diff options
Diffstat (limited to 'powerpc-cpu/sysdeps/powerpc/powerpc64')
32 files changed, 4095 insertions, 0 deletions
diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/970/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/970/Implies new file mode 100644 index 0000000000..ac431fa96e --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/970/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4 diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/970/fpu/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/970/fpu/Implies new file mode 100644 index 0000000000..558a5fb174 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/970/fpu/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4/fpu diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/Makefile b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/Makefile new file mode 100644 index 0000000000..60aa508ba4 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/Makefile @@ -0,0 +1,6 @@ +# Makefile fragment for POWER4/5/5+. + +ifeq ($(subdir),string) +CFLAGS-wordcopy.c += --param max-variable-expansions-in-unroller=2 --param max-unroll-times=2 -funroll-loops -fpeel-loops -ftree-loop-linear +CFLAGS-memmove.c += --param max-variable-expansions-in-unroller=2 --param max-unroll-times=2 -funroll-loops -fpeel-loops -ftree-loop-linear +endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/Makefile b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/Makefile new file mode 100644 index 0000000000..89dfa5ef35 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/Makefile @@ -0,0 +1,5 @@ +# Makefile fragment for POWER4/5/5+ platforms with FPU. + +ifeq ($(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/powerpc64/power4/fpu/mpa.c b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/mpa.c new file mode 100644 index 0000000000..4a232e27bf --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/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/powerpc64/power4/fpu/slowpow.c b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/slowpow.c new file mode 100644 index 0000000000..ad147a89a6 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/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/powerpc64/power4/memcmp.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcmp.S new file mode 100644 index 0000000000..8f74ca7044 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcmp.S @@ -0,0 +1,981 @@ +/* Optimized strcmp implementation for PowerPC64. + Copyright (C) 2003, 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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., 51 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#include <sysdep.h> +#include <bp-sym.h> +#include <bp-asm.h> + +/* int [r3] memcmp (const char *s1 [r3], const char *s2 [r4], size_t size [r5]) */ + +EALIGN (BP_SYM(memcmp), 4, 0) + CALL_MCOUNT 3 + +#define rTMP r0 +#define rRTN r3 +#define rSTR1 r3 /* first string arg */ +#define rSTR2 r4 /* second string arg */ +#define rN r5 /* max string length */ +/* Note: The Bounded pointer support in this code is broken. This code + was inherited from PPC32 and and that support was never completed. + Current PPC gcc does not support -fbounds-check or -fbounded-pointers. */ +#define rWORD1 r6 /* current word in s1 */ +#define rWORD2 r7 /* current word in s2 */ +#define rWORD3 r8 /* next word in s1 */ +#define rWORD4 r9 /* next word in s2 */ +#define rWORD5 r10 /* next word in s1 */ +#define rWORD6 r11 /* next word in s2 */ +#define rBITDIF r12 /* bits that differ in s1 & s2 words */ +#define rWORD7 r30 /* next word in s1 */ +#define rWORD8 r31 /* next word in s2 */ + + xor rTMP, rSTR2, rSTR1 + cmpldi cr6, rN, 0 + cmpldi cr1, rN, 12 + clrldi. rTMP, rTMP, 61 + clrldi rBITDIF, rSTR1, 61 + cmpldi cr5, rBITDIF, 0 + beq- cr6, L(zeroLength) + dcbt 0,rSTR1 + dcbt 0,rSTR2 +/* If less than 8 bytes or not aligned, use the unalligned + byte loop. */ + blt cr1, L(bytealigned) + std rWORD8,-8(r1) + cfi_offset(rWORD8,-8) + std rWORD7,-16(r1) + cfi_offset(rWORD7,-16) + bne L(unaligned) +/* At this point we know both strings have the same alignment and the + compare length is at least 8 bytes. rBITDIF containes the low order + 3 bits of rSTR1 and cr5 contains the result of the logical compare + of rBITDIF to 0. If rBITDIF == 0 then we are already double word + aligned and can perform the DWaligned loop. + + Otherwise we know the two strings have the same alignment (but not + yet DW). So we can force the string addresses to the next lower DW + boundary and special case this first DW word using shift left to + ellimiate bits preceeding the first byte. Since we want to join the + normal (DWaligned) compare loop, starting at the second double word, + we need to adjust the length (rN) and special case the loop + versioning for the first DW. This insures that the loop count is + correct and the first DW (shifted) is in the expected resister pair. */ + .align 4 +L(samealignment): + clrrdi rSTR1, rSTR1, 3 + clrrdi rSTR2, rSTR2, 3 + beq cr5, L(DWaligned) + add rN, rN, rBITDIF + sldi r11, rBITDIF, 3 + srdi rTMP, rN, 5 /* Divide by 32 */ + andi. rBITDIF, rN, 24 /* Get the DW remainder */ + ld rWORD1, 0(rSTR1) + ld rWORD2, 0(rSTR2) + cmpldi cr1, rBITDIF, 16 + cmpldi cr7, rN, 32 + clrldi rN, rN, 61 + beq L(dPs4) + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + bgt cr1, L(dPs3) + beq cr1, L(dPs2) + +/* Remainder is 8 */ + .align 3 +L(dsP1): + sld rWORD5, rWORD1, r11 + sld rWORD6, rWORD2, r11 + cmpld cr5, rWORD5, rWORD6 + blt cr7, L(dP1x) +/* Do something useful in this cycle since we have to branch anyway. */ + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + cmpld cr0, rWORD1, rWORD2 + b L(dP1e) +/* Remainder is 16 */ + .align 4 +L(dPs2): + sld rWORD5, rWORD1, r11 + sld rWORD6, rWORD2, r11 + cmpld cr6, rWORD5, rWORD6 + blt cr7, L(dP2x) +/* Do something useful in this cycle since we have to branch anyway. */ + ld rWORD7, 8(rSTR1) + ld rWORD8, 8(rSTR2) + cmpld cr5, rWORD7, rWORD8 + b L(dP2e) +/* Remainder is 24 */ + .align 4 +L(dPs3): + sld rWORD3, rWORD1, r11 + sld rWORD4, rWORD2, r11 + cmpld cr1, rWORD3, rWORD4 + b L(dP3e) +/* Count is a multiple of 32, remainder is 0 */ + .align 4 +L(dPs4): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + sld rWORD1, rWORD1, r11 + sld rWORD2, rWORD2, r11 + cmpld cr0, rWORD1, rWORD2 + b L(dP4e) + +/* At this point we know both strings are double word aligned and the + compare length is at least 8 bytes. */ + .align 4 +L(DWaligned): + andi. rBITDIF, rN, 24 /* Get the DW remainder */ + srdi rTMP, rN, 5 /* Divide by 32 */ + cmpldi cr1, rBITDIF, 16 + cmpldi cr7, rN, 32 + clrldi rN, rN, 61 + beq L(dP4) + bgt cr1, L(dP3) + beq cr1, L(dP2) + +/* Remainder is 8 */ + .align 4 +L(dP1): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ +/* Normally we'd use rWORD7/rWORD8 here, but since we might exit early + (8-15 byte compare), we want to use only volitile registers. This + means we can avoid restoring non-volitile registers since we did not + change any on the early exit path. The key here is the non-early + exit path only cares about the condition code (cr5), not about which + register pair was used. */ + ld rWORD5, 0(rSTR1) + ld rWORD6, 0(rSTR2) + cmpld cr5, rWORD5, rWORD6 + blt cr7, L(dP1x) + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + cmpld cr0, rWORD1, rWORD2 +L(dP1e): + ld rWORD3, 16(rSTR1) + ld rWORD4, 16(rSTR2) + cmpld cr1, rWORD3, rWORD4 + ld rWORD5, 24(rSTR1) + ld rWORD6, 24(rSTR2) + cmpld cr6, rWORD5, rWORD6 + bne cr5, L(dLcr5) + bne cr0, L(dLcr0) + + ldu rWORD7, 32(rSTR1) + ldu rWORD8, 32(rSTR2) + bne cr1, L(dLcr1) + cmpld cr5, rWORD7, rWORD8 + bdnz L(dLoop) + bne cr6, L(dLcr6) + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + .align 3 +L(dP1x): + sldi. r12, rN, 3 + bne cr5, L(dLcr5) + subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */ + bne L(d00) + li rRTN, 0 + blr + +/* Remainder is 16 */ + .align 4 +L(dP2): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + ld rWORD5, 0(rSTR1) + ld rWORD6, 0(rSTR2) + cmpld cr6, rWORD5, rWORD6 + blt cr7, L(dP2x) + ld rWORD7, 8(rSTR1) + ld rWORD8, 8(rSTR2) + cmpld cr5, rWORD7, rWORD8 +L(dP2e): + ld rWORD1, 16(rSTR1) + ld rWORD2, 16(rSTR2) + cmpld cr0, rWORD1, rWORD2 + ld rWORD3, 24(rSTR1) + ld rWORD4, 24(rSTR2) + cmpld cr1, rWORD3, rWORD4 + addi rSTR1, rSTR1, 8 + addi rSTR2, rSTR2, 8 + bne cr6, L(dLcr6) + bne cr5, L(dLcr5) + b L(dLoop2) +/* Again we are on a early exit path (16-23 byte compare), we want to + only use volitile registers and avoid restoring non-volitile + registers. */ + .align 4 +L(dP2x): + ld rWORD3, 8(rSTR1) + ld rWORD4, 8(rSTR2) + cmpld cr5, rWORD3, rWORD4 + sldi. r12, rN, 3 + bne cr6, L(dLcr6) + addi rSTR1, rSTR1, 8 + addi rSTR2, rSTR2, 8 + bne cr5, L(dLcr5) + subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */ + bne L(d00) + li rRTN, 0 + blr + +/* Remainder is 24 */ + .align 4 +L(dP3): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + ld rWORD3, 0(rSTR1) + ld rWORD4, 0(rSTR2) + cmpld cr1, rWORD3, rWORD4 +L(dP3e): + ld rWORD5, 8(rSTR1) + ld rWORD6, 8(rSTR2) + cmpld cr6, rWORD5, rWORD6 + blt cr7, L(dP3x) + ld rWORD7, 16(rSTR1) + ld rWORD8, 16(rSTR2) + cmpld cr5, rWORD7, rWORD8 + ld rWORD1, 24(rSTR1) + ld rWORD2, 24(rSTR2) + cmpld cr0, rWORD1, rWORD2 + addi rSTR1, rSTR1, 16 + addi rSTR2, rSTR2, 16 + bne cr1, L(dLcr1) + bne cr6, L(dLcr6) + b L(dLoop1) +/* Again we are on a early exit path (24-31 byte compare), we want to + only use volitile registers and avoid restoring non-volitile + registers. */ + .align 4 +L(dP3x): + ld rWORD1, 16(rSTR1) + ld rWORD2, 16(rSTR2) + cmpld cr5, rWORD1, rWORD2 + sldi. r12, rN, 3 + bne cr1, L(dLcr1) + addi rSTR1, rSTR1, 16 + addi rSTR2, rSTR2, 16 + bne cr6, L(dLcr6) + subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */ + bne cr5, L(dLcr5) + bne L(d00) + li rRTN, 0 + blr + +/* Count is a multiple of 32, remainder is 0 */ + .align 4 +L(dP4): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + ld rWORD1, 0(rSTR1) + ld rWORD2, 0(rSTR2) + cmpld cr0, rWORD1, rWORD2 +L(dP4e): + ld rWORD3, 8(rSTR1) + ld rWORD4, 8(rSTR2) + cmpld cr1, rWORD3, rWORD4 + ld rWORD5, 16(rSTR1) + ld rWORD6, 16(rSTR2) + cmpld cr6, rWORD5, rWORD6 + ldu rWORD7, 24(rSTR1) + ldu rWORD8, 24(rSTR2) + cmpld cr5, rWORD7, rWORD8 + bne cr0, L(dLcr0) + bne cr1, L(dLcr1) + bdz- L(d24) /* Adjust CTR as we start with +4 */ +/* This is the primary loop */ + .align 4 +L(dLoop): + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + cmpld cr1, rWORD3, rWORD4 + bne cr6, L(dLcr6) +L(dLoop1): + ld rWORD3, 16(rSTR1) + ld rWORD4, 16(rSTR2) + cmpld cr6, rWORD5, rWORD6 + bne cr5, L(dLcr5) +L(dLoop2): + ld rWORD5, 24(rSTR1) + ld rWORD6, 24(rSTR2) + cmpld cr5, rWORD7, rWORD8 + bne cr0, L(dLcr0) +L(dLoop3): + ldu rWORD7, 32(rSTR1) + ldu rWORD8, 32(rSTR2) + bne- cr1, L(dLcr1) + cmpld cr0, rWORD1, rWORD2 + bdnz+ L(dLoop) + +L(dL4): + cmpld cr1, rWORD3, rWORD4 + bne cr6, L(dLcr6) + cmpld cr6, rWORD5, rWORD6 + bne cr5, L(dLcr5) + cmpld cr5, rWORD7, rWORD8 +L(d44): + bne cr0, L(dLcr0) +L(d34): + bne cr1, L(dLcr1) +L(d24): + bne cr6, L(dLcr6) +L(d14): + sldi. r12, rN, 3 + bne cr5, L(dLcr5) +L(d04): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */ + beq L(zeroLength) +/* At this point we have a remainder of 1 to 7 bytes to compare. Since + we are aligned it is safe to load the whole double word, and use + shift right double to elliminate bits beyond the compare length. */ +L(d00): + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + srd rWORD1, rWORD1, rN + srd rWORD2, rWORD2, rN + cmpld cr5, rWORD1, rWORD2 + bne cr5, L(dLcr5x) + li rRTN, 0 + blr + .align 4 +L(dLcr0): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgtlr cr0 + li rRTN, -1 + blr + .align 4 +L(dLcr1): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgtlr cr1 + li rRTN, -1 + blr + .align 4 +L(dLcr6): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgtlr cr6 + li rRTN, -1 + blr + .align 4 +L(dLcr5): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) +L(dLcr5x): + li rRTN, 1 + bgtlr cr5 + li rRTN, -1 + blr + + .align 4 +L(bytealigned): + mtctr rN /* Power4 wants mtctr 1st in dispatch group */ + beq- cr6, L(zeroLength) + +/* We need to prime this loop. This loop is swing modulo scheduled + to avoid pipe delays. The dependent instruction latencies (load to + compare to conditional branch) is 2 to 3 cycles. In this loop each + dispatch group ends in a branch and takes 1 cycle. Effectively + the first iteration of the loop only serves to load operands and + branches based on compares are delayed until the next loop. + + So we must precondition some registers and condition codes so that + we don't exit the loop early on the first iteration. */ + + lbz rWORD1, 0(rSTR1) + lbz rWORD2, 0(rSTR2) + bdz- L(b11) + cmpld cr0, rWORD1, rWORD2 + lbz rWORD3, 1(rSTR1) + lbz rWORD4, 1(rSTR2) + bdz- L(b12) + cmpld cr1, rWORD3, rWORD4 + lbzu rWORD5, 2(rSTR1) + lbzu rWORD6, 2(rSTR2) + bdz- L(b13) + .align 4 +L(bLoop): + lbzu rWORD1, 1(rSTR1) + lbzu rWORD2, 1(rSTR2) + bne- cr0, L(bLcr0) + + cmpld cr6, rWORD5, rWORD6 + bdz- L(b3i) + + lbzu rWORD3, 1(rSTR1) + lbzu rWORD4, 1(rSTR2) + bne- cr1, L(bLcr1) + + cmpld cr0, rWORD1, rWORD2 + bdz- L(b2i) + + lbzu rWORD5, 1(rSTR1) + lbzu rWORD6, 1(rSTR2) + bne- cr6, L(bLcr6) + + cmpld cr1, rWORD3, rWORD4 + bdnz+ L(bLoop) + +/* We speculatively loading bytes before we have tested the previous + bytes. But we must avoid overrunning the length (in the ctr) to + prevent these speculative loads from causing a segfault. In this + case the loop will exit early (before the all pending bytes are + tested. In this case we must complete the pending operations + before returning. */ +L(b1i): + bne- cr0, L(bLcr0) + bne- cr1, L(bLcr1) + b L(bx56) + .align 4 +L(b2i): + bne- cr6, L(bLcr6) + bne- cr0, L(bLcr0) + b L(bx34) + .align 4 +L(b3i): + bne- cr1, L(bLcr1) + bne- cr6, L(bLcr6) + b L(bx12) + .align 4 +L(bLcr0): + li rRTN, 1 + bgtlr cr0 + li rRTN, -1 + blr +L(bLcr1): + li rRTN, 1 + bgtlr cr1 + li rRTN, -1 + blr +L(bLcr6): + li rRTN, 1 + bgtlr cr6 + li rRTN, -1 + blr + +L(b13): + bne- cr0, L(bx12) + bne- cr1, L(bx34) +L(bx56): + sub rRTN, rWORD5, rWORD6 + blr + nop +L(b12): + bne- cr0, L(bx12) +L(bx34): + sub rRTN, rWORD3, rWORD4 + blr +L(b11): +L(bx12): + sub rRTN, rWORD1, rWORD2 + blr + .align 4 +L(zeroLengthReturn): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) +L(zeroLength): + li rRTN, 0 + blr + + .align 4 +/* At this point we know the strings have different alignment and the + compare length is at least 8 bytes. rBITDIF containes the low order + 3 bits of rSTR1 and cr5 contains the result of the logical compare + of rBITDIF to 0. If rBITDIF == 0 then rStr1 is double word + aligned and can perform the DWunaligned loop. + + Otherwise we know that rSTR1 is not aready DW aligned yet. + So we can force the string addresses to the next lower DW + boundary and special case this first DW word using shift left to + ellimiate bits preceeding the first byte. Since we want to join the + normal (DWaligned) compare loop, starting at the second double word, + we need to adjust the length (rN) and special case the loop + versioning for the first DW. This insures that the loop count is + correct and the first DW (shifted) is in the expected resister pair. */ +#define rSHL r29 /* Unaligned shift left count. */ +#define rSHR r28 /* Unaligned shift right count. */ +#define rB r27 /* Left rotation temp for rWORD2. */ +#define rD r26 /* Left rotation temp for rWORD4. */ +#define rF r25 /* Left rotation temp for rWORD6. */ +#define rH r24 /* Left rotation temp for rWORD8. */ +#define rA r0 /* Right rotation temp for rWORD2. */ +#define rC r12 /* Right rotation temp for rWORD4. */ +#define rE r0 /* Right rotation temp for rWORD6. */ +#define rG r12 /* Right rotation temp for rWORD8. */ +L(unaligned): + std r29,-24(r1) + cfi_offset(r29,-24) + clrldi rSHL, rSTR2, 61 + beq- cr6, L(duzeroLength) + std r28,-32(r1) + cfi_offset(r28,-32) + beq cr5, L(DWunaligned) + std r27,-40(r1) + cfi_offset(r27,-40) +/* Adjust the logical start of rSTR2 ro compensate for the extra bits + in the 1st rSTR1 DW. */ + sub r27, rSTR2, rBITDIF +/* But do not attempt to address the DW before that DW that contains + the actual start of rSTR2. */ + clrrdi rSTR2, rSTR2, 3 + std r26,-48(r1) + cfi_offset(r26,-48) +/* Compute the leaft/right shift counts for the unalign rSTR2, + compensating for the logical (DW aligned) start of rSTR1. */ + clrldi rSHL, r27, 61 + clrrdi rSTR1, rSTR1, 3 + std r25,-56(r1) + cfi_offset(r25,-56) + sldi rSHL, rSHL, 3 + cmpld cr5, r27, rSTR2 + add rN, rN, rBITDIF + sldi r11, rBITDIF, 3 + std r24,-64(r1) + cfi_offset(r24,-64) + subfic rSHR, rSHL, 64 + srdi rTMP, rN, 5 /* Divide by 32 */ + andi. rBITDIF, rN, 24 /* Get the DW remainder */ +/* We normally need to load 2 DWs to start the unaligned rSTR2, but in + this special case those bits may be discarded anyway. Also we + must avoid loading a DW where none of the bits are part of rSTR2 as + this may cross a page boundary and cause a page fault. */ + li rWORD8, 0 + blt cr5, L(dus0) + ld rWORD8, 0(rSTR2) + la rSTR2, 8(rSTR2) + sld rWORD8, rWORD8, rSHL + +L(dus0): + ld rWORD1, 0(rSTR1) + ld rWORD2, 0(rSTR2) + cmpldi cr1, rBITDIF, 16 + cmpldi cr7, rN, 32 + srd rG, rWORD2, rSHR + clrldi rN, rN, 61 + beq L(duPs4) + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + or rWORD8, rG, rWORD8 + bgt cr1, L(duPs3) + beq cr1, L(duPs2) + +/* Remainder is 8 */ + .align 4 +L(dusP1): + sld rB, rWORD2, rSHL + sld rWORD7, rWORD1, r11 + sld rWORD8, rWORD8, r11 + bge cr7, L(duP1e) +/* At this point we exit early with the first double word compare + complete and remainder of 0 to 7 bytes. See L(du14) for details on + how we handle the remaining bytes. */ + cmpld cr5, rWORD7, rWORD8 + sldi. rN, rN, 3 + bne cr5, L(duLcr5) + cmpld cr7, rN, rSHR + beq L(duZeroReturn) + li rA, 0 + ble cr7, L(dutrim) + ld rWORD2, 8(rSTR2) + srd rA, rWORD2, rSHR + b L(dutrim) +/* Remainder is 16 */ + .align 4 +L(duPs2): + sld rH, rWORD2, rSHL + sld rWORD5, rWORD1, r11 + sld rWORD6, rWORD8, r11 + b L(duP2e) +/* Remainder is 24 */ + .align 4 +L(duPs3): + sld rF, rWORD2, rSHL + sld rWORD3, rWORD1, r11 + sld rWORD4, rWORD8, r11 + b L(duP3e) +/* Count is a multiple of 32, remainder is 0 */ + .align 4 +L(duPs4): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + or rWORD8, rG, rWORD8 + sld rD, rWORD2, rSHL + sld rWORD1, rWORD1, r11 + sld rWORD2, rWORD8, r11 + b L(duP4e) + +/* At this point we know rSTR1 is double word aligned and the + compare length is at least 8 bytes. */ + .align 4 +L(DWunaligned): + std r27,-40(r1) + cfi_offset(r27,-40) + clrrdi rSTR2, rSTR2, 3 + std r26,-48(r1) + cfi_offset(r26,-48) + srdi rTMP, rN, 5 /* Divide by 32 */ + std r25,-56(r1) + cfi_offset(r25,-56) + andi. rBITDIF, rN, 24 /* Get the DW remainder */ + std r24,-64(r1) + cfi_offset(r24,-64) + sldi rSHL, rSHL, 3 + ld rWORD6, 0(rSTR2) + ldu rWORD8, 8(rSTR2) + cmpldi cr1, rBITDIF, 16 + cmpldi cr7, rN, 32 + clrldi rN, rN, 61 + subfic rSHR, rSHL, 64 + sld rH, rWORD6, rSHL + beq L(duP4) + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + bgt cr1, L(duP3) + beq cr1, L(duP2) + +/* Remainder is 8 */ + .align 4 +L(duP1): + srd rG, rWORD8, rSHR + ld rWORD7, 0(rSTR1) + sld rB, rWORD8, rSHL + or rWORD8, rG, rH + blt cr7, L(duP1x) +L(duP1e): + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + cmpld cr5, rWORD7, rWORD8 + srd rA, rWORD2, rSHR + sld rD, rWORD2, rSHL + or rWORD2, rA, rB + ld rWORD3, 16(rSTR1) + ld rWORD4, 16(rSTR2) + cmpld cr0, rWORD1, rWORD2 + srd rC, rWORD4, rSHR + sld rF, rWORD4, rSHL + bne cr5, L(duLcr5) + or rWORD4, rC, rD + ld rWORD5, 24(rSTR1) + ld rWORD6, 24(rSTR2) + cmpld cr1, rWORD3, rWORD4 + srd rE, rWORD6, rSHR + sld rH, rWORD6, rSHL + bne cr0, L(duLcr0) + or rWORD6, rE, rF + cmpld cr6, rWORD5, rWORD6 + b L(duLoop3) + .align 4 +/* At this point we exit early with the first double word compare + complete and remainder of 0 to 7 bytes. See L(du14) for details on + how we handle the remaining bytes. */ +L(duP1x): + cmpld cr5, rWORD7, rWORD8 + sldi. rN, rN, 3 + bne cr5, L(duLcr5) + cmpld cr7, rN, rSHR + beq L(duZeroReturn) + li rA, 0 + ble cr7, L(dutrim) + ld rWORD2, 8(rSTR2) + srd rA, rWORD2, rSHR + b L(dutrim) +/* Remainder is 16 */ + .align 4 +L(duP2): + srd rE, rWORD8, rSHR + ld rWORD5, 0(rSTR1) + or rWORD6, rE, rH + sld rH, rWORD8, rSHL +L(duP2e): + ld rWORD7, 8(rSTR1) + ld rWORD8, 8(rSTR2) + cmpld cr6, rWORD5, rWORD6 + srd rG, rWORD8, rSHR + sld rB, rWORD8, rSHL + or rWORD8, rG, rH + blt cr7, L(duP2x) + ld rWORD1, 16(rSTR1) + ld rWORD2, 16(rSTR2) + cmpld cr5, rWORD7, rWORD8 + bne cr6, L(duLcr6) + srd rA, rWORD2, rSHR + sld rD, rWORD2, rSHL + or rWORD2, rA, rB + ld rWORD3, 24(rSTR1) + ld rWORD4, 24(rSTR2) + cmpld cr0, rWORD1, rWORD2 + bne cr5, L(duLcr5) + srd rC, rWORD4, rSHR + sld rF, rWORD4, rSHL + or rWORD4, rC, rD + addi rSTR1, rSTR1, 8 + addi rSTR2, rSTR2, 8 + cmpld cr1, rWORD3, rWORD4 + b L(duLoop2) + .align 4 +L(duP2x): + cmpld cr5, rWORD7, rWORD8 + addi rSTR1, rSTR1, 8 + addi rSTR2, rSTR2, 8 + bne cr6, L(duLcr6) + sldi. rN, rN, 3 + bne cr5, L(duLcr5) + cmpld cr7, rN, rSHR + beq L(duZeroReturn) + li rA, 0 + ble cr7, L(dutrim) + ld rWORD2, 8(rSTR2) + srd rA, rWORD2, rSHR + b L(dutrim) + +/* Remainder is 24 */ + .align 4 +L(duP3): + srd rC, rWORD8, rSHR + ld rWORD3, 0(rSTR1) + sld rF, rWORD8, rSHL + or rWORD4, rC, rH +L(duP3e): + ld rWORD5, 8(rSTR1) + ld rWORD6, 8(rSTR2) + cmpld cr1, rWORD3, rWORD4 + srd rE, rWORD6, rSHR + sld rH, rWORD6, rSHL + or rWORD6, rE, rF + ld rWORD7, 16(rSTR1) + ld rWORD8, 16(rSTR2) + cmpld cr6, rWORD5, rWORD6 + bne cr1, L(duLcr1) + srd rG, rWORD8, rSHR + sld rB, rWORD8, rSHL + or rWORD8, rG, rH + blt cr7, L(duP3x) + ld rWORD1, 24(rSTR1) + ld rWORD2, 24(rSTR2) + cmpld cr5, rWORD7, rWORD8 + bne cr6, L(duLcr6) + srd rA, rWORD2, rSHR + sld rD, rWORD2, rSHL + or rWORD2, rA, rB + addi rSTR1, rSTR1, 16 + addi rSTR2, rSTR2, 16 + cmpld cr0, rWORD1, rWORD2 + b L(duLoop1) + .align 4 +L(duP3x): + addi rSTR1, rSTR1, 16 + addi rSTR2, rSTR2, 16 + bne cr1, L(duLcr1) + cmpld cr5, rWORD7, rWORD8 + bne cr6, L(duLcr6) + sldi. rN, rN, 3 + bne cr5, L(duLcr5) + cmpld cr7, rN, rSHR + beq L(duZeroReturn) + li rA, 0 + ble cr7, L(dutrim) + ld rWORD2, 8(rSTR2) + srd rA, rWORD2, rSHR + b L(dutrim) + +/* Count is a multiple of 32, remainder is 0 */ + .align 4 +L(duP4): + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group */ + srd rA, rWORD8, rSHR + ld rWORD1, 0(rSTR1) + sld rD, rWORD8, rSHL + or rWORD2, rA, rH +L(duP4e): + ld rWORD3, 8(rSTR1) + ld rWORD4, 8(rSTR2) + cmpld cr0, rWORD1, rWORD2 + srd rC, rWORD4, rSHR + sld rF, rWORD4, rSHL + or rWORD4, rC, rD + ld rWORD5, 16(rSTR1) + ld rWORD6, 16(rSTR2) + cmpld cr1, rWORD3, rWORD4 + bne cr0, L(duLcr0) + srd rE, rWORD6, rSHR + sld rH, rWORD6, rSHL + or rWORD6, rE, rF + ldu rWORD7, 24(rSTR1) + ldu rWORD8, 24(rSTR2) + cmpld cr6, rWORD5, rWORD6 + bne cr1, L(duLcr1) + srd rG, rWORD8, rSHR + sld rB, rWORD8, rSHL + or rWORD8, rG, rH + cmpld cr5, rWORD7, rWORD8 + bdz- L(du24) /* Adjust CTR as we start with +4 */ +/* This is the primary loop */ + .align 4 +L(duLoop): + ld rWORD1, 8(rSTR1) + ld rWORD2, 8(rSTR2) + cmpld cr1, rWORD3, rWORD4 + bne cr6, L(duLcr6) + srd rA, rWORD2, rSHR + sld rD, rWORD2, rSHL + or rWORD2, rA, rB +L(duLoop1): + ld rWORD3, 16(rSTR1) + ld rWORD4, 16(rSTR2) + cmpld cr6, rWORD5, rWORD6 + bne cr5, L(duLcr5) + srd rC, rWORD4, rSHR + sld rF, rWORD4, rSHL + or rWORD4, rC, rD +L(duLoop2): + ld rWORD5, 24(rSTR1) + ld rWORD6, 24(rSTR2) + cmpld cr5, rWORD7, rWORD8 + bne cr0, L(duLcr0) + srd rE, rWORD6, rSHR + sld rH, rWORD6, rSHL + or rWORD6, rE, rF +L(duLoop3): + ldu rWORD7, 32(rSTR1) + ldu rWORD8, 32(rSTR2) + cmpld cr0, rWORD1, rWORD2 + bne- cr1, L(duLcr1) + srd rG, rWORD8, rSHR + sld rB, rWORD8, rSHL + or rWORD8, rG, rH + bdnz+ L(duLoop) + +L(duL4): + bne cr1, L(duLcr1) + cmpld cr1, rWORD3, rWORD4 + bne cr6, L(duLcr6) + cmpld cr6, rWORD5, rWORD6 + bne cr5, L(duLcr5) + cmpld cr5, rWORD7, rWORD8 +L(du44): + bne cr0, L(duLcr0) +L(du34): + bne cr1, L(duLcr1) +L(du24): + bne cr6, L(duLcr6) +L(du14): + sldi. rN, rN, 3 + bne cr5, L(duLcr5) +/* At this point we have a remainder of 1 to 7 bytes to compare. We use + shift right double to elliminate bits beyond the compare length. + This allows the use of double word subtract to compute the final + result. + + However it may not be safe to load rWORD2 which may be beyond the + string length. So we compare the bit length of the remainder to + the right shift count (rSHR). If the bit count is less than or equal + we do not need to load rWORD2 (all significant bits are already in + rB). */ + cmpld cr7, rN, rSHR + beq L(duZeroReturn) + li rA, 0 + ble cr7, L(dutrim) + ld rWORD2, 8(rSTR2) + srd rA, rWORD2, rSHR + .align 4 +L(dutrim): + ld rWORD1, 8(rSTR1) + ld rWORD8,-8(r1) + subfic rN, rN, 64 /* Shift count is 64 - (rN * 8). */ + or rWORD2, rA, rB + ld rWORD7,-16(r1) + ld r29,-24(r1) + srd rWORD1, rWORD1, rN + srd rWORD2, rWORD2, rN + ld r28,-32(r1) + ld r27,-40(r1) + li rRTN, 0 + cmpld cr0, rWORD1, rWORD2 + ld r26,-48(r1) + ld r25,-56(r1) + beq cr0, L(dureturn24) + li rRTN, 1 + ld r24,-64(r1) + bgtlr cr0 + li rRTN, -1 + blr + .align 4 +L(duLcr0): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgt cr0, L(dureturn29) + ld r29,-24(r1) + ld r28,-32(r1) + li rRTN, -1 + b L(dureturn27) + .align 4 +L(duLcr1): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgt cr1, L(dureturn29) + ld r29,-24(r1) + ld r28,-32(r1) + li rRTN, -1 + b L(dureturn27) + .align 4 +L(duLcr6): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgt cr6, L(dureturn29) + ld r29,-24(r1) + ld r28,-32(r1) + li rRTN, -1 + b L(dureturn27) + .align 4 +L(duLcr5): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) + li rRTN, 1 + bgt cr5, L(dureturn29) + ld r29,-24(r1) + ld r28,-32(r1) + li rRTN, -1 + b L(dureturn27) + .align 3 +L(duZeroReturn): + li rRTN,0 + .align 4 +L(dureturn): + ld rWORD8,-8(r1) + ld rWORD7,-16(r1) +L(dureturn29): + ld r29,-24(r1) + ld r28,-32(r1) +L(dureturn27): + ld r27,-40(r1) +L(dureturn26): + ld r26,-48(r1) +L(dureturn25): + ld r25,-56(r1) +L(dureturn24): + ld r24,-64(r1) + blr +L(duzeroLength): + li rRTN,0 + blr + +END (BP_SYM (memcmp)) +libc_hidden_builtin_def (memcmp) +weak_alias (memcmp, bcmp) diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcopy.h b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcopy.h new file mode 100644 index 0000000000..9a4ff79f4a --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcopy.h @@ -0,0 +1 @@ +#include "../../powerpc32/power4/memcopy.h" diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcpy.