From a2e0c2fc3d074c914d06700fec0ac49676443324 Mon Sep 17 00:00:00 2001 From: Dibyendu Majumdar Date: Sat, 20 Jun 2020 13:51:59 +0100 Subject: [PATCH] issue #191 #193 backport random number generator from Lua 5.4 --- build-utils/cmake-64bit-nojit-assertions.bat | 2 +- include/lauxlib.h | 4 + include/luaconf.h | 46 +- src/lmathlib.c | 433 +++++++++++++++++-- src/lstrlib.c | 59 ++- src/lvm.c | 26 +- tests/lua53/math.lua | 220 +++++++--- tests/lua53/strings.lua | 6 +- 8 files changed, 659 insertions(+), 137 deletions(-) diff --git a/build-utils/cmake-64bit-nojit-assertions.bat b/build-utils/cmake-64bit-nojit-assertions.bat index a9f2bb3..a2f6af6 100644 --- a/build-utils/cmake-64bit-nojit-assertions.bat +++ b/build-utils/cmake-64bit-nojit-assertions.bat @@ -1,4 +1,4 @@ mkdir nojit64a cd nojit64a -cmake -DCMAKE_INSTALL_PREFIX=c:\Software\ravi -DCMAKE_BUILD_TYPE=Debug -DLTESTS=ON -G "Visual Studio 15 2017 Win64" .. +cmake -DCMAKE_INSTALL_PREFIX=c:\Software\ravi -DCMAKE_BUILD_TYPE=Debug -DLTESTS=ON .. cd .. \ No newline at end of file diff --git a/include/lauxlib.h b/include/lauxlib.h index 02a3035..73f7d2a 100644 --- a/include/lauxlib.h +++ b/include/lauxlib.h @@ -143,6 +143,10 @@ LUALIB_API void (luaL_requiref) (lua_State *L, const char *modname, #define luaL_loadbuffer(L,s,sz,n) luaL_loadbufferx(L,s,sz,n,NULL) +/* push the value used to represent failure/error */ +#define luaL_pushfail(L) lua_pushnil(L) + + /* ** {====================================================== ** Generic Buffer manipulation diff --git a/include/luaconf.h b/include/luaconf.h index a448e5c..ee75124 100644 --- a/include/luaconf.h +++ b/include/luaconf.h @@ -421,14 +421,14 @@ @@ LUA_NUMBER is the floating-point type used by Lua. @@ LUAI_UACNUMBER is the result of a 'default argument promotion' @@ over a floating number. -@@ l_mathlim(x) corrects limit name 'x' to the proper float type +@@ l_floatatt(x) corrects float attribute 'x' to the proper float type ** by prefixing it with one of FLT/DBL/LDBL. @@ LUA_NUMBER_FRMLEN is the length modifier for writing floats. @@ LUA_NUMBER_FMT is the format for writing floats. @@ lua_number2str converts a float to a string. @@ l_mathop allows the addition of an 'l' or 'f' to all math operations. @@ l_floor takes the floor of a float. -@@ lua_str2number converts a decimal numeric string to a number. +@@ lua_str2number converts a decimal numeral to a number. */ @@ -440,12 +440,13 @@ l_sprintf((s), sz, LUA_NUMBER_FMT, (LUAI_UACNUMBER)(n)) /* -@@ lua_numbertointeger converts a float number to an integer, or -** returns 0 if float is not within the range of a lua_Integer. -** (The range comparisons are tricky because of rounding. The tests -** here assume a two-complement representation, where MININTEGER always -** has an exact representation as a float; MAXINTEGER may not have one, -** and therefore its conversion to float may have an ill-defined value.) +@@ lua_numbertointeger converts a float number with an integral value +** to an integer, or returns 0 if float is not within the range of +** a lua_Integer. (The range comparisons are tricky because of +** rounding. The tests here assume a two-complement representation, +** where MININTEGER always has an exact representation as a float; +** MAXINTEGER may not have one, and therefore its conversion to float +** may have an ill-defined value.) */ #define lua_numbertointeger(n,p) \ ((n) >= (LUA_NUMBER)(LUA_MININTEGER) && \ @@ -459,7 +460,7 @@ #define LUA_NUMBER float -#define l_mathlim(n) (FLT_##n) +#define l_floatatt(n) (FLT_##n) #define LUAI_UACNUMBER double @@ -475,7 +476,7 @@ #define LUA_NUMBER long double -#define l_mathlim(n) (LDBL_##n) +#define l_floatatt(n) (LDBL_##n) #define LUAI_UACNUMBER long double @@ -490,7 +491,7 @@ #define LUA_NUMBER double -#define l_mathlim(n) (DBL_##n) +#define l_floatatt(n) (DBL_##n) #define LUAI_UACNUMBER double @@ -515,11 +516,13 @@ @@ LUA_UNSIGNED is the unsigned version of LUA_INTEGER. ** @@ LUAI_UACINT is the result of a 'default argument promotion' -@@ over a lUA_INTEGER. +@@ over a LUA_INTEGER. @@ LUA_INTEGER_FRMLEN is the length modifier for reading/writing integers. @@ LUA_INTEGER_FMT is the format for writing integers. @@ LUA_MAXINTEGER is the maximum value for a LUA_INTEGER. @@ LUA_MININTEGER is the minimum value for a LUA_INTEGER. +@@ LUA_MAXUNSIGNED is the