From b14ef0e24bcb70d70360690130b2bf2a0c8f8d33 Mon Sep 17 00:00:00 2001 From: Edmund Grimley Evans Date: Fri, 13 Feb 2015 18:58:31 +0000 Subject: Add arm64 (AArch64) as a target architecture. --- lib/Makefile | 12 ++ lib/lib-arm64.c | 652 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ lib/testfp.c | 510 ++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 1174 insertions(+) create mode 100644 lib/lib-arm64.c create mode 100644 lib/testfp.c (limited to 'lib') diff --git a/lib/Makefile b/lib/Makefile index 37d4711..6e192b9 100644 --- a/lib/Makefile +++ b/lib/Makefile @@ -28,6 +28,11 @@ ifndef TARGET # native library ifeq ($(ARCH),arm) TARGET = arm XCC = $(CC) + else + ifeq ($(ARCH),arm64) + TARGET = arm64 + else + endif endif endif endif @@ -49,6 +54,7 @@ X86_64_O = libtcc1.o alloca86_64.o ARM_O = libtcc1.o armeabi.o alloca-arm.o WIN32_O = $(I386_O) crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o WIN64_O = $(X86_64_O) crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o +ARM64_O = lib-arm64.o # build TCC runtime library to contain PIC code, so it can be linked # into shared libraries @@ -86,6 +92,11 @@ ifeq "$(TARGET)" "arm" OBJ = $(addprefix $(DIR)/,$(ARM_O)) TGT = -DTCC_TARGET_ARM XCC ?= $(TCC) -B$(TOP) +else +ifeq "$(TARGET)" "arm64" + OBJ = $(addprefix $(DIR)/,$(ARM64_O)) + TGT = -DTCC_TARGET_ARM64 + XCC ?= $(TCC) -B$(TOP) else $(error libtcc1.a not supported on target '$(TARGET)') endif @@ -93,6 +104,7 @@ endif endif endif endif +endif XFLAGS = $(CPPFLAGS) $(CFLAGS) $(PICFLAGS) $(TGT) diff --git a/lib/lib-arm64.c b/lib/lib-arm64.c new file mode 100644 index 0000000..fd73506 --- /dev/null +++ b/lib/lib-arm64.c @@ -0,0 +1,652 @@ +/* + * TCC runtime library for arm64. + * + * Copyright (c) 2015 Edmund Grimley Evans + * + * Copying and distribution of this file, with or without modification, + * are permitted in any medium without royalty provided the copyright + * notice and this notice are preserved. This file is offered as-is, + * without any warranty. + */ + +#include +#include + +void __clear_cache(char *beg, char *end) +{ +#warning __clear_cache not yet implemented +} + +typedef struct { + uint64_t x0, x1; +} u128_t; + +static long double f3_zero(int sgn) +{ + long double f; + u128_t x = { 0, (uint64_t)sgn << 63 }; + memcpy(&f, &x, 16); + return f; +} + +static long double f3_infinity(int sgn) +{ + long double f; + u128_t x = { 0, (uint64_t)sgn << 63 | 0x7fff000000000000 }; + memcpy(&f, &x, 16); + return f; +} + +static long double f3_NaN(void) +{ + long double f; +#if 0 + // ARM's default NaN usually has just the top fraction bit set: + u128_t x = { 0, 0x7fff800000000000 }; +#else + // GCC's library sets all fraction bits: + u128_t x = { -1, 0x7fffffffffffffff }; +#endif + memcpy(&f, &x, 16); + return f; +} + +static int fp3_convert_NaN(long double *f, int sgn, u128_t mnt) +{ + u128_t x = { mnt.x0, + mnt.x1 | 0x7fff800000000000 | (uint64_t)sgn << 63 }; + memcpy(f, &x, 16); + return 1; +} + +static int fp3_detect_NaNs(long double *f, + int a_sgn, int a_exp, u128_t a, + int b_sgn, int b_exp, u128_t b) +{ + // Detect signalling NaNs: + if (a_exp == 32767 && (a.x0 | a.x1 << 16) && !