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcpy.S new file mode 100644 index 0000000000..9910ebda82 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/memcpy.S @@ -0,0 +1,417 @@ +/* Optimized memcpy implementation for PowerPC64. + Copyright (C) 2003, 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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., 51 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#include <sysdep.h> +#include <bp-sym.h> +#include <bp-asm.h> + +/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]); + Returns 'dst'. + + Memcpy handles short copies (< 32-bytes) using a binary move blocks + (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled + with the appropriate combination of byte and halfword load/stores. + There is minimal effort to optimize the alignment of short moves. + The 64-bit implementations of POWER3 and POWER4 do a reasonable job + of handling unligned load/stores that do not cross 32-byte boundries. + + Longer moves (>= 32-bytes) justify the effort to get at least the + destination doubleword (8-byte) aligned. Further optimization is + posible when both source and destination are doubleword aligned. + Each case has a optimized unrolled loop. */ + +EALIGN (BP_SYM (memcpy), 5, 0) + CALL_MCOUNT 3 + + cmpldi cr1,5,31 + neg 0,3 + std 3,-16(1) + std 31,-8(1) + cfi_offset(31,-8) + andi. 11,3,7 /* check alignement of dst. */ + clrldi 0,0,61 /* Number of bytes until the 1st doubleword of dst. */ + clrldi 10,4,61 /* check alignement of src. */ + cmpldi cr6,5,8 + ble- cr1,.L2 /* If move < 32 bytes use short move code. */ + cmpld cr6,10,11 + mr 12,4 + srdi 9,5,3 /* Number of full double words remaining. */ + mtcrf 0x01,0 + mr 31,5 + beq .L0 + + subf 31,0,5 + /* Move 0-7 bytes as needed to get the destination doubleword alligned. */ +1: bf 31,2f + lbz 6,0(12) + addi 12,12,1 + stb 6,0(3) + addi 3,3,1 +2: bf 30,4f + lhz 6,0(12) + addi 12,12,2 + sth 6,0(3) + addi 3,3,2 +4: bf 29,0f + lwz 6,0(12) + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +0: + clrldi 10,12,61 /* check alignement of src again. */ + srdi 9,31,3 /* Number of full double words remaining. */ + + /* Copy doublewords from source to destination, assumpting the + destination is aligned on a doubleword boundary. + + At this point we know there are at least 25 bytes left (32-7) to copy. + The next step is to determine if the source is also doubleword aligned. + If not branch to the unaligned move code at .L6. which uses + a load, shift, store strategy. + + Otherwise source and destination are doubleword aligned, and we can + the optimized doubleword copy loop. */ +.L0: + clrldi 11,31,61 + mtcrf 0x01,9 + cmpldi cr1,11,0 + bne- cr6,.L6 /* If source is not DW aligned. */ + + /* Move doublewords where destination and source are DW aligned. + Use a unrolled loop to copy 4 doubleword (32-bytes) per iteration. + If the the copy is not an exact multiple of 32 bytes, 1-3 + doublewords are copied as needed to set up the main loop. After + the main loop exits there may be a tail of 1-7 bytes. These byte are + copied a word/halfword/byte at a time as needed to preserve alignment. */ + + srdi 8,31,5 + cmpldi cr1,9,4 + cmpldi cr6,11,0 + mr 11,12 + + bf 30,1f + ld 6,0(12) + ld 7,8(12) + addi 11,12,16 + mtctr 8 + std 6,0(3) + std 7,8(3) + addi 10,3,16 + bf 31,4f + ld 0,16(12) + std 0,16(3) + blt cr1,3f + addi 11,12,24 + addi 10,3,24 + b 4f + .align 4 +1: + mr 10,3 + mtctr 8 + bf 31,4f + ld 6,0(12) + addi 11,12,8 + std 6,0(3) + addi 10,3,8 + + .align 4 +4: + ld 6,0(11) + ld 7,8(11) + ld 8,16(11) + ld 0,24(11) + addi 11,11,32 +2: + std 6,0(10) + std 7,8(10) + std 8,16(10) + std 0,24(10) + addi 10,10,32 + bdnz 4b +3: + + rldicr 0,31,0,60 + mtcrf 0x01,31 + beq cr6,0f +.L9: + add 3,3,0 + add 12,12,0 + +/* At this point we have a tail of 0-7 bytes and we know that the + destiniation is double word aligned. */ +4: bf 29,2f + lwz 6,0(12) + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +2: bf 30,1f + lhz 6,0(12) + addi 12,12,2 + sth 6,0(3) + addi 3,3,2 +1: bf 31,0f + lbz 6,0(12) + stb 6,0(3) +0: + /* Return original dst pointer. */ + ld 31,-8(1) + ld 3,-16(1) + blr + +/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31 + bytes. Each case is handled without loops, using binary (1,2,4,8) + tests. + + In the short (0-8 byte) case no attempt is made to force alignment + of either source or destination. The hardware will handle the + unaligned load/stores with small delays for crossing 32- 64-byte, and + 4096-byte boundaries. Since these short moves are unlikely to be + unaligned or cross these boundaries, the overhead to force + alignment is not justified. + + The longer (9-31 byte) move is more likely to cross 32- or 64-byte + boundaries. Since only loads are sensitive to the 32-/64-byte + boundaries it is more important to align the source then the + destination. If the source is not already word aligned, we first + move 1-3 bytes as needed. Since we are only word aligned we don't + use double word load/stores to insure that all loads are aligned. + While the destination and stores may still be unaligned, this + is only an issue for page (4096 byte boundary) crossing, which + should be rare for these short moves. The hardware handles this + case automatically with a small delay. */ + + .align 4 +.L2: + mtcrf 0x01,5 + neg 8,4 + clrrdi 11,4,2 + andi. 0,8,3 + ble cr6,.LE8 /* Handle moves of 0-8 bytes. */ +/* At least 9 bytes left. Get the source word aligned. */ + cmpldi cr1,5,16 + mr 10,5 + mr 12,4 + cmpldi cr6,0,2 + beq .L3 /* If the source is already word aligned skip this. */ +/* Copy 1-3 bytes to get source address word aligned. */ + lwz 6,0(11) + subf 10,0,5 + add 12,4,0 + blt cr6,5f + srdi 7,6,16 + bgt cr6,3f + sth 6,0(3) + b 7f + .align 4 +3: + stb 7,0(3) + sth 6,1(3) + b 7f + .align 4 +5: + stb 6,0(3) +7: + cmpldi cr1,10,16 + add 3,3,0 + mtcrf 0x01,10 + .align 4 +.L3: +/* At least 6 bytes left and the source is word aligned. */ + blt cr1,8f +16: /* Move 16 bytes. */ + lwz 6,0(12) + lwz 7,4(12) + stw 6,0(3) + lwz 6,8(12) + stw 7,4(3) + lwz 7,12(12) + addi 12,12,16 + stw 6,8(3) + stw 7,12(3) + addi 3,3,16 +8: /* Move 8 bytes. */ + bf 28,4f + lwz 6,0(12) + lwz 7,4(12) + addi 12,12,8 + stw 6,0(3) + stw 7,4(3) + addi 3,3,8 +4: /* Move 4 bytes. */ + bf 29,2f + lwz 6,0(12) + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +2: /* Move 2-3 bytes. */ + bf 30,1f + lhz 6,0(12) + sth 6,0(3) + bf 31,0f + lbz 7,2(12) + stb 7,2(3) + ld 3,-16(1) + blr +1: /* Move 1 byte. */ + bf 31,0f + lbz 6,0(12) + stb 6,0(3) +0: + /* Return original dst pointer. */ + ld 3,-16(1) + blr + +/* Special case to copy 0-8 bytes. */ + .align 4 +.LE8: + mr 12,4 + bne cr6,4f +/* Would have liked to use use ld/std here but the 630 processors are + slow for load/store doubles that are not at least word aligned. + Unaligned Load/Store word execute with only a 1 cycle penaltity. */ + lwz 6,0(4) + lwz 7,4(4) + stw 6,0(3) + stw 7,4(3) + /* Return original dst pointer. */ + ld 3,-16(1) + blr + .align 4 +4: bf 29,2b + lwz 6,0(4) + stw 6,0(3) +6: + bf 30,5f + lhz 7,4(4) + sth 7,4(3) + bf 31,0f + lbz 8,6(4) + stb 8,6(3) + ld 3,-16(1) + blr + .align 4 +5: + bf 31,0f + lbz 6,4(4) + stb 6,4(3) + .align 4 +0: + /* Return original dst pointer. */ + ld 3,-16(1) + blr + + .align 4 +.L6: + + /* Copy doublewords where the destination is aligned but the source is + not. Use aligned doubleword loads from the source, shifted to realign + the data, to allow aligned destination stores. */ + addi 11,9,-1 /* loop DW count is one less than total */ + subf 5,10,12 + sldi 10,10,3 + mr 4,3 + srdi 8,11,2 /* calculate the 32 byte loop count */ + ld 6,0(5) + mtcrf 0x01,11 + cmpldi cr6,9,4 + mtctr 8 + ld 7,8(5) + subfic 9,10,64 + bf 30,1f + + /* there are at least two DWs to copy */ + sld 0,6,10 + srd 8,7,9 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sld 0,7,10 + srd 8,6,9 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,8f /* if total DWs = 3, then bypass loop */ + bf 31,4f + /* there is a third DW to copy */ + sld 0,6,10 + srd 8,7,9 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,8f /* if total DWs = 4, then bypass loop */ + b 4f + .align 4 +1: + sld 0,6,10 + srd 8,7,9 + addi 5,5,16 + or 0,0,8 + bf 31,4f + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +4: sld 0,6,10 + srd 8,7,9 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sld 0,7,10 + srd 8,6,9 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sld 0,6,10 + srd 8,7,9 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sld 0,7,10 + srd 8,6,9 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ 4b + .align 4 +8: + /* calculate and store the final DW */ + sld 0,6,10 + srd 8,7,9 + or 0,0,8 + std 0,0(4) +3: + rldicr 0,31,0,60 + mtcrf 0x01,31 + bne cr1,.L9 /* If the tail is 0 bytes we are done! */ + /* Return original dst pointer. */ + ld 31,-8(1) + ld 3,-16(1) + blr +END_GEN_TB (BP_SYM (memcpy),TB_TOCLESS) +libc_hidden_builtin_def (memcpy) diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/strncmp.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/strncmp.S new file mode 100644 index 0000000000..7a1665d2bc --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/strncmp.S @@ -0,0 +1,180 @@ +/* Optimized strcmp implementation for PowerPC64. + Copyright (C) 2003, 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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., 51 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#include <sysdep.h> +#include <bp-sym.h> +#include <bp-asm.h> + +/* See strlen.s for comments on how the end-of-string testing works. */ + +/* int [r3] strncmp (const char *s1 [r3], const char *s2 [r4], size_t size [r5]) */ + +EALIGN (BP_SYM(strncmp), 4, 0) + CALL_MCOUNT 3 + +#define rTMP r0 +#define rRTN r3 +#define rSTR1 r3 /* first string arg */ +#define rSTR2 r4 /* second string arg */ +#define rN r5 /* max string length */ +/* Note: The Bounded pointer support in this code is broken. This code + was inherited from PPC32 and and that support was never completed. + Current PPC gcc does not support -fbounds-check or -fbounded-pointers. */ +#define rWORD1 r6 /* current word in s1 */ +#define rWORD2 r7 /* current word in s2 */ +#define rWORD3 r10 +#define rWORD4 r11 +#define rFEFE r8 /* constant 0xfefefefefefefeff (-0x0101010101010101) */ +#define r7F7F r9 /* constant 0x7f7f7f7f7f7f7f7f */ +#define rNEG r10 /* ~(word in s1 | 0x7f7f7f7f7f7f7f7f) */ +#define rBITDIF r11 /* bits that differ in s1 & s2 words */ + + dcbt 0,rSTR1 + or rTMP, rSTR2, rSTR1 + lis r7F7F, 0x7f7f + dcbt 0,rSTR2 + clrldi. rTMP, rTMP, 61 + cmpldi cr1, rN, 0 + lis rFEFE, -0x101 + bne L(unaligned) +/* We are doubleword alligned so set up for two loops. first a double word + loop, then fall into the byte loop if any residual. */ + srdi. rTMP, rN, 3 + clrldi rN, rN, 61 + addi rFEFE, rFEFE, -0x101 + addi r7F7F, r7F7F, 0x7f7f + cmpldi cr1, rN, 0 + beq L(unaligned) + + mtctr rTMP /* Power4 wants mtctr 1st in dispatch group. */ + ld rWORD1, 0(rSTR1) + ld rWORD2, 0(rSTR2) + sldi rTMP, rFEFE, 32 + insrdi r7F7F, r7F7F, 32, 0 + add rFEFE, rFEFE, rTMP + b L(g1) + +L(g0): + ldu rWORD1, 8(rSTR1) + bne- cr1, L(different) + ldu rWORD2, 8(rSTR2) +L(g1): add rTMP, rFEFE, rWORD1 + nor rNEG, r7F7F, rWORD1 + bdz L(tail) + and. rTMP, rTMP, rNEG + cmpd cr1, rWORD1, rWORD2 + beq+ L(g0) + +/* OK. We've hit the end of the string. We need to be careful that + we don't compare two strings as different because of gunk beyond + the end of the strings... */ + +L(endstring): + and rTMP, r7F7F, rWORD1 + beq cr1, L(equal) + add rTMP, rTMP, r7F7F + xor. rBITDIF, rWORD1, rWORD2 + + andc rNEG, rNEG, rTMP + blt- L(highbit) + cntlzd rBITDIF, rBITDIF + cntlzd rNEG, rNEG + addi rNEG, rNEG, 7 + cmpd cr1, rNEG, rBITDIF + sub rRTN, rWORD1, rWORD2 + blt- cr1, L(equal) + sradi rRTN, rRTN, 63 + ori rRTN, rRTN, 1 + blr +L(equal): + li rRTN, 0 + blr + +L(different): + ldu rWORD1, -8(rSTR1) + xor. rBITDIF, rWORD1, rWORD2 + sub rRTN, rWORD1, rWORD2 + blt- L(highbit) + sradi rRTN, rRTN, 63 + ori rRTN, rRTN, 1 + blr +L(highbit): + srdi rWORD2, rWORD2, 56 + srdi rWORD1, rWORD1, 56 + sub rRTN, rWORD1, rWORD2 + blr + + +/* Oh well. In this case, we just do a byte-by-byte comparison. */ + .align 4 +L(tail): + and. rTMP, rTMP, rNEG + cmpd cr1, rWORD1, rWORD2 + bne- L(endstring) + addi rSTR1, rSTR1, 8 + bne- cr1, L(different) + addi rSTR2, rSTR2, 8 + cmpldi cr1, rN, 0 +L(unaligned): + mtctr rN /* Power4 wants mtctr 1st in dispatch group */ + ble cr1, L(ux) +L(uz): + lbz rWORD1, 0(rSTR1) + lbz rWORD2, 0(rSTR2) + .align 4 +L(u1): + cmpdi cr1, rWORD1, 0 + bdz L(u4) + cmpd rWORD1, rWORD2 + beq- cr1, L(u4) + lbzu rWORD3, 1(rSTR1) + lbzu rWORD4, 1(rSTR2) + bne- L(u4) + cmpdi cr1, rWORD3, 0 + bdz L(u3) + cmpd rWORD3, rWORD4 + beq- cr1, L(u3) + lbzu rWORD1, 1(rSTR1) + lbzu rWORD2, 1(rSTR2) + bne- L(u3) + cmpdi cr1, rWORD1, 0 + bdz L(u4) + cmpd rWORD1, rWORD2 + beq- cr1, L(u4) + lbzu rWORD3, 1(rSTR1) + lbzu rWORD4, 1(rSTR2) + bne- L(u4) + cmpdi cr1, rWORD3, 0 + bdz L(u3) + cmpd rWORD3, rWORD4 + beq- cr1, L(u3) + lbzu rWORD1, 1(rSTR1) + lbzu rWORD2, 1(rSTR2) + beq+ L(u1) + +L(u3): sub rRTN, rWORD3, rWORD4 + blr +L(u4): sub rRTN, rWORD1, rWORD2 + blr +L(ux): + li rRTN, 0 + blr +END (BP_SYM (strncmp)) +libc_hidden_builtin_def (strncmp) + diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/wordcopy.c b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/wordcopy.c new file mode 100644 index 0000000000..f427b48e7a --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/wordcopy.c @@ -0,0 +1 @@ +#include "../../powerpc32/power4/wordcopy.c" diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/Implies new file mode 100644 index 0000000000..ac431fa96e --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4 diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/Implies new file mode 100644 index 0000000000..558a5fb174 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4/fpu diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceil.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceil.S new file mode 100644 index 0000000000..1e4851db9e --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceil.S @@ -0,0 +1,38 @@ +/* ceil function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + + .machine "power5" +EALIGN (__ceil, 4, 0) + CALL_MCOUNT 0 + frip fp1, fp1 + blr + END (__ceil) + +weak_alias (__ceil, ceil) + +#ifdef NO_LONG_DOUBLE +weak_alias (__ceil, ceill) +strong_alias (__ceil, __ceill) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0) +compat_symbol (libm, __ceil, ceill, GLIBC_2_0) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceilf.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceilf.S new file mode 100644 index 0000000000..38c51d5ff1 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_ceilf.S @@ -0,0 +1,31 @@ +/* ceilf function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> + + .machine "power5" +EALIGN (__ceilf, 4, 0) + CALL_MCOUNT 0 + frip fp1, fp1 /* The rounding instructions are double. */ + frsp fp1, fp1 /* But we need to set ooverflow for float. */ + blr + END (__ceilf) + +weak_alias (__ceilf, ceilf) + diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floor.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floor.S new file mode 100644 index 0000000000..86f07c8b59 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floor.S @@ -0,0 +1,38 @@ +/* floor function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + + .machine "power5" +EALIGN (__floor, 4, 0) + CALL_MCOUNT 0 + frim fp1, fp1 + blr + END (__floor) + +weak_alias (__floor, floor) + +#ifdef NO_LONG_DOUBLE +weak_alias (__floor, floorl) +strong_alias (__floor, __floorl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0) +compat_symbol (libm, __floor, floorl, GLIBC_2_0) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floorf.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floorf.S new file mode 100644 index 0000000000..e7bdccf513 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_floorf.S @@ -0,0 +1,31 @@ +/* floorf function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> + + .machine "power5" +EALIGN (__floorf, 4, 0) + CALL_MCOUNT 0 + frim fp1, fp1 /* The rounding instructions are double. */ + frsp fp1, fp1 /* But we need to set ooverflow for float. */ + blr + END (__floorf) + +weak_alias (__floorf, floorf) + diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_llround.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_llround.S new file mode 100644 index 0000000000..6a022769f4 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_llround.S @@ -0,0 +1,59 @@ +/* llround function. POWER5+, PowerPC64 version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + +/* long long [r3] llround (float x [fp1]) + IEEE 1003.1 llround function. IEEE specifies "round to the nearest + integer value, rounding halfway cases away from zero, regardless of + the current rounding mode." However PowerPC Architecture defines + "round to Nearest" as "Choose the best approximation. In case of a + tie, choose the one that is even (least significant bit o).". + So we pre-round using the V2.02 Floating Round to Integer Nearest + instruction before we use Floating Convert to Integer Word with + round to zero instruction. */ + + .machine "power5" +EALIGN (__llround, 4, 0) + CALL_MCOUNT 0 + frin fp2, fp1 /* Round to nearest +-0.5. */ + fctidz fp3, fp2 /* Convert To Integer DW round toward 0. */ + stfd fp3, -16(r1) + nop /* Insure the following load is in a different dispatch group */ + nop /* to avoid pipe stall on POWER4&5. */ + nop + ld r3, -16(r1) + blr + END (__llround) + +strong_alias (__llround, __lround) +weak_alias (__llround, llround) +weak_alias (__lround, lround) + +#ifdef NO_LONG_DOUBLE +weak_alias (__llround, llroundl) +strong_alias (__llround, __llroundl) +weak_alias (__lround, lroundl) +strong_alias (__lround, __lroundl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __llround, llroundl, GLIBC_2_1) +compat_symbol (libm, __lround, lroundl, GLIBC_2_1) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_round.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_round.S new file mode 100644 index 0000000000..7082734bb5 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_round.S @@ -0,0 +1,38 @@ +/* round function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + + .machine "power5" +EALIGN (__round, 4, 0) + CALL_MCOUNT 0 + frin fp1, fp1 + blr + END (__round) + +weak_alias (__round, round) + +#ifdef NO_LONG_DOUBLE +weak_alias (__round, roundl) +strong_alias (__round, __roundl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __round, roundl, GLIBC_2_1) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_roundf.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_roundf.S new file mode 100644 index 0000000000..30b83be9e7 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_roundf.S @@ -0,0 +1,31 @@ +/* roundf function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> + + .machine "power5" +EALIGN (__roundf, 4, 0) + CALL_MCOUNT 0 + frin fp1, fp1 /* The rounding instructions are double. */ + frsp fp1, fp1 /* But we need to set ooverflow for float. */ + blr + END (__roundf) + +weak_alias (__roundf, roundf) + diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_trunc.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_trunc.S new file mode 100644 index 0000000000..4e6e4d790e --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_trunc.S @@ -0,0 +1,38 @@ +/* trunc function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + + .machine "power5" +EALIGN (__trunc, 4, 0) + CALL_MCOUNT 0 + friz fp1, fp1 + blr + END (__trunc) + +weak_alias (__trunc, trunc) + +#ifdef NO_LONG_DOUBLE +weak_alias (__trunc, truncl) +strong_alias (__trunc, __truncl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __trunc, truncl, GLIBC_2_1) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_truncf.