maximum value for a LUA_UNSIGNED. +@@ LUA_UNSIGNEDBITS is the number of bits in a LUA_UNSIGNED. @@ lua_integer2str converts an integer to a string. */ @@ -540,6 +543,9 @@ #define LUA_UNSIGNED unsigned LUAI_UACINT +#define LUA_UNSIGNEDBITS (sizeof(LUA_UNSIGNED) * CHAR_BIT) + + /* now the variable definitions */ #if LUA_INT_TYPE == LUA_INT_INT /* { int */ @@ -550,6 +556,8 @@ #define LUA_MAXINTEGER INT_MAX #define LUA_MININTEGER INT_MIN +#define LUA_MAXUNSIGNED UINT_MAX + #elif LUA_INT_TYPE == LUA_INT_LONG /* }{ long */ #define LUA_INTEGER long @@ -558,6 +566,8 @@ #define LUA_MAXINTEGER LONG_MAX #define LUA_MININTEGER LONG_MIN +#define LUA_MAXUNSIGNED ULONG_MAX + #elif LUA_INT_TYPE == LUA_INT_LONGLONG /* }{ long long */ /* use presence of macro LLONG_MAX as proxy for C99 compliance */ @@ -570,6 +580,8 @@ #define LUA_MAXINTEGER LLONG_MAX #define LUA_MININTEGER LLONG_MIN +#define LUA_MAXUNSIGNED ULLONG_MAX + #elif defined(LUA_USE_WINDOWS) /* }{ */ /* in Windows, can use specific Windows types */ @@ -579,6 +591,8 @@ #define LUA_MAXINTEGER _I64_MAX #define LUA_MININTEGER _I64_MIN +#define LUA_MAXUNSIGNED _UI64_MAX + #else /* }{ */ #error "Compiler does not support 'long long'. Use option '-DLUA_32BITS' \ @@ -613,7 +627,7 @@ /* -@@ lua_strx2number converts an hexadecimal numeric string to a number. +@@ lua_strx2number converts a hexadecimal numeral to a number. ** In C99, 'strtod' does that conversion. Otherwise, you can ** leave 'lua_strx2number' undefined and Lua will provide its own ** implementation. @@ -631,7 +645,7 @@ /* -@@ lua_number2strx converts a float to an hexadecimal numeric string. +@@ lua_number2strx converts a float to a hexadecimal numeral. ** In C99, 'sprintf' (with format specifiers '%a'/'%A') does that. ** Otherwise, you can leave 'lua_number2strx' undefined and Lua will ** provide its own implementation. @@ -677,7 +691,7 @@ /* @@ lua_getlocaledecpoint gets the locale "radix character" (decimal point). ** Change that if you do not want to use C locales. (Code using this -** macro must include header 'locale.h'.) +** macro must include the header 'locale.h'.) */ #if !defined(lua_getlocaledecpoint) #define lua_getlocaledecpoint() (localeconv()->decimal_point[0]) @@ -718,7 +732,7 @@ ** {================================================================== ** Macros that affect the API and must be stable (that is, must be the ** same when you compile Lua and when you compile code that links to -** Lua). You probably do not want/need to change them. +** Lua). ** ===================================================================== */ diff --git a/src/lmathlib.c b/src/lmathlib.c index 7ef7e59..afd7607 100644 --- a/src/lmathlib.c +++ b/src/lmathlib.c @@ -1,5 +1,5 @@ /* -** $Id: lmathlib.c,v 1.119.1.1 2017/04/19 17:20:42 roberto Exp $ +** $Id: lmathlib.c $ ** Standard mathematical library ** See Copyright Notice in lua.h */ @@ -10,8 +10,11 @@ #include "lprefix.h" -#include +#include +#include #include +#include +#include #include "lua.h" @@ -23,19 +26,6 @@ #define PI (l_mathop(3.141592653589793238462643383279502884)) -#if !defined(l_rand) /* { */ -#if defined(LUA_USE_POSIX) -#define l_rand() random() -#define l_srand(x) srandom(x) -#define L_RANDMAX 2147483647 /* (2^31 - 1), following POSIX */ -#else -#define l_rand() rand() -#define l_srand(x) srand(x) -#define L_RANDMAX RAND_MAX -#endif -#endif /* } */ - - static int math_abs (lua_State *L) { if (lua_isinteger(L, 1)) { lua_Integer n = lua_tointeger(L, 1); @@ -87,7 +77,7 @@ static int math_toint (lua_State *L) { lua_pushinteger(L, n); else { luaL_checkany(L, 1); - lua_pushnil(L); /* value is not convertible to integer */ + luaL_pushfail(L); /* value is not convertible to integer */ } return 1; } @@ -239,22 +229,347 @@ static int math_max (lua_State *L) { return 1; } + +static int math_type (lua_State *L) { + if (lua_type(L, 1) == LUA_TNUMBER) + lua_pushstring(L, (lua_isinteger(L, 1)) ? "integer" : "float"); + else { + luaL_checkany(L, 1); + luaL_pushfail(L); + } + return 1; +} + + + +/* +** {================================================================== +** Pseudo-Random Number Generator based on 'xoshiro256**'. +** =================================================================== +*/ + +/* number of binary digits in the mantissa of a float */ +#define FIGS l_floatatt(MANT_DIG) + +#if FIGS > 