + } + else { + if ((n=__acr(x,y,p)) == 1) {sub_magnitudes(x,y,z,p); Z[0] = X[0]; } + else if (n == -1) {sub_magnitudes(y,x,z,p); Z[0] = Y[0]; } + else Z[0] = ZERO; + } + return; +} + + +/* Subtract two multiple precision numbers. *z is set to *x - *y. x&y may */ +/* overlap but not x&z or y&z. One guard digit is used. The error is */ +/* less than one ulp. *x & *y are left unchanged. */ + +void __sub(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + int n; + + if (X[0] == ZERO) {__cpy(y,z,p); Z[0] = -Z[0]; return; } + else if (Y[0] == ZERO) {__cpy(x,z,p); return; } + + if (X[0] != Y[0]) { + if (__acr(x,y,p) > 0) {add_magnitudes(x,y,z,p); Z[0] = X[0]; } + else {add_magnitudes(y,x,z,p); Z[0] = -Y[0]; } + } + else { + if ((n=__acr(x,y,p)) == 1) {sub_magnitudes(x,y,z,p); Z[0] = X[0]; } + else if (n == -1) {sub_magnitudes(y,x,z,p); Z[0] = -Y[0]; } + else Z[0] = ZERO; + } + return; +} + + +/* Multiply two multiple precision numbers. *z is set to *x * *y. x&y */ +/* may overlap but not x&z or y&z. In case p=1,2,3 the exact result is */ +/* truncated to p digits. In case p>3 the error is bounded by 1.001 ulp. */ +/* *x & *y are left unchanged. */ + +void __mul(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + long i, i1, i2, j, k, k2; + long p2 = p; + double u, zk, zk2; + + /* Is z=0? */ + if (X[0]*Y[0]==ZERO) + { Z[0]=ZERO; return; } + + /* Multiply, add and carry */ + k2 = (p2<3) ? p2+p2 : p2+3; + zk = Z[k2]=ZERO; + for (k=k2; k>1; ) { + if (k > p2) {i1=k-p2; i2=p2+1; } + else {i1=1; i2=k; } +#if 1 + /* rearange this inner loop to allow the fmadd instructions to be + independent and execute in parallel on processors that have + dual symetrical FP pipelines. */ + if (i1 < (i2-1)) + { + /* make sure we have at least 2 iterations */ + if (((i2 - i1) & 1L) == 1L) + { + /* Handle the odd iterations case. */ + zk2 = x->d[i2-1]*y->d[i1]; + } + else + zk2 = zero.d; + /* Do two multiply/adds per loop iteration, using independent + accumulators; zk and zk2. */ + for (i=i1,j=i2-1; i<i2-1; i+=2,j-=2) + { + zk += x->d[i]*y->d[j]; + zk2 += x->d[i+1]*y->d[j-1]; + } + zk += zk2; /* final sum. */ + } + else + { + /* Special case when iterations is 1. */ + zk += x->d[i1]*y->d[i1]; + } +#else + /* The orginal code. */ + for (i=i1,j=i2-1; i<i2; i++,j--) zk += X[i]*Y[j]; +#endif + + u = (zk + CUTTER)-CUTTER; + if (u > zk) u -= RADIX; + Z[k] = zk - u; + zk = u*RADIXI; + --k; + } + Z[k] = zk; + + /* Is there a carry beyond the most significant digit? */ + if (Z[1] == ZERO) { + for (i=1; i<=p2; i++) Z[i]=Z[i+1]; + EZ = EX + EY - 1; } + else + EZ = EX + EY; + + Z[0] = X[0] * Y[0]; + return; +} + + +/* Invert a multiple precision number. Set *y = 1 / *x. */ +/* Relative error bound = 1.001*r**(1-p) for p=2, 1.063*r**(1-p) for p=3, */ +/* 2.001*r**(1-p) for p>3. */ +/* *x=0 is not permissible. *x is left unchanged. */ + +void __inv(const mp_no *x, mp_no *y, int p) { + long i; +#if 0 + int l; +#endif + double t; + mp_no z,w; + static const int np1[] = {0,0,0,0,1,2,2,2,2,3,3,3,3,3,3,3,3,3, + 4,4,4,4,4,4,4,4,4,4,4,4,4,4,4}; + const mp_no mptwo = {1,{1.0,2.