(a.x1 >> 47 & 1)) + return fp3_convert_NaN(f, a_sgn, a); + if (b_exp == 32767 && (b.x0 | b.x1 << 16) && !(b.x1 >> 47 & 1)) + return fp3_convert_NaN(f, b_sgn, b); + + // Detect quiet NaNs: + if (a_exp == 32767 && (a.x0 | a.x1 << 16)) + return fp3_convert_NaN(f, a_sgn, a); + if (b_exp == 32767 && (b.x0 | b.x1 << 16)) + return fp3_convert_NaN(f, b_sgn, b); + + return 0; +} + +static void f3_unpack(int *sgn, int32_t *exp, u128_t *mnt, long double f) +{ + u128_t x; + memcpy(&x, &f, 16); + *sgn = x.x1 >> 63; + *exp = x.x1 >> 48 & 32767; + x.x1 = x.x1 << 16 >> 16; + if (*exp) + x.x1 |= (uint64_t)1 << 48; + else + *exp = 1; + *mnt = x; +} + +static u128_t f3_normalise(int32_t *exp, u128_t mnt) +{ + int sh; + if (!(mnt.x0 | mnt.x1)) + return mnt; + if (!mnt.x1) { + mnt.x1 = mnt.x0; + mnt.x0 = 0; + *exp -= 64; + } + for (sh = 32; sh; sh >>= 1) { + if (!(mnt.x1 >> (64 - sh))) { + mnt.x1 = mnt.x1 << sh | mnt.x0 >> (64 - sh); + mnt.x0 = mnt.x0 << sh; + *exp -= sh; + } + } + return mnt; +} + +static u128_t f3_sticky_shift(int32_t sh, u128_t x) +{ + if (sh >= 128) { + x.x0 = !!(x.x0 | x.x1); + x.x1 = 0; + return x; + } + if (sh >= 64) { + x.x0 = x.x1 | !!x.x0; + x.x1 = 0; + sh -= 64; + } + if (sh > 0) { + x.x0 = x.x0 >> sh | x.x1 << (64 - sh) | !!(x.x0 << (64 - sh)); + x.x1 = x.x1 >> sh; + } + return x; +} + +static long double f3_round(int sgn, int32_t exp, u128_t x) +{ + long double f; + int error; + + if (exp > 0) { + x = f3_sticky_shift(13, x); + } + else { + x = f3_sticky_shift(14 - exp, x); + exp = 0; + } + + error = x.x0 & 3; + x.x0 = x.x0 >> 2 | x.x1 << 62; + x.x1 = x.x1 >> 2; + + if (error == 3 || ((error == 2) & (x.x0 & 1))) { + if (!++x.x0) { + ++x.x1; + if (x.x1 == (uint64_t)1 << 48) + exp = 1; + else if (x.x1 == (uint64_t)1 << 49) { + ++exp; + x.x0 = x.x0 >> 1 | x.x1 << 63; + x.x1 = x.x1 >> 1; + } + } + } + + if (exp >= 32767) + return f3_infinity(sgn); + + x.x1 = x.x1 << 16 >> 16 | (uint64_t)exp << 48 | (uint64_t)sgn << 63; + memcpy(&f, &x, 16); + return f; +} + +static long double f3_add(long double fa, long double fb, int neg) +{ + u128_t a, b, x; + int32_t a_exp, b_exp, x_exp; + int a_sgn, b_sgn, x_sgn; + long double fx; + + f3_unpack(&a_sgn, &a_exp, &a, fa); + f3_unpack(&b_sgn, &b_exp, &b, fb); + + if (fp3_detect_NaNs(&fx, a_sgn, a_exp, a, b_sgn, b_exp, b)) + return fx; + + b_sgn ^= neg; + + // Handle infinities and zeroes: + if (a_exp == 32767 && b_exp == 32767 && a_sgn != b_sgn) + return f3_NaN(); + if (a_exp == 32767) + return f3_infinity(a_sgn); + if (b_exp == 32767) + return f3_infinity(b_sgn); + if (!