64 +/* there are only 64 random bits; use them all */ +#undef FIGS +#define FIGS 64 +#endif + + +/* +** LUA_RAND32 forces the use of 32-bit integers in the implementation +** of the PRN generator (mainly for testing). +*/ +#if !defined(LUA_RAND32) && !defined(Rand64) + +/* try to find an integer type with at least 64 bits */ + +#if (ULONG_MAX >> 31 >> 31) >= 3 + +/* 'long' has at least 64 bits */ +#define Rand64 unsigned long + +#elif !defined(LUA_USE_C89) && defined(LLONG_MAX) + +/* there is a 'long long' type (which must have at least 64 bits) */ +#define Rand64 unsigned long long + +#elif (LUA_MAXUNSIGNED >> 31 >> 31) >= 3 + +/* 'lua_Integer' has at least 64 bits */ +#define Rand64 lua_Unsigned + +#endif + +#endif + + +#if defined(Rand64) /* { */ + +/* +** Standard implementation, using 64-bit integers. +** If 'Rand64' has more than 64 bits, the extra bits do not interfere +** with the 64 initial bits, except in a right shift. Moreover, the +** final result has to discard the extra bits. +*/ + +/* avoid using extra bits when needed */ +#define trim64(x) ((x) & 0xffffffffffffffffu) + + +/* rotate left 'x' by 'n' bits */ +static Rand64 rotl (Rand64 x, int n) { + return (x << n) | (trim64(x) >> (64 - n)); +} + +static Rand64 nextrand (Rand64 *state) { + Rand64 state0 = state[0]; + Rand64 state1 = state[1]; + Rand64 state2 = state[2] ^ state0; + Rand64 state3 = state[3] ^ state1; + Rand64 res = rotl(state1 * 5, 7) * 9; + state[0] = state0 ^ state3; + state[1] = state1 ^ state2; + state[2] = state2 ^ (state1 << 17); + state[3] = rotl(state3, 45); + return res; +} + + +/* must take care to not shift stuff by more than 63 slots */ + + +/* +** Convert bits from a random integer into a float in the +** interval [0,1), getting the higher FIG bits from the +** random unsigned integer and converting that to a float. +*/ + +/* must throw out the extra (64 - FIGS) bits */ +#define shift64_FIG (64 - FIGS) + +/* to scale to [0, 1), multiply by scaleFIG = 2^(-FIGS) */ +#define scaleFIG (l_mathop(0.5) / ((Rand64)1 << (FIGS - 1))) + +static lua_Number I2d (Rand64 x) { + return (lua_Number)(trim64(x) >> shift64_FIG) * scaleFIG; +} + +/* convert a 'Rand64' to a 'lua_Unsigned' */ +#define I2UInt(x) ((lua_Unsigned)trim64(x)) + +/* convert a 'lua_Unsigned' to a 'Rand64' */ +#define Int2I(x) ((Rand64)(x)) + + +#else /* no 'Rand64' }{ */ + +/* get an integer with at least 32 bits */ +#if LUAI_IS32INT +typedef unsigned int lu_int32; +#else +typedef unsigned long lu_int32; +#endif + + /* -** This function uses 'double' (instead of 'lua_Number') to ensure that -** all bits from 'l_rand' can be represented, and that 'RANDMAX + 1.0' -** will keep full precision (ensuring that 'r' is always less than 1.0.) +** Use two 32-bit integers to represent a 64-bit quantity. */ +typedef struct Rand64 { + lu_int32 h; /* higher half */ + lu_int32 l; /* lower half */ +} Rand64; + + +/* +** If 'lu_int32' has more than 32 bits, the extra bits do not interfere +** with the 32 initial bits, except in a right shift and comparisons. +** Moreover, the final result has to discard the extra bits. +*/ + +/* avoid using extra bits when needed */ +#define trim32(x) ((x) & 0xffffffffu) + + +/* +** basic operations on 'Rand64' values +*/ + +/* build a new Rand64 value */ +static Rand64 packI (lu_int32 h, lu_int32 l) { + Rand64 result; + result.h = h; + result.l = l; + return result; +} + +/* return i << n */ +static Rand64 Ishl (Rand64 i, int n) { + lua_assert(n > 0 && n < 32); + return packI((i.h << n) | (trim32(i.l) >> (32 - n)), i.l << n); +} + +/* i1 ^= i2 */ +static void Ixor (Rand64 *i1, Rand64 i2) { + i1->h ^= i2.h; + i1->l ^= i2.l; +} + +/* return i1 + i2 */ +static Rand64 Iadd (Rand64 i1, Rand64 i2) { + Rand64 result = packI(i1.h + i2.h, i1.l + i2.l); + if (trim32(result.l) < trim32(i1.l)) /* carry? */ + result.h++; + return result; +} + +/* return i * 5 */ +static Rand64 times5 (Rand64 i) { + return Iadd(Ishl(i, 2), i); /* i * 5 == (i << 2) + i */ +} + +/* return i * 9 */ +static Rand64 times9 (Rand64 i) { + return Iadd(Ishl(i, 3), i); /* i * 9 == (i << 3) + i */ +} + +/* return 'i' rotated left 'n' bits */ +static Rand64 rotl (Rand64 i, int n) { + lua_assert(n > 0 && n < 32); + return packI((i.h << n) | (trim32(i.l) >> (32 - n)), + (trim32(i.h) >> (32 - n)) | (i.l << n)); +} + +/* for offsets