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0, + 0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0, + 0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0, + 0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0}}; + + __cpy(x,&z,p); z.e=0; __mp_dbl(&z,&t,p); + t=ONE/t; __dbl_mp(t,y,p); EY -= EX; + + for (i=0; i<np1[p]; i++) { + __cpy(y,&w,p); + __mul(x,&w,y,p); + __sub(&mptwo,y,&z,p); + __mul(&w,&z,y,p); + } + return; +} + + +/* Divide one multiple precision number by another.Set *z = *x / *y. *x & *y */ +/* are left unchanged. x&y may overlap but not x&z or y&z. */ +/* Relative error bound = 2.001*r**(1-p) for p=2, 2.063*r**(1-p) for p=3 */ +/* and 3.001*r**(1-p) for p>3. *y=0 is not permissible. */ + +void __dvd(const mp_no *x, const mp_no *y, mp_no *z, int p) { + + mp_no w; + + if (X[0] == ZERO) Z[0] = ZERO; + else {__inv(y,&w,p); __mul(x,&w,z,p);} + return; +} diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/slowpow.c b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/slowpow.c new file mode 100644 index 0000000000..ad147a89a6 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/fpu/slowpow.c @@ -0,0 +1,94 @@ +/* + * IBM Accurate Mathematical Library + * written by International Business Machines Corp. + * Copyright (C) 2001, 2006 Free Software Foundation + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU Lesser General Public License as published by + * the Free Software Foundation; either version 2.1 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + */ +/*************************************************************************/ +/* MODULE_NAME:slowpow.c */ +/* */ +/* FUNCTION:slowpow */ +/* */ +/*FILES NEEDED:mpa.h */ +/* mpa.c mpexp.c mplog.c halfulp.c */ +/* */ +/* Given two IEEE double machine numbers y,x , routine computes the */ +/* correctly rounded (to nearest) value of x^y. Result calculated by */ +/* multiplication (in halfulp.c) or if result isn't accurate enough */ +/* then routine converts x and y into multi-precision doubles and */ +/* recompute. */ +/*************************************************************************/ + +#include "mpa.h" +#include "math_private.h" + +void __mpexp (mp_no * x, mp_no * y, int p); +void __mplog (mp_no * x, mp_no * y, int p); +double ulog (double); +double __halfulp (double x, double y); + +double +__slowpow (double x, double y, double z) +{ + double res, res1; + long double ldw, ldz, ldpp; + static const long double ldeps = 0x4.0p-96; + + res = __halfulp (x, y); /* halfulp() returns -10 or x^y */ + if (res >= 0) + return res; /* if result was really computed by halfulp */ + /* else, if result was not really computed by halfulp */ + + /* Compute pow as long double, 106 bits */ + ldz = __ieee754_logl ((long double) x); + ldw = (long double) y *ldz; + ldpp = __ieee754_expl (ldw); + res = (double) (ldpp + ldeps); + res1 = (double) (ldpp - ldeps); + + if (res != res1) /* if result still not accurate enough */ + { /* use mpa for higher persision. */ + mp_no mpx, mpy, mpz, mpw, mpp, mpr, mpr1; + static const mp_no eps = { -3, {1.0, 4.0} }; + int p; + + p = 10; /* p=precision 240 bits */ + __dbl_mp (x, &mpx, p); + __dbl_mp (y, &mpy, p); + __dbl_mp (z, &mpz, p); + __mplog (&mpx, &mpz, p); /* log(x) = z */ + __mul (&mpy, &mpz, &mpw, p); /* y * z =w */ + __mpexp (&mpw, &mpp, p); /* e^w =pp */ + __add (&mpp, &eps, &mpr, p); /* pp+eps =r */ + __mp_dbl (&mpr, &res, p); + __sub (&mpp, &eps, &mpr1, p); /* pp -eps =r1 */ + __mp_dbl (&mpr1, &res1, p); /* converting into double precision */ + if (res == res1) + return res; + + /* if we get here result wasn't calculated exactly, continue for + more exact calculation using 768 bits. */ + p = 32; + __dbl_mp (x, &mpx, p); + __dbl_mp (y, &mpy, p); + __dbl_mp (z, &mpz, p); + __mplog (&mpx, &mpz, p); /* log(c)=z */ + __mul (&mpy, &mpz, &mpw, p); /* y*z =w */ + __mpexp (&mpw, &mpp, p); /* e^w=pp */ + __mp_dbl (&mpp, &res, p); /* converting into double precision */ + } + return res; +} diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/memcopy.h