(a.x0 | a.x1 | b.x0 | b.x1)) + return f3_zero(a_sgn & b_sgn); + + a.x1 = a.x1 << 3 | a.x0 >> 61; + a.x0 = a.x0 << 3; + b.x1 = b.x1 << 3 | b.x0 >> 61; + b.x0 = b.x0 << 3; + + if (a_exp <= b_exp) { + a = f3_sticky_shift(b_exp - a_exp, a); + a_exp = b_exp; + } + else { + b = f3_sticky_shift(a_exp - b_exp, b); + b_exp = a_exp; + } + + x_sgn = a_sgn; + x_exp = a_exp; + if (a_sgn == b_sgn) { + x.x0 = a.x0 + b.x0; + x.x1 = a.x1 + b.x1 + (x.x0 < a.x0); + } + else { + x.x0 = a.x0 - b.x0; + x.x1 = a.x1 - b.x1 - (x.x0 > a.x0); + if (x.x1 >> 63) { + x_sgn ^= 1; + x.x0 = -x.x0; + x.x1 = -x.x1 - !!x.x0; + } + } + + if (!(x.x0 | x.x1)) + return f3_zero(0); + + x = f3_normalise(&x_exp, x); + + return f3_round(x_sgn, x_exp + 12, x); +} + +long double __addtf3(long double a, long double b) +{ + return f3_add(a, b, 0); +} + +long double __subtf3(long double a, long double b) +{ + return f3_add(a, b, 1); +} + +long double __multf3(long double fa, long double fb) +{ + u128_t a, b, x; + int32_t a_exp, b_exp, x_exp; + int a_sgn, b_sgn, x_sgn; + long double fx; + + f3_unpack(&a_sgn, &a_exp, &a, fa); + f3_unpack(&b_sgn, &b_exp, &b, fb); + + if (fp3_detect_NaNs(&fx, a_sgn, a_exp, a, b_sgn, b_exp, b)) + return fx; + + // Handle infinities and zeroes: + if ((a_exp == 32767 && !(b.x0 | b.x1)) || + (b_exp == 32767 && !(a.x0 | a.x1))) + return f3_NaN(); + if (a_exp == 32767 || b_exp == 32767) + return f3_infinity(a_sgn ^ b_sgn); + if (!(a.x0 | a.x1) || !(b.x0 | b.x1)) + return f3_zero(a_sgn ^ b_sgn); + + a = f3_normalise(&a_exp, a); + b = f3_normalise(&b_exp, b); + + x_sgn = a_sgn ^ b_sgn; + x_exp = a_exp + b_exp - 16352; + + { + // Convert to base (1 << 30), discarding bottom 6 bits, which are zero, + // so there are (32, 30, 30, 30) bits in (a3, a2, a1, a0): + uint64_t a0 = a.x0 << 28 >> 34; + uint64_t b0 = b.x0 << 28 >> 34; + uint64_t a1 = a.x0 >> 36 | a.x1 << 62 >> 34; + uint64_t b1 = b.x0 >> 36 | b.x1 << 62 >> 34; + uint64_t a2 = a.x1 << 32 >> 34; + uint64_t b2 = b.x1 << 32 >> 34; + uint64_t a3 = a.x1 >> 32; + uint64_t b3 = b.x1 >> 32; + // Use 16 small multiplications and additions that do not overflow: + uint64_t x0 = a0 * b0; + uint64_t x1 = (x0 >> 30) + a0 * b1 + a1 * b0; + uint64_t x2 = (x1 >> 30) + a0 * b2 + a1 * b1 + a2 * b0; + uint64_t x3 = (x2 >> 30) + a0 * b3 + a1 * b2 + a2 * b1 + a3 * b0; + uint64_t x4 = (x3 >> 30) + a1 * b3 + a2 * b2 + a3 * b1; + uint64_t x5 = (x4 >> 30) + a2 * b3 + a3 * b2; + uint64_t x6 = (x5 >> 30) + a3 * b3; + // We now have (64, 30, 30, ...) bits in (x6, x5, x4, ...). + // Take the top 128 bits, setting bottom bit if any lower bits were set: + uint64_t y0 = (x5 << 34 | x4 << 34 >> 30 | x3 << 34 >> 60 | + !!(x3 << 38 | (x2 | x1 | x0) << 34)); + uint64_t y1 = x6; + // Top bit may be zero. Renormalise: + if (!