X-Git-Url: https://git.xonotic.org/?p=xonotic%2Fgmqcc.git;a=blobdiff_plain;f=fold.c;h=4f6ca5f349c61d9d1c668fc58bdcb62eb845f9ab;hp=7b87dc280ad55378cb82a38e6c2cd0acf300dbe4;hb=e3577912c86755592035786a8fdceedf5331f9e0;hpb=b08195e2da48e041a4f027d93e944bc4715169ec diff --git a/fold.c b/fold.c index 7b87dc2..4f6ca5f 100644 --- a/fold.c +++ b/fold.c @@ -34,15 +34,16 @@ #define FOLD_TINYNESS SFLOAT_TBEFORE /* - * The constant folder is also responsible for validating if the constant - * expressions produce valid results. We cannot trust the FPU control - * unit for these exceptions because setting FPU control words might not - * work. Systems can set and enforce FPU modes of operation. It's also valid - * for libc's to simply ignore FPU exceptions. For instance ARM CPUs in - * glibc. We implement some trivial and IEE 754 conformant functions which - * emulate those operations. This is an entierly optional compiler feature - * which shouldn't be enabled for anything other than performing strict - * passes on constant expressions since it's quite slow. + * Comparing float values is an unsafe operation when the operands to the + * comparison are floating point values that are inexact. For instance 1/3 is an + * inexact value. The FPU is meant to raise exceptions when these sorts of things + * happen, including division by zero, underflows and overflows. The C standard + * library provides us with the header to gain access to the floating- + * point environment and lets us set the rounding mode and check for these exceptions. + * The problem is the standard C library allows an implementation to leave these + * stubbed out and does not require they be implemented. Furthermore, depending + * on implementations there is no control over the FPU. This is an IEE 754 + * conforming implementation in software to compensate. */ typedef uint32_t sfloat_t; @@ -78,7 +79,7 @@ typedef struct { sfloat_tdetect_t tiny; } sfloat_state_t; -/* Count of leading zero bits before the most-significand 1 bit. */ +/* Counts the number of leading zero bits before the most-significand one bit. */ #ifdef _MSC_VER /* MSVC has an intrinsic for this */ static GMQCC_INLINE uint32_t sfloat_clz(uint32_t x) { @@ -115,7 +116,7 @@ typedef struct { #endif /* The value of a NaN */ -#define SFLOAT_NAN 0xFFC00000 +#define SFLOAT_NAN 0xFFFFFFFF /* Test if NaN */ #define SFLOAT_ISNAN(A) \ (0xFF000000 < (uint32_t)((A) << 1)) @@ -126,8 +127,12 @@ typedef struct { #define SFLOAT_RAISE(STATE, FLAGS) \ ((STATE)->exceptionflags = (sfloat_exceptionflags_t)((STATE)->exceptionflags | (FLAGS))) /* - * Shifts `A' right `COUNT' bits. Non-zero bits are stored in LSB. Size - * sets the arbitrarly-large limit. + * Shifts `A' right by the number of bits given in `COUNT'. If any non-zero bits + * are shifted off they are forced into the least significand bit of the result + * by setting it to one. As a result of this, the value of `COUNT' can be + * arbitrarily large; if `COUNT' is greater than 32, the result will be either + * zero or one, depending on whether `A' is a zero or non-zero. The result is + * stored into the value pointed by `Z'. */ #define SFLOAT_SHIFT(SIZE, A, COUNT, Z) \ *(Z) = ((COUNT) == 0) \ @@ -135,6 +140,7 @@ typedef struct { : (((COUNT) < (SIZE)) \ ? ((A) >> (COUNT)) | (((A) << ((-(COUNT)) & ((SIZE) - 1))) != 0) \ : ((A) != 0)) + /* Extract fractional component */ #define SFLOAT_EXTRACT_FRAC(X) \ ((uint32_t)((X) & 0x007FFFFF)) @@ -144,21 +150,32 @@ typedef struct { /* Extract sign bit */ #define SFLOAT_EXTRACT_SIGN(X) \ ((X) >> 31) -/* Normalize