larger than 32, rotate right by 64 - offset */ +static Rand64 rotl1 (Rand64 i, int n) { + lua_assert(n > 32 && n < 64); + n = 64 - n; + return packI((trim32(i.h) >> n) | (i.l << (32 - n)), + (i.h << (32 - n)) | (trim32(i.l) >> n)); +} + +/* +** implementation of 'xoshiro256**' algorithm on 'Rand64' values +*/ +static Rand64 nextrand (Rand64 *state) { + Rand64 res = times9(rotl(times5(state[1]), 7)); + Rand64 t = Ishl(state[1], 17); + Ixor(&state[2], state[0]); + Ixor(&state[3], state[1]); + Ixor(&state[1], state[2]); + Ixor(&state[0], state[3]); + Ixor(&state[2], t); + state[3] = rotl1(state[3], 45); + return res; +} + + +/* +** Converts a 'Rand64' into a float. +*/ + +/* an unsigned 1 with proper type */ +#define UONE ((lu_int32)1) + + +#if FIGS <= 32 + +/* 2^(-FIGS) */ +#define scaleFIG (l_mathop(0.5) / (UONE << (FIGS - 1))) + +/* +** get up to 32 bits from higher half, shifting right to +** throw out the extra bits. +*/ +static lua_Number I2d (Rand64 x) { + lua_Number h = (lua_Number)(trim32(x.h) >> (32 - FIGS)); + return h * scaleFIG; +} + +#else /* 32 < FIGS <= 64 */ + +/* must take care to not shift stuff by more than 31 slots */ + +/* 2^(-FIGS) = 1.0 / 2^30 / 2^3 / 2^(FIGS-33) */ +#define scaleFIG \ + ((lua_Number)1.0 / (UONE << 30) / 8.0 / (UONE << (FIGS - 33))) + +/* +** use FIGS - 32 bits from lower half, throwing out the other +** (32 - (FIGS - 32)) = (64 - FIGS) bits +*/ +#define shiftLOW (64 - FIGS) + +/* +** higher 32 bits go after those (FIGS - 32) bits: shiftHI = 2^(FIGS - 32) +*/ +#define shiftHI ((lua_Number)(UONE << (FIGS - 33)) * 2.0) + + +static lua_Number I2d (Rand64 x) { + lua_Number h = (lua_Number)trim32(x.h) * shiftHI; + lua_Number l = (lua_Number)(trim32(x.l) >> shiftLOW); + return (h + l) * scaleFIG; +} + +#endif + + +/* convert a 'Rand64' to a 'lua_Unsigned' */ +static lua_Unsigned I2UInt (Rand64 x) { + return ((lua_Unsigned)trim32(x.h) << 31 << 1) | (lua_Unsigned)trim32(x.l); +} + +/* convert a 'lua_Unsigned' to a 'Rand64' */ +static Rand64 Int2I (lua_Unsigned n) { + return packI((lu_int32)(n >> 31 >> 1), (lu_int32)n); +} + +#endif /* } */ + + +/* +** A state uses four 'Rand64' values. +*/ +typedef struct { + Rand64 s[4]; +} RanState; + + +/* +** Project the random integer 'ran' into the interval [0, n]. +** Because 'ran' has 2^B possible values, the projection can only be +** uniform when the size of the interval is a power of 2 (exact +** division). Otherwise, to get a uniform projection into [0, n], we +** first compute 'lim', the smallest Mersenne number not smaller than +** 'n'. We then project 'ran' into the interval [0, lim]. If the result +** is inside [0, n], we are done. Otherwise, we try with another 'ran', +** until we have a result inside the interval. +*/ +static lua_Unsigned project (lua_Unsigned ran, lua_Unsigned n, + RanState *state) { + if ((n & (n + 1)) == 0) /* is 'n + 1' a power of 2? */ + return ran & n; /* no bias */ + else { + lua_Unsigned lim = n; + /* compute the smallest (2^b - 1) not smaller than 'n' */ + lim |= (lim >> 1); + lim |= (lim >> 2); + lim |= (lim >> 4); + lim |= (lim >> 8); + lim |= (lim >> 16); +#if (LUA_MAXUNSIGNED >> 31) >= 3 + lim |= (lim >> 32); /* integer type has more than 32 bits */ +#endif + lua_assert((lim & (lim + 1)) == 0 /* 'lim + 1' is a power of 2, */ + && lim >= n /* not smaller than 'n', */ + && (lim >> 1) < n); /* and it is the smallest one */ + while ((ran &= lim) > n) /* project 'ran' into [0..lim] */ + ran = I2UInt(nextrand(state->s)); /* not inside [0..n]? try again */ + return ran; + } +} + + static int math_random (lua_State *L) { lua_Integer low, up; - double r = (double)l_rand() * (1.0 / ((double)L_RANDMAX + 1.0)); + lua_Unsigned p; + RanState *state = (RanState *)lua_touserdata(L, lua_upvalueindex(1)); + Rand64 rv = nextrand(state->s); /* next pseudo-random value */ switch (lua_gettop(L)) { /* check number of arguments */ case 0: { /* no arguments */ - lua_pushnumber(L, (lua_Number)r); /* Number between 0 and 1 */ + lua_pushnumber(L, I2d(rv)); /* float between 0 and 1 */ return 1; } case 1: { /* only upper limit */ low = 1; up = luaL_checkinteger(L, 1); + if (up == 0) { /* single 0 as argument? */ + lua_pushinteger(L, I2UInt(rv)); /* full random integer */ + return 1; + } break; } case 2: { /* lower and upper limits */ @@ -266,35 +581,72 @@ static int