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_truncf.S new file mode 100644 index 0000000000..61d9cb372d --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5+/fpu/s_truncf.S @@ -0,0 +1,31 @@ +/* truncf function. PowerPC64/power5+ version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> + + .machine "power5" +EALIGN (__truncf, 4, 0) + CALL_MCOUNT 0 + friz fp1, fp1 /* The rounding instructions are double. */ + frsp fp1, fp1 /* But we need to set ooverflow for float. */ + blr + END (__truncf) + +weak_alias (__truncf, truncf) + diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/Implies new file mode 100644 index 0000000000..ac431fa96e --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4 diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/fpu/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/fpu/Implies new file mode 100644 index 0000000000..558a5fb174 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power5/fpu/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power4/fpu diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/Implies new file mode 100644 index 0000000000..0f35f37c77 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/Implies @@ -0,0 +1 @@ +powerpc/powerpc64/power5+ diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/fpu/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/fpu/Implies new file mode 100644 index 0000000000..885e39c003 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/fpu/Implies @@ -0,0 +1,2 @@ +powerpc/powerpc64/power5+/fpu +powerpc/powerpc64/power4/fpu diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/memcpy.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/memcpy.S new file mode 100644 index 0000000000..98798e5d6b --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/memcpy.S @@ -0,0 +1,1002 @@ +/* Optimized memcpy implementation for PowerPC64. + Copyright (C) 2003, 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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., 51 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#include <sysdep.h> +#include <bp-sym.h> +#include <bp-asm.h> + +/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]); + Returns 'dst'. + + Memcpy handles short copies (< 32-bytes) using a binary move blocks + (no loops) of lwz/stw. The tail (remaining 1-3) bytes is handled + with the appropriate combination of byte and halfword load/stores. + There is minimal effort to optimize the alignment of short moves. + The 64-bit implementations of POWER3 and POWER4 do a reasonable job + of handling unligned load/stores that do not cross 32-byte boundries. + + Longer moves (>= 32-bytes) justify the effort to get at least the + destination doubleword (8-byte) aligned. Further optimization is + posible when both source and destination are doubleword aligned. + Each case has a optimized unrolled loop. */ + +EALIGN (BP_SYM (memcpy), 5, 0) + CALL_MCOUNT 3 + + cmpldi cr1,5,31 + neg 0,3 + std 3,-16(1) + std 31,-8(1) + cfi_offset(31,-8) + andi. 11,3,7 /* check alignement of dst. */ + clrldi 0,0,61 /* Number of bytes until the 1st doubleword of dst. */ + clrldi 10,4,61 /* check alignement of src. */ + cmpldi cr6,5,8 + ble- cr1,.L2 /* If move < 32 bytes use short move code. */ + cmpld cr6,10,11 + mr 12,4 + srdi 9,5,3 /* Number of full double words remaining. */ + mtcrf 0x01,0 + mr 31,5 + beq .L0 + + subf 31,0,5 + /* Move 0-7 bytes as needed to get the destination doubleword alligned. */ +1: bf 31,2f + lbz 6,0(12) + addi 12,12,1 + stb 6,0(3) + addi 3,3,1 +2: bf 30,4f + lhz 6,0(12) + addi 12,12,2 + sth 6,0(3) + addi 3,3,2 +4: bf 29,0f + lwz 6,0(12) + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +0: + clrldi 10,12,61 /* check alignement of src again. */ + srdi 9,31,3 /* Number of full double words remaining. */ + + /* Copy doublewords from source to destination, assumpting the + destination is aligned on a doubleword boundary. + + At this point we know there are at least 25 bytes left (32-7) to copy. + The next step is to determine if the source is also doubleword aligned. + If not branch to the unaligned move code at .L6. which uses + a load, shift, store strategy. + + Otherwise source and destination are doubleword aligned, and we can + the optimized doubleword copy loop. */ +.L0: + clrldi 11,31,61 + mtcrf 0x01,9 + cmpldi cr1,11,0 + bne- cr6,.L6 /* If source is not DW aligned. */ + + /* Move doublewords where destination and source are DW aligned. + Use a unrolled loop to copy 4 doubleword (32-bytes) per iteration. + If the the copy is not an exact multiple of 32 bytes, 1-3 + doublewords are copied as needed to set up the main loop. After + the main loop exits there may be a tail of 1-7 bytes. These byte + are copied a word/halfword/byte at a time as needed to preserve + alignment. + + For POWER6 the L1 is store-through and the L2 is store-in. The + L2 is clocked at half CPU clock so we can store 16 bytes every + other cycle. POWER6 also has a load/store bypass so we can do + load, load, store, store every 2 cycles. + + For POWER6 unaligned loads will take a 20+ cycle hicup for any + L1 cache miss that crosses a 32- or 128-byte boundary. Store + is more forgiving and does not take a hicup until page or + segment boundaries. So we require doubleword alignment for + the source but may take a risk and only require word alignment + for the destination. */ + + srdi 8,31,5 + cmpldi cr1,9,4 + cmpldi cr6,11,0 + mr 11,12 + + bf 30,1f + ld 6,0(12) + ld 7,8(12) + addi 11,12,16 + mtctr 8 + std 6,0(3) + std 7,8(3) + addi 10,3,16 + bf 31,4f + ld 0,16(12) + std 0,16(3) + blt cr1,3f + addi 11,12,24 + addi 10,3,24 + b 4f + .align 4 +1: + mr 10,3 + mtctr 8 + bf 31,4f + ld 6,0(12) + addi 11,12,8 + std 6,0(3) + addi 10,3,8 + + .align 4 +4: + ld 6,0(11) + ld 7,8(11) + std 6,0(10) + std 7,8(10) + ld 8,16(11) + ld 0,24(11) + std 8,16(10) + std 0,24(10) + bdz 3f + + ld 6,0+32(11) + ld 7,8+32(11) + std 6,0+32(10) + std 7,8+32(10) + ld 8,16+32(11) + ld 0,24+32(11) + std 8,16+32(10) + std 0,24+32(10) + bdz 3f + + ld 6,0+64(11) + ld 7,8+64(11) + std 6,0+64(10) + std 7,8+64(10) + ld 8,16+64(11) + ld 0,24+64(11) + std 8,16+64(10) + std 0,24+64(10) + bdz 3f + + ld 6,0+96(11) + ld 7,8+96(11) + std 6,0+96(10) + std 7,8+96(10) + ld 8,16+96(11) + ld 0,24+96(11) + addi 11,11,128 + std 8,16+96(10) + std 0,24+96(10) + addi 10,10,128 + bdnz 4b +3: + + rldicr 0,31,0,60 + mtcrf 0x01,31 + beq cr6,0f +.L9: + add 3,3,0 + add 12,12,0 + +/* At this point we have a tail of 0-7 bytes and we know that the + destiniation is double word aligned. */ +4: bf 29,2f + lwz 6,0(12) + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +2: bf 30,1f + lhz 6,0(12) + addi 12,12,2 + sth 6,0(3) + addi 3,3,2 +1: bf 31,0f + lbz 6,0(12) + stb 6,0(3) +0: + /* Return original dst pointer. */ + ld 31,-8(1) + ld 3,-16(1) + blr + +/* Copy up to 31 bytes. This divided into two cases 0-8 bytes and 9-31 + bytes. Each case is handled without loops, using binary (1,2,4,8) + tests. + + In the short (0-8 byte) case no attempt is made to force alignment + of either source or destination. The hardware will handle the + unaligned load/stores with small delays for crossing 32- 128-byte, + and 4096-byte boundaries. Since these short moves are unlikely to be + unaligned or cross these boundaries, the overhead to force + alignment is not justified. + + The longer (9-31 byte) move is more likely to cross 32- or 128-byte + boundaries. Since only loads are sensitive to the 32-/128-byte + boundaries it is more important to align the source then the + destination. If the source is not already word aligned, we first + move 1-3 bytes as needed. Since we are only word aligned we don't + use double word load/stores to insure that all loads are aligned. + While the destination and stores may still be unaligned, this + is only an issue for page (4096 byte boundary) crossing, which + should be rare for these short moves. The hardware handles this + case automatically with a small (~20 cycle) delay. */ + .align 4 +.L2: + mtcrf 0x01,5 + neg 8,4 + clrrdi 11,4,2 + andi. 0,8,3 + ble cr6,.LE8 /* Handle moves of 0-8 bytes. */ +/* At least 9 bytes left. Get the source word aligned. */ + cmpldi cr1,5,16 + mr 10,5 + mr 12,4 + cmpldi cr6,0,2 + beq L(dus_tail) /* If the source is already word aligned skip this. */ +/* Copy 1-3 bytes to get source address word aligned. */ + lwz 6,0(11) + subf 10,0,5 + add 12,4,0 + blt cr6,5f + srdi 7,6,16 + bgt cr6,3f + sth 6,0(3) + b 7f + .align 4 +3: + stb 7,0(3) + sth 6,1(3) + b 7f + .align 4 +5: + stb 6,0(3) +7: + cmpldi cr1,10,16 + add 3,3,0 + mtcrf 0x01,10 + .align 4 +L(dus_tail): +/* At least 6 bytes left and the source is word aligned. This allows + some speculative loads up front. */ +/* We need to special case the fall-through because the biggest delays + are due to address computation not being ready in time for the + AGEN. */ + lwz 6,0(12) + lwz 7,4(12) + blt cr1,L(dus_tail8) + cmpldi cr0,10,24 +L(dus_tail16): /* Move 16 bytes. */ + stw 6,0(3) + stw 7,4(3) + lwz 6,8(12) + lwz 7,12(12) + stw 6,8(3) + stw 7,12(3) +/* Move 8 bytes more. */ + bf 28,L(dus_tail16p8) + cmpldi cr1,10,28 + lwz 6,16(12) + lwz 