a subnormal */ +/* + * Normalizes the subnormal value represented by the denormalized significand + * `SA'. The normalized exponent and significand are stored at the locations + * pointed by `Z' and `SZ' respectively. + */ #define SFLOAT_SUBNORMALIZE(SA, Z, SZ) \ - (void)(*(SZ) = (SA) << SFLOAT_CLZ((SA), 8), *(SZ) = 1 - SFLOAT_CLZ((SA), 8)) + (void)(*(SZ) = (SA) << SFLOAT_CLZ((SA), 8), *(Z) = 1 - SFLOAT_CLZ((SA), 8)) /* - * Pack sign, exponent and significand and produce a float. + * Packs the sign `SIGN', exponent `EXP' and significand `SIG' into the value + * giving the result. * - * Integer portions of the significand are added to the exponent. The - * exponent input should be one less than the result exponent whenever - * the significand is normalized since normalized significand will - * always have an integer portion of value one. + * After the shifting into their proper positions, the fields are added together + * to form the result. This means any integer portion of `SIG' will be added + * to the exponent. Similarly, because a properly normalized significand will + * always have an integer portion equal to one, the exponent input `EXP' should + * be one less than the desired result exponent whenever the significant input + * `SIG' is a complete, normalized significand. */ #define SFLOAT_PACK(SIGN, EXP, SIG) \ (sfloat_t)((((uint32_t)(SIGN)) << 31) + (((uint32_t)(EXP)) << 23) + (SIG)) -/* Calculate NaN. If either operands are signaling then raise invalid */ +/* + * Takes two values `a' and `b', one of which is a NaN, and returns the appropriate + * NaN result. If either `a' or `b' is a signaling NaN than an invalid exception is + * raised. + */ static sfloat_t sfloat_propagate_nan(sfloat_state_t *state, sfloat_t a, sfloat_t b) { bool isnan_a = SFLOAT_ISNAN(a); bool issnan_a = SFLOAT_ISSNAN(a); @@ -169,23 +186,33 @@ static sfloat_t sfloat_propagate_nan(sfloat_state_t *state, sfloat_t a, sfloat_t b |= 0x00400000; if (issnan_a | issnan_b) - SFLOAT_RAISE(state, SFLOAT_INEXACT); - if (issnan_a) { - if (issnan_b) - goto larger; - return isnan_b ? b : a; - } else if (isnan_a) { - if (issnan_b | !isnan_b) - return a; -larger: - if ((uint32_t)(a << 1) < (uint32_t)(b << 1)) return b; - if ((uint32_t)(b << 1) < (uint32_t)(a << 1)) return a; - return (a < b) ? a : b; - } + SFLOAT_RAISE(state, SFLOAT_INVALID); + if (isnan_a) + return (issnan_a & isnan_b) ? b : a; return b; } -/* Round and pack */ +/* + * Takes an abstract value having sign `sign_z', exponent `exp_z', and significand + * `sig_z' and returns the appropriate value corresponding to the abstract input. + * + * The abstract value is simply rounded and packed into the format. If the abstract + * input cannot be represented exactly an inexact exception is raised. If the + * abstract input is too large, the overflow and inexact exceptions are both raised + * and an infinity or maximal finite value is returned. If the abstract value is + * too small, the value is rounded to a subnormal and the underflow and inexact + * exceptions are only raised if the value cannot be represented exactly with + * a subnormal. + * + * The input significand `sig_z' has it's binary point between bits 30 and 29, + * this is seven bits to the left of its usual location. The shifted significand + * must be normalized or smaller than this. If it's not normalized then the exponent + * `exp_z' must be zero; in that case, the result returned is a subnormal number + * and is must not require rounding. In the more usual case where the significand + * is normalized, the exponent must be one less than the *true* exponent. + * + * The handling of underflow and overflow is otherwise in alignment with IEC/IEEE. + */ static sfloat_t SFLOAT_PACK_round(sfloat_state_t *state, bool sign_z, int16_t exp_z, uint32_t sig_z) { sfloat_roundingmode_t mode = state->roundingmode; bool even = !!