math_random (lua_State *L) { } /* random integer in the interval [low, up] */ luaL_argcheck(L, low <= up, 1, "interval is empty"); - luaL_argcheck(L, low >= 0 || up <= LUA_MAXINTEGER + low, 1, - "interval too large"); - r *= (double)(up - low) + 1.0; - lua_pushinteger(L, (lua_Integer)r + low); + /* project random integer into the interval [0, up - low] */ + p = project(I2UInt(rv), (lua_Unsigned)up - (lua_Unsigned)low, state); + lua_pushinteger(L, p + (lua_Unsigned)low); return 1; } -static int math_randomseed (lua_State *L) { - l_srand((unsigned int)(lua_Integer)luaL_checknumber(L, 1)); - (void)l_rand(); /* discard first value to avoid undesirable correlations */ - return 0; +static void setseed (lua_State *L, Rand64 *state, + lua_Unsigned n1, lua_Unsigned n2) { + int i; + state[0] = Int2I(n1); + state[1] = Int2I(0xff); /* avoid a zero state */ + state[2] = Int2I(n2); + state[3] = Int2I(0); + for (i = 0; i < 16; i++) + nextrand(state); /* discard initial values to "spread" seed */ + lua_pushinteger(L, n1); + lua_pushinteger(L, n2); } -static int math_type (lua_State *L) { - if (lua_type(L, 1) == LUA_TNUMBER) { - if (lua_isinteger(L, 1)) - lua_pushliteral(L, "integer"); - else - lua_pushliteral(L, "float"); +/* +** Set a "random" seed. To get some randomness, use the current time +** and the address of 'L' (in case the machine does address space layout +** randomization). +*/ +static void randseed (lua_State *L, RanState *state) { + lua_Unsigned seed1 = (lua_Unsigned)time(NULL); + lua_Unsigned seed2 = (lua_Unsigned)(size_t)L; + setseed(L, state->s, seed1, seed2); +} + + +static int math_randomseed (lua_State *L) { + RanState *state = (RanState *)lua_touserdata(L, lua_upvalueindex(1)); + if (lua_isnone(L, 1)) { + randseed(L, state); } else { - luaL_checkany(L, 1); - lua_pushnil(L); + lua_Integer n1 = luaL_checkinteger(L, 1); + lua_Integer n2 = luaL_optinteger(L, 2, 0); + setseed(L, state->s, n1, n2); } - return 1; + return 2; /* return seeds */ +} + + +static const luaL_Reg randfuncs[] = { + {"random", math_random}, + {"randomseed", math_randomseed}, + {NULL, NULL} +}; + + +/* +** Register the random functions and initialize their state. +*/ +static void setrandfunc (lua_State *L) { + RanState *state = (RanState *)lua_newuserdata(L, sizeof(RanState)); + randseed(L, state); /* initialize with a "random" seed */ + lua_pop(L, 2); /* remove pushed seeds */ + luaL_setfuncs(L, randfuncs, 1); } +/* }================================================================== */ + /* ** {================================================================== @@ -367,8 +719,6 @@ static const luaL_Reg mathlib[] = { {"min", math_min}, {"modf", math_modf}, {"rad", math_rad}, - {"random", math_random}, - {"randomseed", math_randomseed}, {"sin", math_sin}, {"sqrt", math_sqrt}, {"tan", math_tan}, @@ -384,6 +734,8 @@ static const luaL_Reg mathlib[] = { {"log10", math_log10}, #endif /* placeholders */ + {"random", NULL}, + {"randomseed", NULL}, {"pi", NULL}, {"huge", NULL}, {"maxinteger", NULL}, @@ -405,6 +757,7 @@ LUAMOD_API int luaopen_math (lua_State *L) { lua_setfield(L, -2, "maxinteger"); lua_pushinteger(L, LUA_MININTEGER); lua_setfield(L, -2, "mininteger"); + setrandfunc(L); return 1; } diff --git a/src/lstrlib.c b/src/lstrlib.c index 7c29293..908693e 100644 --- a/src/lstrlib.c +++ b/src/lstrlib.c @@ -14,6 +14,7 @@ #include #include #include +#include #include #include #include @@ -969,7 +970,7 @@ static int str_gsub (lua_State *L) { ** to nibble boundaries by making what is left after that first digit a ** multiple of 4. */ -#define L_NBFD ((l_mathlim(MANT_DIG) - 1)%4 + 1) +#define L_NBFD ((l_floatatt(MANT_DIG) - 1)%4 + 1) /* @@ -1032,13 +1033,23 @@ static int lua_number2strx (lua_State *L, char *buff, int sz, /* -** Maximum size of each formatted item. This maximum size is produced +** Maximum size for items formatted with '%f'. This size is produced ** by format('%.99f', -maxfloat), and is equal to 99 + 3 ('-', '.', ** and '\0') + number of decimal digits to represent maxfloat (which -** is maximum exponent + 1). (99+3+1 then rounded to 120 for "extra -** expenses", such as locale-dependent stuff) +** is maximum exponent + 1). (99+3+1, adding some extra, 110) +*/ +#define MAX_ITEMF (110 + l_floatatt(MAX_10_EXP)) + + +/* +** All formats except '%f' do not need that large limit. The