7,20(12) + stw 6,16(3) + stw 7,20(3) +/* Move 4 bytes more. */ + bf 29,L(dus_tail16p4) + lwz 6,24(12) + stw 6,24(3) + addi 12,12,28 + addi 3,3,28 + bgt cr1,L(dus_tail2) + /* exactly 28 bytes. Return original dst pointer and exit. */ + ld 3,-16(1) + blr + .align 4 +L(dus_tail16p8): /* less then 8 bytes left. */ + beq cr1,L(dus_tailX) /* exactly 16 bytes, early exit. */ + cmpldi cr1,10,20 + bf 29,L(dus_tail16p2) +/* Move 4 bytes more. */ + lwz 6,16(12) + stw 6,16(3) + addi 12,12,20 + addi 3,3,20 + bgt cr1,L(dus_tail2) + /* exactly 20 bytes. Return original dst pointer and exit. */ + ld 3,-16(1) + blr + .align 4 +L(dus_tail16p4): /* less then 4 bytes left. */ + addi 12,12,24 + addi 3,3,24 + bgt cr0,L(dus_tail2) + /* exactly 24 bytes. Return original dst pointer and exit. */ + ld 3,-16(1) + blr + .align 4 +L(dus_tail16p2): /* 16 bytes moved, less then 4 bytes left. */ + addi 12,12,16 + addi 3,3,16 + b L(dus_tail2) + + .align 4 +L(dus_tail8): /* Move 8 bytes. */ +/* r6, r7 already loaded speculatively. */ + cmpldi cr1,10,8 + cmpldi cr0,10,12 + bf 28,L(dus_tail4) + stw 6,0(3) + stw 7,4(3) +/* Move 4 bytes more. */ + bf 29,L(dus_tail8p4) + lwz 6,8(12) + stw 6,8(3) + addi 12,12,12 + addi 3,3,12 + bgt cr0,L(dus_tail2) + /* exactly 12 bytes. Return original dst pointer and exit. */ + ld 3,-16(1) + blr + .align 4 +L(dus_tail8p4): /* less then 4 bytes left. */ + addi 12,12,8 + addi 3,3,8 + bgt cr1,L(dus_tail2) + /* exactly 8 bytes. Return original dst pointer and exit. */ + ld 3,-16(1) + blr + + .align 4 +L(dus_tail4): /* Move 4 bytes. */ +/* r6 already loaded speculatively. If we are here we know there is + more then 4 bytes left. So there is no need to test. */ + addi 12,12,4 + stw 6,0(3) + addi 3,3,4 +L(dus_tail2): /* Move 2-3 bytes. */ + bf 30,L(dus_tail1) + lhz 6,0(12) + sth 6,0(3) + bf 31,L(dus_tailX) + lbz 7,2(12) + stb 7,2(3) + ld 3,-16(1) + blr +L(dus_tail1): /* Move 1 byte. */ + bf 31,L(dus_tailX) + lbz 6,0(12) + stb 6,0(3) +L(dus_tailX): + /* Return original dst pointer. */ + ld 3,-16(1) + blr + +/* Special case to copy 0-8 bytes. */ + .align 4 +.LE8: + mr 12,4 + bne cr6,L(dus_4) +/* Exactly 8 bytes. We may cross a 32-/128-byte boundry and take a ~20 + cycle delay. This case should be rare and any attempt to avoid this + would take most of 20 cycles any way. */ + ld 6,0(4) + std 6,0(3) + /* Return original dst pointer. */ + ld 3,-16(1) + blr + .align 4 +L(dus_4): + bf 29,L(dus_tail2) + lwz 6,0(4) + stw 6,0(3) + bf 30,L(dus_5) + lhz 7,4(4) + sth 7,4(3) + bf 31,L(dus_0) + lbz 8,6(4) + stb 8,6(3) + ld 3,-16(1) + blr + .align 4 +L(dus_5): + bf 31,L(dus_0) + lbz 6,4(4) + stb 6,4(3) +L(dus_0): + /* Return original dst pointer. */ + ld 3,-16(1) + blr + + .align 4 +.L6: + + /* Copy doublewords where the destination is aligned but the source is + not. Use aligned doubleword loads from the source, shifted to realign + the data, to allow aligned destination stores. */ + addi 11,9,-1 /* loop DW count is one less than total */ + subf 5,10,12 /* Move source addr to previous full double word. */ + cmpldi cr5, 10, 2 + cmpldi cr0, 10, 4 + mr 4,3 + srdi 8,11,2 /* calculate the 32 byte loop count */ + ld 6,0(5) /* pre load 1st full doubleword. */ + mtcrf 0x01,11 + cmpldi cr6,9,4 + mtctr 8 + ld 7,8(5) /* pre load 2nd full doubleword. */ + bge cr0, L(du4_do) + blt cr5, L(du1_do) + beq cr5, L(du2_do) + b L(du3_do) + + .align 4 +L(du1_do): + bf 30,L(du1_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 8 + srdi 8,7, 64-8 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 8 + srdi 8,6, 64-8 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du1_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du1_loop) + /* there is a third DW to copy */ + sldi 0,6, 8 + srdi 8,7, 64-8 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du1_fini) /* if total DWs = 4, then bypass loop */ + b L(du1_loop) + .align 4 +L(du1_1dw): + sldi 0,6, 8 + srdi 8,7, 64-8 + addi 5,5,16 + or 0,0,8 + bf 31,L(du1_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du1_loop): + sldi 0,6, 8 + srdi 8,7, 64-8 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 8 + srdi 8,6, 64-8 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 8 + srdi 8,7, 64-8 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 8 + srdi 8,6, 64-8 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du1_loop) + .align 4 +L(du1_fini): + /* calculate and store the final DW */ + sldi 0,6, 8 + srdi 8,7, 64-8 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du2_do): + bf 30,L(du2_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 16 + srdi 8,7, 64-16 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 16 + srdi 8,6, 64-16 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du2_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du2_loop) + /* there is a third DW to copy */ + sldi 0,6, 16 + srdi 8,7, 64-16 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du2_fini) /* if total DWs = 4, then bypass loop */ + b L(du2_loop) + .align 4 +L(du2_1dw): + sldi 0,6, 16 + srdi 8,7, 64-16 + addi 5,5,16 + or 0,0,8 + bf 31,L(du2_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du2_loop): + sldi 0,6, 16 + srdi 8,7, 64-16 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 16 + srdi 8,6, 64-16 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 16 + srdi 8,7, 64-16 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 16 + srdi 8,6, 64-16 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du2_loop) + .align 4 +L(du2_fini): + /* calculate and store the final DW */ + sldi 0,6, 16 + srdi 8,7, 64-16 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du3_do): + bf 30,L(du3_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 24 + srdi 8,7, 64-24 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 24 + srdi 8,6, 64-24 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du3_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du3_loop) + /* there is a third DW to copy */ + sldi 0,6, 24 + srdi 8,7, 64-24 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du3_fini) /* if total DWs = 4, then bypass loop */ + b L(du3_loop) + .align 4 +L(du3_1dw): + sldi 0,6, 24 + srdi 8,7, 64-24 + addi 5,5,16 + or 0,0,8 + bf 31,L(du3_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du3_loop): + sldi 0,6, 24 + srdi 8,7, 64-24 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 24 + srdi 8,6, 64-24 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 24 + srdi 8,7, 64-24 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 24 + srdi 8,6, 64-24 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du3_loop) + .align 4 +L(du3_fini): + /* calculate and store the final DW */ + sldi 0,6, 24 + srdi 8,7, 64-24 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du4_do): + cmpldi cr5, 10, 6 + beq cr0, L(du4_dox) + blt cr5, L(du5_do) + beq cr5, L(du6_do) + b L(du7_do) +L(du4_dox): + bf 30,L(du4_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 32 + srdi 8,7, 64-32 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 32 + srdi 8,6, 64-32 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du4_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du4_loop) + /* there is a third DW to copy */ + sldi 0,6, 32 + srdi 8,7, 64-32 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du4_fini) /* if total DWs = 4, then bypass loop */ + b L(du4_loop) + .align 4 +L(du4_1dw): + sldi 0,6, 32 + srdi 8,7, 64-32 + addi 5,5,16 + or 0,0,8 + bf 31,L(du4_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du4_loop): + sldi 0,6, 32 + srdi 8,7, 64-32 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 32 + srdi 8,6, 64-32 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 32 + srdi 8,7, 64-32 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 32 + srdi 8,6, 64-32 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du4_loop) + .align 4 +L(du4_fini): + /* calculate and store the final DW */ + sldi 0,6, 32 + srdi 8,7, 64-32 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du5_do): + bf 30,L(du5_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 40 + srdi 8,7, 64-40 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 40 + srdi 8,6, 64-40 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du5_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du5_loop) + /* there is a third DW to copy */ + sldi 0,6, 40 + srdi 8,7, 64-40 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du5_fini) /* if total DWs = 4, then bypass loop */ + b L(du5_loop) + .align 4 +L(du5_1dw): + sldi 0,6, 40 + srdi 8,7, 64-40 + addi 5,5,16 + or 0,0,8 + bf 31,L(du5_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du5_loop): + sldi 0,6, 40 + srdi 8,7, 64-40 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 40 + srdi 8,6, 64-40 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 40 + srdi 8,7, 64-40 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 40 + srdi 8,6, 64-40 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du5_loop) + .align 4 +L(du5_fini): + /* calculate and store the final DW */ + sldi 0,6, 40 + srdi 8,7, 64-40 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du6_do): + bf 30,L(du6_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 48 + srdi 8,7, 64-48 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 48 + srdi 8,6, 64-48 