(mode == SFLOAT_ROUND_NEAREST_EVEN); @@ -222,13 +249,6 @@ static sfloat_t SFLOAT_PACK_round(sfloat_state_t *state, bool sign_z, int16_t ex SFLOAT_RAISE(state, SFLOAT_UNDERFLOW); } } - - /* - * Significand has point between bits 30 and 29, 7 bits to the left of - * the usual place. This shifted significand has to be normalized - * or smaller, if it isn't the exponent must be zero, in which case - * no rounding occurs since the result will be a subnormal. - */ if (bits) SFLOAT_RAISE(state, SFLOAT_INEXACT); sig_z = (sig_z + increment) >> 7; @@ -238,12 +258,24 @@ static sfloat_t SFLOAT_PACK_round(sfloat_state_t *state, bool sign_z, int16_t ex return SFLOAT_PACK(sign_z, exp_z, sig_z); } -/* Normalized round and pack */ +/* + * Takes an abstract value having sign `sign_z', exponent `exp_z' and significand + * `sig_z' and returns the appropriate value corresponding to the abstract input. + * This function is exactly like `PACK_round' except the significand does not have + * to be normalized. + * + * Bit 31 of the significand must be zero and the exponent must be one less than + * the *true* exponent. + */ static sfloat_t SFLOAT_PACK_normal(sfloat_state_t *state, bool sign_z, int16_t exp_z, uint32_t sig_z) { unsigned char c = SFLOAT_CLZ(sig_z, 1); return SFLOAT_PACK_round(state, sign_z, exp_z - c, sig_z << c); } +/* + * Returns the result of adding the absolute values of `a' and `b'. The sign + * `sign_z' is ignored if the result is a NaN. + */ static sfloat_t sfloat_add_impl(sfloat_state_t *state, sfloat_t a, sfloat_t b, bool sign_z) { int16_t exp_a = SFLOAT_EXTRACT_EXP(a); int16_t exp_b = SFLOAT_EXTRACT_EXP(b); @@ -291,6 +323,11 @@ end: return SFLOAT_PACK_round(state, sign_z, exp_z, sig_z); } +/* + * Returns the result of subtracting the absolute values of `a' and `b'. If the + * sign `sign_z' is one, the difference is negated before being returned. The + * sign is ignored if the result is a NaN. + */ static sfloat_t sfloat_sub_impl(sfloat_state_t *state, sfloat_t a, sfloat_t b, bool sign_z) { int16_t exp_a = SFLOAT_EXTRACT_EXP(a); int16_t exp_b = SFLOAT_EXTRACT_EXP(b); @@ -470,6 +507,12 @@ static sfloat_t sfloat_div(sfloat_state_t *state, sfloat_t a, sfloat_t b) { return SFLOAT_PACK_round(state, sign_z, exp_z, sig_z); } +static sfloat_t sfloat_neg(sfloat_state_t *state, sfloat_t a) { + sfloat_cast_t neg; + neg.f = -1; + return sfloat_mul(state, a, neg.s); +} + static GMQCC_INLINE void sfloat_check(lex_ctx_t ctx, sfloat_state_t *state, const char *vec) { /* Exception comes from vector component */ if (vec) { @@ -607,8 +650,31 @@ static GMQCC_INLINE vec3_t vec3_sub(lex_ctx_t ctx, vec3_t a, vec3_t b) { return out; } -static GMQCC_INLINE vec3_t vec3_neg(vec3_t a) { - vec3_t out; +static GMQCC_INLINE vec3_t vec3_neg(lex_ctx_t ctx, vec3_t a) { + vec3_t out; + sfloat_cast_t v[3]; + sfloat_state_t s[3]; + + if (!OPTS_FLAG(ARITHMETIC_EXCEPTIONS)) + goto end; + + v[0].f = a.x; + v[1].f = a.y; + v[2].f = a.z; + + sfloat_init(&s[0]); + sfloat_init(&s[1]); + sfloat_init(&s[2]); + + sfloat_neg(&s[0], v[0].s); + sfloat_neg(&s[1], v[1].s); + sfloat_neg(&s[2], v[2].s); + + sfloat_check(ctx, &s[0], NULL); + sfloat_check(ctx, &s[1], NULL); + sfloat_check(ctx, &s[2], NULL); + +end: out.x = -a.x; out.y = -a.y; out.z = -a.z; @@ -977,6 +1043,58 @@ ast_expression *fold_constgen_string(fold_t *fold, const char *str, bool transla return (ast_expression*)out; } +typedef union { + void (*callback)(void); + sfloat_t (*binary)(sfloat_state_t *, sfloat_t, sfloat_t); + sfloat_t (*unary)(sfloat_state_t *, sfloat_t); +} float_check_callback_t; + +static bool fold_check_except_float_impl(void (*callback)(void), + fold_t *fold, + ast_value *a, + ast_value *b) +{ + float_check_callback_t call; + sfloat_state_t s; + sfloat_cast_t ca; + + if (!OPTS_FLAG(ARITHMETIC_EXCEPTIONS) && !OPTS_WARN(WARN_INEXACT_COMPARES)) + return false; + + call.callback = callback; + sfloat_init(&s); + ca.f = fold_immvalue_float(a); + if (b) { + sfloat_cast_t cb; + cb.f = fold_immvalue_float(b); + call.binary(&s, ca.s, cb.s); + } else { + call.unary(&s, ca.s); + } + + if (s.exceptionflags == 0) + return false; + + if (!OPTS_FLAG(ARITHMETIC_EXCEPTIONS)) + goto inexact_possible; + + sfloat_check(fold_ctx(fold), &s, NULL); + +inexact_possible: + return s.exceptionflags & SFLOAT_INEXACT; +} + +#define fold_check_except_float(CALLBACK, FOLD, A, B) \ + fold_check_except_float_impl(((void (*)(void))(CALLBACK)), (FOLD), (A), (B)) + +static bool fold_check_inexact_float(fold_t *fold, ast_value *a, ast_value *b) { + lex_ctx_t ctx = fold_ctx(fold); + if (!OPTS_WARN(WARN_INEXACT_COMPARES)) + return false; + if (!a->inexact && !b->inexact) + return false; + return compile_warning(ctx, WARN_INEXACT_COMPARES, "inexact value in comparison"); +} static GMQCC_INLINE ast_expression *fold_op_mul_vec(fold_t *fold, vec3_t vec, ast_value *sel, const char *set) { /* @@ -1017,11 +1135,14 @@ static GMQCC_INLINE ast_expression *fold_op_mul_vec(fold_t *fold, vec3_t vec, as static GMQCC_INLINE ast_expression *fold_op_neg(fold_t *fold, ast_value *a) { if (isfloat(a)) { - if (fold_can_1(a)) - return fold_constgen_float(fold, -fold_immvalue_float(a), false); + if (fold_can_1(a)) { + /* Negation can produce inexact as well */ + bool inexact = fold_check_except_float(&sfloat_neg, fold, a, NULL); + return fold_constgen_float(fold, -fold_immvalue_float(a), inexact); + } } else if (isvector(a)) { if (fold_can_1(a)) - return fold_constgen_vector(fold, vec3_neg(fold_immvalue_vector(a))); + return fold_constgen_vector(fold, vec3_neg(fold_ctx(fold), fold_immvalue_vector(a))); } return NULL; } @@ -1044,44 +1165,6 @@ static GMQCC_INLINE ast_expression *fold_op_not(fold_t *fold, ast_value *a) { return NULL; } -static bool fold_check_except_float(sfloat_t (*callback)(sfloat_state_t *, sfloat_t, sfloat_t), - fold_t *fold, - ast_value *a, - ast_value *b) -{ - sfloat_state_t s; - sfloat_cast_t ca; - sfloat_cast_t cb; - - if (!OPTS_FLAG(ARITHMETIC_EXCEPTIONS) && !OPTS_WARN(WARN_INEXACT_COMPARES)) - return false; - - sfloat_init(&s); - ca.f = fold_immvalue_float(a); - cb.f = fold_immvalue_float(b); - - callback(&s, ca.s, cb.s); - if (s.exceptionflags == 0) - return false; - - if (!OPTS_FLAG(ARITHMETIC_EXCEPTIONS)) - goto inexact_possible; - - sfloat_check(fold_ctx(fold), &s, NULL); - -inexact_possible: - return s.exceptionflags & SFLOAT_INEXACT; -} - -static bool fold_check_inexact_float(fold_t *fold, ast_value *a, ast_value *b) { - lex_ctx_t ctx = fold_ctx(fold); - if (!OPTS_WARN(WARN_INEXACT_COMPARES)) - return false; - if (!a->inexact && !b->inexact) - return false; - return compile_warning(ctx, WARN_INEXACT_COMPARES, "inexact value in comparison"); -} - static GMQCC_INLINE ast_expression *fold_op_add(fold_t *fold, ast_value *a, ast_value *b) { if (isfloat(a)) { if (fold_can_2(a, b)) {