other +** float formats use exponents, so that they fit in the 99 limit for +** significant digits; 's' for large strings and 'q' add items directly +** to the buffer; all integer formats also fit in the 99 limit. The +** worst case are floats: they may need 99 significant digits, plus +** '0x', '-', '.', 'e+XXXX', and '\0'. Adding some extra, 120. */ -#define MAX_ITEM (120 + l_mathlim(MAX_10_EXP)) +#define MAX_ITEM 120 /* valid flags in a format specification */ @@ -1173,36 +1184,51 @@ static int str_format (lua_State *L) { luaL_addchar(&b, *strfrmt++); /* %% */ else { /* format item */ char form[MAX_FORMAT]; /* to store the format ('%...') */ - char *buff = luaL_prepbuffsize(&b, MAX_ITEM); /* to put formatted item */ + int maxitem = MAX_ITEM; + char *buff = luaL_prepbuffsize(&b, maxitem); /* to put formatted item */ int nb = 0; /* number of bytes in added item */ if (++arg > top) - luaL_argerror(L, arg, "no value"); + return luaL_argerror(L, arg, "no value"); strfrmt = scanformat(L, strfrmt, form); switch (*strfrmt++) { case 'c': { - nb = l_sprintf(buff, MAX_ITEM, form, (int)luaL_checkinteger(L, arg)); + nb = l_sprintf(buff, maxitem, form, (int)luaL_checkinteger(L, arg)); break; } case 'd': case 'i': case 'o': case 'u': case 'x': case 'X': { lua_Integer n = luaL_checkinteger(L, arg); addlenmod(form, LUA_INTEGER_FRMLEN); - nb = l_sprintf(buff, MAX_ITEM, form, (LUAI_UACINT)n); + nb = l_sprintf(buff, maxitem, form, (LUAI_UACINT)n); break; } case 'a': case 'A': addlenmod(form, LUA_NUMBER_FRMLEN); - nb = lua_number2strx(L, buff, MAX_ITEM, form, + nb = lua_number2strx(L, buff, maxitem, form, luaL_checknumber(L, arg)); break; - case 'e': case 'E': case 'f': - case 'g': case 'G': { + case 'f': + maxitem = MAX_ITEMF; /* extra space for '%f' */ + buff = luaL_prepbuffsize(&b, maxitem); + /* FALLTHROUGH */ + case 'e': case 'E': case 'g': case 'G': { lua_Number n = luaL_checknumber(L, arg); addlenmod(form, LUA_NUMBER_FRMLEN); - nb = l_sprintf(buff, MAX_ITEM, form, (LUAI_UACNUMBER)n); + nb = l_sprintf(buff, maxitem, form, (LUAI_UACNUMBER)n); + break; + } + case 'p': { + const void *p = lua_topointer(L, arg); + if (p == NULL) { /* avoid calling 'printf' with argument NULL */ + p = "(null)"; /* result */ + form[strlen(form) - 1] = 's'; /* format it as a string */ + } + nb = l_sprintf(buff, maxitem, form, p); break; } case 'q': { + if (form[2] != '\0') /* modifiers? */ + return luaL_error(L, "specifier '%%q' cannot have modifiers"); addliteral(L, &b, arg); break; } @@ -1218,18 +1244,17 @@ static int str_format (lua_State *L) { luaL_addvalue(&b); /* keep entire string */ } else { /* format the string into 'buff' */ - nb = l_sprintf(buff, MAX_ITEM, form, s); + nb = l_sprintf(buff, maxitem, form, s); lua_pop(L, 1); /* remove result from 'luaL_tolstring' */ } } break; } default: { /* also treat cases 'pnLlh' */ - return luaL_error(L, "invalid option '%%%c' to 'format'", - *(strfrmt - 1)); + return luaL_error(L, "invalid conversion '%s' to 'format'", form); } } - lua_assert(nb < MAX_ITEM); + lua_assert(nb < maxitem); luaL_addsize(&b, nb); } } diff --git a/src/lvm.c b/src/lvm.c index 0dcf5c1..7ce6156 100644 --- a/src/lvm.c +++ b/src/lvm.c @@ -42,27 +42,33 @@ /* -** 'l_intfitsf' checks whether a given integer can be converted to a -** float without rounding. Used in comparisons. Left undefined if -** all integers fit in a float precisely. +** 'l_intfitsf' checks whether a given integer is in the range that +** can be converted to a float without rounding. Used in comparisons. */ #if !defined(l_intfitsf) /* number of bits in the mantissa of a float */ -#define NBM (l_mathlim(MANT_DIG)) +#define NBM (l_floatatt(MANT_DIG)) /* -** Check whether some integers may not fit in a float, that is, whether -** (maxinteger >> NBM) > 0 (that implies (1 << NBM) <= maxinteger). -** (The shifts are done in parts to avoid shifting by more than the size +** Check whether some integers may not fit in a float, testing whether +** (maxinteger >> NBM) > 0. (That implies (1 << NBM) <= maxinteger.) +** (The shifts are done in parts, to avoid shifting by more than the size ** of an integer. In a worst case, NBM == 113 for long double and -** sizeof(integer) == 32.) +** sizeof(long) == 32.) */ #if ((((LUA_MAXINTEGER >> (NBM / 4)) >> (NBM / 4)) >> (NBM / 4)) \ >> (NBM - (3 * (NBM / 4)))) > 0 -#define