b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/memcopy.h new file mode 100644 index 0000000000..c05208da55 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/memcopy.h @@ -0,0 +1,113 @@ +/* memcopy.h -- definitions for memory copy functions. Generic C version. + Copyright (C) 1991, 1992, 1993, 1997, 2004, 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Torbjorn Granlund (tege@sics.se). + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +/* The strategy of the memory functions is: + + 1. Copy bytes until the destination pointer is aligned. + + 2. Copy words in unrolled loops. If the source and destination + are not aligned in the same way, use word memory operations, + but shift and merge two read words before writing. + + 3. Copy the few remaining bytes. + + This is fast on processors that have at least 10 registers for + allocation by GCC, and that can access memory at reg+const in one + instruction. + + I made an "exhaustive" test of this memmove when I wrote it, + exhaustive in the sense that I tried all alignment and length + combinations, with and without overlap. */ + +#include <sysdeps/generic/memcopy.h> + +/* The macros defined in this file are: + + BYTE_COPY_FWD(dst_beg_ptr, src_beg_ptr, nbytes_to_copy) + + BYTE_COPY_BWD(dst_end_ptr, src_end_ptr, nbytes_to_copy) + + WORD_COPY_FWD(dst_beg_ptr, src_beg_ptr, nbytes_remaining, nbytes_to_copy) + + WORD_COPY_BWD(dst_end_ptr, src_end_ptr, nbytes_remaining, nbytes_to_copy) + + MERGE(old_word, sh_1, new_word, sh_2) + [I fail to understand. I feel stupid. --roland] +*/ + + +/* Threshold value for when to enter the unrolled loops. */ +#undef OP_T_THRES +#define OP_T_THRES 16 + +/* Copy exactly NBYTES bytes from SRC_BP to DST_BP, + without any assumptions about alignment of the pointers. */ +#undef BYTE_COPY_FWD +#define BYTE_COPY_FWD(dst_bp, src_bp, nbytes) \ + do \ + { \ + size_t __nbytes = (nbytes); \ + if (__nbytes & 1) \ + { \ + ((byte *) dst_bp)[0] = ((byte *) src_bp)[0]; \ + src_bp += 1; \ + dst_bp += 1; \ + __nbytes -= 1; \ + } \ + while (__nbytes > 0) \ + { \ + byte __x = ((byte *) src_bp)[0]; \ + byte __y = ((byte *) src_bp)[1]; \ + src_bp += 2; \ + __nbytes -= 2; \ + ((byte *) dst_bp)[0] = __x; \ + ((byte *) dst_bp)[1] = __y; \ + dst_bp += 2; \ + } \ + } while (0) + +/* Copy exactly NBYTES_TO_COPY bytes from SRC_END_PTR to DST_END_PTR, + beginning at the bytes right before the pointers and continuing towards + smaller addresses. Don't assume anything about alignment of the + pointers. */ +#undef BYTE_COPY_BWD +#define BYTE_COPY_BWD(dst_ep, src_ep, nbytes) \ + do \ + { \ + size_t __nbytes = (nbytes); \ + if (__nbytes & 1) \ + { \ + src_ep -= 1; \ + dst_ep -= 1; \ + ((byte *) dst_ep)[0] = ((byte *) src_ep)[0]; \ + __nbytes -= 1; \ + } \ + while (__nbytes > 0) \ + { \ + byte __x, __y; \ + src_ep -= 2; \ + __y = ((byte *) src_ep)[1]; \ + __x = ((byte *) src_ep)[0]; \ + dst_ep -= 2; \ + __nbytes -= 2; \ + ((byte *) dst_ep)[1] = __y; \ + ((byte *) dst_ep)[0] = __x; \ + } \ + } while (0) diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/wordcopy.c b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/wordcopy.c new file mode 100644 index 0000000000..f71b41dc42 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc32/power4/wordcopy.c @@ -0,0 +1,209 @@ +/* _memcopy.c -- subroutines for memory copy functions. + Copyright (C) 1991, 1996 Free Software Foundation, Inc. + This file is part of the GNU C Library. + Contributed by Torbjorn Granlund (tege@sics.se). + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, write to the Free + Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA + 02111-1307 USA. */ + +/* BE VERY CAREFUL IF YOU CHANGE THIS CODE...! */ + +#include <stddef.h> +#include <memcopy.h> + +/* _wordcopy_fwd_aligned -- Copy block beginning at SRCP to + block beginning at DSTP with LEN `op_t' words (not LEN bytes!). + Both SRCP and DSTP should be aligned for memory operations on `op_t's. */ + +void +_wordcopy_fwd_aligned (dstp, srcp, len) + long int dstp; + long int srcp; + size_t len; +{ + op_t a0, a1; + + if (len & 1) + { + ((op_t *) dstp)[0] = ((op_t *) srcp)[0]; + + if (len == 1) + return; + srcp += OPSIZ; + dstp += OPSIZ; + len -= 1; + } + + do + { + a0 = ((op_t *) srcp)[0]; + a1 = ((op_t *) srcp)[1]; + ((op_t *) dstp)[0] = a0; + ((op_t *) dstp)[1] = a1; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); +} + +/* _wordcopy_fwd_dest_aligned -- Copy block beginning at SRCP to + block beginning at DSTP with LEN `op_t' words (not LEN bytes!). + DSTP should be aligned for memory operations on `op_t's, but SRCP must + *not* be aligned. */ + +void +_wordcopy_fwd_dest_aligned (dstp, srcp, len) + long int dstp; + long int srcp; + size_t len; +{ + op_t a0, a1, a2; + int sh_1, sh_2; + + /* Calculate how to shift a word read at the memory operation + aligned srcp to make it aligned for copy. */ + + sh_1 = 8 * (srcp % OPSIZ); + sh_2 = 8 * OPSIZ - sh_1; + + /* Make SRCP aligned by rounding it down to the beginning of the `op_t' + it points in the middle of. */ + srcp &= -OPSIZ; + a0 = ((op_t *) srcp)[0]; + + if (len & 1) + { + a1 = ((op_t *) srcp)[1]; + ((op_t *) dstp)[0] = MERGE (a0, sh_1, a1, sh_2); + + if (len == 1) + return; + + a0 = a1; + srcp += OPSIZ; + dstp += OPSIZ; + len -= 1; + } + + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, sh_1, a1, sh_2); + ((op_t *) dstp)[1] = MERGE (a1, sh_1, a2, sh_2); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); +} + +/* _wordcopy_bwd_aligned -- Copy block finishing right before + SRCP to block finishing right before DSTP with LEN `op_t' words + (not LEN bytes!). Both SRCP and DSTP should be aligned for memory + operations on `op_t's. */ + +void +_wordcopy_bwd_aligned (dstp, srcp, len) + long int dstp; + long int srcp; + size_t len; +{ + op_t a0, a1; + + if (len & 1) + { + srcp -= OPSIZ; + dstp -= OPSIZ; + ((op_t *) dstp)[0] = ((op_t *) srcp)[0]; + + if (len == 1) + return; + len -= 1; + } + + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = a1; + ((op_t *) dstp)[0] = a0; + + len -= 2; + } + while (len != 0); +} + +/* _wordcopy_bwd_dest_aligned -- Copy block finishing right + before SRCP to block finishing right before DSTP with LEN `op_t' + words (not LEN bytes!). DSTP should be aligned for memory + operations on `op_t', but SRCP must *not* be aligned. */ + +void +_wordcopy_bwd_dest_aligned (dstp, srcp, len) + long int dstp; + long int srcp; + size_t len; +{ + op_t a0, a1, a2; + int sh_1, sh_2; + + /* Calculate how to shift a word read at the memory operation + aligned srcp to make it aligned for copy. */ + + sh_1 = 8 * (srcp % OPSIZ); + sh_2 = 8 * OPSIZ - sh_1; + + /* Make srcp aligned by rounding it down to the beginning of the op_t + it points in the middle of. */ + srcp &= -OPSIZ; + a2 = ((op_t *) srcp)[0]; + + if (len & 1) + { + srcp -= OPSIZ; + dstp -= OPSIZ; + a1 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[0] = MERGE (a1, sh_1, a2, sh_2); + + if (len == 1) + return; + + a2 = a1; + len -= 1; + } + + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, sh_1, a2, sh_2); + ((op_t *) dstp)[0] = MERGE (a0, sh_1, a1, sh_2); + a2 = a0; + + len -= 2; + } + while (len != 0); +} |