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du6_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du6_loop) + /* there is a third DW to copy */ + sldi 0,6, 48 + srdi 8,7, 64-48 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du6_fini) /* if total DWs = 4, then bypass loop */ + b L(du6_loop) + .align 4 +L(du6_1dw): + sldi 0,6, 48 + srdi 8,7, 64-48 + addi 5,5,16 + or 0,0,8 + bf 31,L(du6_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du6_loop): + sldi 0,6, 48 + srdi 8,7, 64-48 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 48 + srdi 8,6, 64-48 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 48 + srdi 8,7, 64-48 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 48 + srdi 8,6, 64-48 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du6_loop) + .align 4 +L(du6_fini): + /* calculate and store the final DW */ + sldi 0,6, 48 + srdi 8,7, 64-48 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du7_do): + bf 30,L(du7_1dw) + + /* there are at least two DWs to copy */ + sldi 0,6, 56 + srdi 8,7, 64-56 + or 0,0,8 + ld 6,16(5) + std 0,0(4) + sldi 0,7, 56 + srdi 8,6, 64-56 + or 0,0,8 + ld 7,24(5) + std 0,8(4) + addi 4,4,16 + addi 5,5,32 + blt cr6,L(du7_fini) /* if total DWs = 3, then bypass loop */ + bf 31,L(du7_loop) + /* there is a third DW to copy */ + sldi 0,6, 56 + srdi 8,7, 64-56 + or 0,0,8 + std 0,0(4) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + addi 4,4,8 + beq cr6,L(du7_fini) /* if total DWs = 4, then bypass loop */ + b L(du7_loop) + .align 4 +L(du7_1dw): + sldi 0,6, 56 + srdi 8,7, 64-56 + addi 5,5,16 + or 0,0,8 + bf 31,L(du7_loop) + mr 6,7 + ld 7,0(5) + addi 5,5,8 + std 0,0(4) + addi 4,4,8 + .align 4 +/* copy 32 bytes at a time */ +L(du7_loop): + sldi 0,6, 56 + srdi 8,7, 64-56 + or 0,0,8 + ld 6,0(5) + std 0,0(4) + sldi 0,7, 56 + srdi 8,6, 64-56 + or 0,0,8 + ld 7,8(5) + std 0,8(4) + sldi 0,6, 56 + srdi 8,7, 64-56 + or 0,0,8 + ld 6,16(5) + std 0,16(4) + sldi 0,7, 56 + srdi 8,6, 64-56 + or 0,0,8 + ld 7,24(5) + std 0,24(4) + addi 5,5,32 + addi 4,4,32 + bdnz+ L(du7_loop) + .align 4 +L(du7_fini): + /* calculate and store the final DW */ + sldi 0,6, 56 + srdi 8,7, 64-56 + or 0,0,8 + std 0,0(4) + b L(du_done) + + .align 4 +L(du_done): + rldicr 0,31,0,60 + mtcrf 0x01,31 + bne cr1,.L9 /* If the tail is 0 bytes we are done! */ + /* Return original dst pointer. */ + ld 31,-8(1) + ld 3,-16(1) + blr +END_GEN_TB (BP_SYM (memcpy),TB_TOCLESS) +libc_hidden_builtin_def (memcpy) diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/wordcopy.c b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/wordcopy.c new file mode 100644 index 0000000000..faddd945b3 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/wordcopy.c @@ -0,0 +1,410 @@ +/* _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; + int align; + + /* Calculate how to shift a word read at the memory operation + aligned srcp to make it aligned for copy. */ + + align = srcp % OPSIZ; + 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; + } + + switch (align) + { + case 1: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 8, a1, (64-8)); + ((op_t *) dstp)[1] = MERGE (a1, 8, a2, (64-8)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 2: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 16, a1, (64-16)); + ((op_t *) dstp)[1] = MERGE (a1, 16, a2, (64-16)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 3: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 24, a1, (64-24)); + ((op_t *) dstp)[1] = MERGE (a1, 24, a2, (64-24)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 4: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 32, a1, (64-32)); + ((op_t *) dstp)[1] = MERGE (a1, 32, a2, (64-32)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 5: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 40, a1, (64-40)); + ((op_t *) dstp)[1] = MERGE (a1, 40, a2, (64-40)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 6: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 48, a1, (64-48)); + ((op_t *) dstp)[1] = MERGE (a1, 48, a2, (64-48)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + case 7: + do + { + a1 = ((op_t *) srcp)[1]; + a2 = ((op_t *) srcp)[2]; + ((op_t *) dstp)[0] = MERGE (a0, 56, a1, (64-56)); + ((op_t *) dstp)[1] = MERGE (a1, 56, a2, (64-56)); + a0 = a2; + + srcp += 2 * OPSIZ; + dstp += 2 * OPSIZ; + len -= 2; + } + while (len != 0); + break; + } + +} + +/* _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; + int align; + + /* Calculate how to shift a word read at the memory operation + aligned srcp to make it aligned for copy. */ + + align = srcp % OPSIZ; + 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; + } + + switch (align) + { + case 1: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 8, a2, (64-8)); + ((op_t *) dstp)[0] = MERGE (a0, 8, a1, (64-8)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 2: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 16, a2, (64-16)); + ((op_t *) dstp)[0] = MERGE (a0, 16, a1, (64-16)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 3: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 24, a2, (64-24)); + ((op_t *) dstp)[0] = MERGE (a0, 24, a1, (64-24)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 4: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 32, a2, (64-32)); + ((op_t *) dstp)[0] = MERGE (a0, 32, a1, (64-32)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 5: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 40, a2, (64-40)); + ((op_t *) dstp)[0] = MERGE (a0, 40, a1, (64-40)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 6: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 48, a2, (64-48)); + ((op_t *) dstp)[0] = MERGE (a0, 48, a1, (64-48)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + case 7: + do + { + srcp -= 2 * OPSIZ; + dstp -= 2 * OPSIZ; + + a1 = ((op_t *) srcp)[1]; + a0 = ((op_t *) srcp)[0]; + ((op_t *) dstp)[1] = MERGE (a1, 56, a2, (64-56)); + ((op_t *) dstp)[0] = MERGE (a0, 56, a1, (64-56)); + a2 = a0; + + len -= 2; + } + while (len != 0); + break; + } +} diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/Implies new file mode 100644 index 0000000000..d993b0cd73 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/Implies @@ -0,0 +1,2 @@ +powerpc/powerpc64/power6 +powerpc/powerpc64/power5+ diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/Implies b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/Implies new file mode 100644 index 0000000000..84510991b2 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/Implies @@ -0,0 +1,3 @@ +powerpc/powerpc64/power6/fpu +powerpc/powerpc64/power5+/fpu +powerpc/powerpc64/power4/fpu diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llrint.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llrint.S new file mode 100644 index 0000000000..4a4a5f8855 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llrint.S @@ -0,0 +1,45 @@ +/* Round double to long int. POWER6x PowerPC64 version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + + .machine "power6" +/* long long int[r3] __llrint (double x[fp1]) */ +ENTRY (__llrint) + CALL_MCOUNT 0 + fctid fp13,fp1 + mftgpr r3,fp13 + blr + END (__llrint) + +strong_alias (__llrint, __lrint) +weak_alias (__llrint, llrint) +weak_alias (__lrint, lrint) + +#ifdef NO_LONG_DOUBLE +strong_alias (__llrint, __llrintl) +weak_alias (__llrint, llrintl) +strong_alias (__lrint, __lrintl) +weak_alias (__lrint, lrintl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __llrint, llrintl, GLIBC_2_1) +compat_symbol (libm, __lrint, lrintl, GLIBC_2_1) +#endif diff --git a/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llround.S b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llround.S new file mode 100644 index 0000000000..149feaf965 --- /dev/null +++ b/powerpc-cpu/sysdeps/powerpc/powerpc64/power6x/fpu/s_llround.S @@ -0,0 +1,55 @@ +/* llround function. POWER6x PowerPC64 version. + Copyright (C) 2006 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + 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. */ + +#include <sysdep.h> +#include <math_ldbl_opt.h> + +/* long long [r3] llround (float x [fp1]) + IEEE 1003.1 llround function. IEEE specifies "round to the nearest + integer value, rounding halfway cases away from zero, regardless of + the current rounding mode." However PowerPC Architecture defines + "round to Nearest" as "Choose the best approximation. In case of a + tie, choose the one that is even (least significant bit o).". + So we pre-round using the V2.02 Floating Round to Integer Nearest + instruction before we use Floating Convert to Integer Word with + round to zero instruction. */ + + .machine "power6" +ENTRY (__llround) + CALL_MCOUNT 0 + frin fp2,fp1 /* Round to nearest +-0.5. */ + fctidz fp3,fp2 /* Convert To Integer DW round toward 0. */ + mftgpr r3,fp3 /* Transfer integer to R3. */ + blr + END (__llround) + +strong_alias (__llround, __lround) +weak_alias (__llround, llround) +weak_alias (__lround, lround) + +#ifdef NO_LONG_DOUBLE +weak_alias (__llround, llroundl) +strong_alias (__llround, __llroundl) +weak_alias (__lround, lroundl) +strong_alias (__lround, __lroundl) +#endif +#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_1) +compat_symbol (libm, __llround, llroundl, GLIBC_2_1) +compat_symbol (libm, __lround, lroundl, GLIBC_2_1) +#endif |