l_intfitsf(i) \ - (-((lua_Integer)1 << NBM) <= (i) && (i) <= ((lua_Integer)1 << NBM)) +/* limit for integers that fit in a float */ +#define MAXINTFITSF ((lua_Unsigned)1 << NBM) + +/* check whether 'i' is in the interval [-MAXINTFITSF, MAXINTFITSF] */ +#define l_intfitsf(i) ((MAXINTFITSF + l_castS2U(i)) <= (2 * MAXINTFITSF)) + +#else /* all integers fit in a float precisely */ + +#define l_intfitsf(i) 1 #endif diff --git a/tests/lua53/math.lua b/tests/lua53/math.lua index 6d29c7c..459534e 100644 --- a/tests/lua53/math.lua +++ b/tests/lua53/math.lua @@ -725,100 +725,220 @@ end print("testing 'math.random'") -math.randomseed(0) + +local random, max, min = math.random, math.max, math.min + +local function testnear (val, ref, tol) + return (math.abs(val - ref) < ref * tol) +end + + +-- low-level!! For the current implementation of random in Lua, +-- the first call after seed 1007 should return 0x7a7040a5a323c9d6 +do + -- all computations should work with 32-bit integers + local h = 0x7a7040a5 -- higher half + local l = 0xa323c9d6 -- lower half + + math.randomseed(1007) + -- get the low 'intbits' of the 64-bit expected result + local res = (h << 32 | l) & ~(~0 << intbits) + assert(random(0) == res) + + math.randomseed(1007, 0) + -- using higher bits to generate random floats; (the '% 2^32' converts + -- 32-bit integers to floats as unsigned) + local res + if floatbits <= 32 then + -- get all bits from the higher half + res = (h >> (32 - floatbits)) % 2^32 + else + -- get 32 bits from the higher half and the rest from the lower half + res = (h % 2^32) * 2^(floatbits - 32) + ((l >> (64 - floatbits)) % 2^32) + end + local rand = random() + assert(eq(rand, 0x0.7a7040a5a323c9d6, 2^-floatbits)) + assert(rand * 2^floatbits == res) +end + +do + -- testing return of 'randomseed' + local x, y = math.randomseed() + local res = math.random(0) + x, y = math.randomseed(x, y) -- should repeat the state + assert(math.random(0) == res) + math.randomseed(x, y) -- again should repeat the state + assert(math.random(0) == res) + -- keep the random seed for following tests +end do -- test random for floats - local max = -math.huge - local min = math.huge - for i = 0, 20000 do - local t = math.random() + local randbits = math.min(floatbits, 64) -- at most 64 random bits + local mult = 2^randbits -- to make random float into an integral + local counts = {} -- counts for bits + for i = 1, randbits do counts[i] = 0 end + local up = -math.huge + local low = math.huge + local rounds = 100 * randbits -- 100 times for each bit + local totalrounds = 0 + ::doagain:: -- will repeat test until we get good statistics + for i = 0, rounds do + local t = random() assert(0 <= t and t < 1) - max = math.max(max, t) - min = math.min(min, t) - if eq(max, 1, 0.001) and eq(min, 0, 0.001) then - goto ok + up = max(up, t) + low = min(low, t) + assert(t * mult % 1 == 0) -- no extra bits + local bit = i % randbits -- bit to be tested + if (t * 2^bit) % 1 >= 0.5 then -- is bit set? + counts[bit + 1] = counts[bit + 1] + 1 -- increment its count + end + end + totalrounds = totalrounds + rounds + if not (eq(up, 1, 0.001) and eq(low, 0, 0.001)) then + goto doagain + end + -- all bit counts should be near 50% + local expected = (totalrounds / randbits / 2) + for i = 1, randbits do + if not testnear(counts[i], expected, 0.10) then + goto doagain + end + end + print(string.format("float random range in %d calls: [%f, %f]", + totalrounds, low, up)) +end + + +do -- test random for full integers + local up = 0 + local low = 0 + local counts = {} -- counts for bits + for i = 1, intbits do counts[i] = 0 end + local rounds = 100 * intbits -- 100 times for each bit + local totalrounds = 0 + ::doagain:: -- will repeat test until we get good statistics + for i = 0, rounds do + local t = random(0) + up = max(up, t) + low = min(low, t) + local bit = i % intbits -- bit to be tested + -- increment its count if it is set + counts[bit + 1] = counts[bit + 1] + ((t >> bit) & 1) + end + totalrounds = totalrounds + rounds + local lim = maxint >> 10 + if not (maxint - up < lim and low - minint < lim) then + goto doagain + end + -- all bit counts should be near 50% + local expected = (totalrounds / intbits / 2) + for i = 1, intbits do + if not testnear(counts[i], expected, 0.10) then + goto doagain end end - -- loop ended without satisfing condition - assert(false) - ::ok:: + print(string.format( + "integer random range in %d calls: [minint + %.0fppm, maxint - %.0fppm]", + totalrounds, (minint - low) / minint * 1e6, + (maxint - up) / maxint * 1e6)) end do - local function aux (p, lim) -- test random for small intervals - local x1, x2 - if #p == 1 then x1 = 1; x2 = p[1] - else x1 = p[1]; x2 = p[2] + -- test distribution for a dice + local count = {0, 0, 0, 0, 0, 0} + local rep = 200 + local totalrep = 0 + ::doagain:: + for i = 1, rep * 6 do + local r = random(6) + count[r] = count[r] + 1 + end + totalrep = totalrep + rep + for i = 1, 6 do + if not testnear(count[i], totalrep, 0.05) then + goto doagain end + end +end + +do + local function aux (x1, x2) -- test random for small intervals local mark = {}; local count = 0 -- to check that all values appeared - for i = 0, lim or 2000 do - local t = math.random(table.unpack(p)) + while true do + local t = random(x1, x2) assert(x1 <= t and t <= x2) if not mark[t] then -- new value mark[t] = true count = count + 1 - end - if count == x2 - x1 + 1 then -- all values appeared; OK - goto ok + if count == x2 - x1 + 1 then -- all values appeared; OK + goto ok + end end end - -- loop ended without satisfing condition - assert(false) ::ok:: end - aux({-10,0}) - aux({6}) - aux({-10, 10}) - aux({minint, minint}) - aux({maxint, maxint}) - aux({minint, minint + 9}) - aux({maxint - 3, maxint}) + aux(-10,0) + aux(1, 6) + aux(1, 2) + aux(1, 13) + aux(1, 31) + aux(1, 32) + aux(1, 33) + aux(-10, 10) + aux(-10,-10) -- unit set + aux(minint, minint) -- unit set + aux(maxint, maxint) -- unit set + aux(minint, minint + 9) + aux(maxint - 3, maxint) end do local function aux(p1, p2) -- test random for large intervals local max = minint local min = maxint - local n = 200 + local n = 100 local mark = {}; local count = 0 -- to count how many different values + ::doagain:: for _ = 1, n do - local t = math.random(p1, p2) - max = math.max(max, t) - min = math.min(min, t) + local t = random(p1, p2) if not mark[t] then -- new value + assert(p1 <= t and t <= p2) + max = math.max(max, t) + min = math.min(min, t) mark[t] = true count = count + 1 end end -- at least 80% of values are different - assert(count >= n * 0.8) + if not (count >= n * 0.8) then + goto doagain + end -- min and max not too far from formal min and max - local diff = (p2 - p1) // 8 - assert(min < p1 + diff and max > p2 - diff) + local diff = (p2 - p1) >> 4 + if not (min < p1 + diff and max > p2 - diff) then + goto doagain + end end aux(0, maxint) aux(1, maxint) + aux(3, maxint // 3) aux(minint, -1) aux(minint // 2, maxint // 2) + aux(minint, maxint) + aux(minint + 1, maxint) + aux(minint, maxint - 1) + aux(0, 1 << (intbits - 5)) end -for i=1,100 do - assert(math.random(maxint) > 0) - assert(math.random(minint, -1) < 0) -end -assert(not pcall(math.random, 1, 2, 3)) -- too many arguments +assert(not pcall(random, 1, 2, 3)) -- too many arguments -- empty interval -assert(not pcall(math.random, minint + 1, minint)) -assert(not pcall(math.random, maxint, maxint - 1)) -assert(not pcall(math.random, maxint, minint)) - --- interval too large -assert(not pcall(math.random, minint, 0)) -assert(not pcall(math.random, -1, maxint)) -assert(not pcall(math.random, minint // 2, maxint // 2 + 1)) +assert(not pcall(random, minint + 1, minint)) +assert(not pcall(random, maxint, maxint - 1)) +assert(not pcall(random, maxint, minint)) + print('OK') diff --git a/tests/lua53/strings.lua b/tests/lua53/strings.lua index c37a8e7..04f2beb 100644 --- a/tests/lua53/strings.lua +++ b/tests/lua53/strings.lua @@ -269,8 +269,8 @@ do print("testing 'format %a %A'") matchhexa(n) end - assert(string.find(string.format("%A", 0.0), "^0X0%.?0?P%+?0$")) - assert(string.find(string.format("%a", -0.0), "^%-0x0%.?0?p%+?0$")) + assert(string.find(string.format("%A", 0.0), "^0X0%.?0*P%+?0$")) + assert(string.find(string.format("%a", -0.0), "^%-0x0%.?0*p%+?0$")) if not _port then -- test inf, -inf, NaN, and -0.0 assert(string.find(string.format("%a", 1/0), "^inf")) @@ -299,7 +299,7 @@ check("%100.3d", "too long") check("%1"..aux..".3d", "too long") check("%1.100d", "too long") check("%10.1"..aux.."004d", "too long") -check("%t", "invalid option") +check("%t", "invalid conversion") check("%"..aux.."d", "repeated flags") check("%d %d", "no value")