(y1 >> 63)) { + y1 = y1 << 1 | y0 >> 63; + y0 = y0 << 1; + --x_exp; + } + x.x0 = y0; + x.x1 = y1; + } + + return f3_round(x_sgn, x_exp, x); +} + +long double __divtf3(long double fa, long double fb) +{ + u128_t a, b, x; + int32_t a_exp, b_exp, x_exp; + int a_sgn, b_sgn, x_sgn, i; + long double fx; + + f3_unpack(&a_sgn, &a_exp, &a, fa); + f3_unpack(&b_sgn, &b_exp, &b, fb); + + if (fp3_detect_NaNs(&fx, a_sgn, a_exp, a, b_sgn, b_exp, b)) + return fx; + + // Handle infinities and zeroes: + if ((a_exp == 32767 && b_exp == 32767) || + (!(a.x0 | a.x1) && !(b.x0 | b.x1))) + return f3_NaN(); + if (a_exp == 32767 || !(b.x0 | b.x1)) + return f3_infinity(a_sgn ^ b_sgn); + if (!(a.x0 | a.x1) || b_exp == 32767) + return f3_zero(a_sgn ^ b_sgn); + + a = f3_normalise(&a_exp, a); + b = f3_normalise(&b_exp, b); + + x_sgn = a_sgn ^ b_sgn; + x_exp = a_exp - b_exp + 16395; + + a.x0 = a.x0 >> 1 | a.x1 << 63; + a.x1 = a.x1 >> 1; + b.x0 = b.x0 >> 1 | b.x1 << 63; + b.x1 = b.x1 >> 1; + x.x0 = 0; + x.x1 = 0; + for (i = 0; i < 116; i++) { + x.x1 = x.x1 << 1 | x.x0 >> 63; + x.x0 = x.x0 << 1; + if (a.x1 > b.x1 || (a.x1 == b.x1 && a.x0 >= b.x0)) { + a.x1 = a.x1 - b.x1 - (a.x0 < b.x0); + a.x0 = a.x0 - b.x0; + x.x0 |= 1; + } + a.x1 = a.x1 << 1 | a.x0 >> 63; + a.x0 = a.x0 << 1; + } + x.x0 |= !!(a.x0 | a.x1); + + x = f3_normalise(&x_exp, x); + + return f3_round(x_sgn, x_exp, x); +} + +long double __extendsftf2(float f) +{ + long double fx; + u128_t x; + uint32_t a; + uint64_t aa; + memcpy(&a, &f, 4); + aa = a; + x.x0 = 0; + if (!(a << 1)) + x.x1 = aa << 32; + else if (a << 1 >> 24 == 255) + x.x1 = (0x7fff000000000000 | aa >> 31 << 63 | aa << 41 >> 16 | + (uint64_t)!!(a << 9) << 47); + else + x.x1 = (aa >> 31 << 63 | ((aa >> 23 & 255) + 16256) << 48 | + aa << 41 >> 16); + memcpy(&fx, &x, 16); + return fx; +} + +long double __extenddftf2(double f) +{ + long double fx; + u128_t x; + uint64_t a; + memcpy(&a, &f, 8); + x.x0 = a << 60; + if (!(a << 1)) + x.x1 = a; + else if (a << 1 >> 53 == 2047) + x.x1 = (0x7fff000000000000 | a >> 63 << 63 | a << 12 >> 16 | + (uint64_t)!!(a << 12) << 47); + else + x.x1 = a >> 63 << 63 | ((a >> 52 & 2047) + 15360) << 48 | a << 12 >> 16; + memcpy(&fx, &x, 16); + return fx; +} + +float __trunctfsf2(long double f) +{ + u128_t mnt; + int32_t exp; + int sgn; + uint32_t x; + float fx; + + f3_unpack(&sgn, &exp, &mnt, f); + + if (exp == 32767 && (mnt.x0 | mnt.x1 << 16)) + x = 0x7fc00000 | (uint32_t)sgn << 31 | (mnt.x1 >> 25 & 0x007fffff); + else if (exp > 16510) + x = 0x7f800000 | (uint32_t)sgn << 31; + else if (exp < 16233) + x = (uint32_t)sgn << 31; + else { + exp -= 16257; + x = mnt.x1 >> 23 | !!(mnt.x0 | mnt.x1 << 41); + if (exp < 0) { + x = x >> -exp | !!(x << (32 + exp)); + exp = 0; + } + if ((x & 3) == 3 || (x & 7) == 6) + x += 4; + x = ((x >> 2) + (exp << 23)) | (uint32_t)sgn << 31; + } + memcpy(&fx, &x, 4); + return fx; +} + +double __trunctfdf2(long double f) +{ + u128_t mnt; + int32_t exp; + int sgn; + uint64_t x; + double fx; + + f3_unpack(&sgn, &exp, &mnt, f); + + if (exp == 32767 && (mnt.x0 | mnt.x1 << 16)) + x = (0x7ff8000000000000 | (uint64_t)sgn << 63 | + mnt.x1 << 16 >> 12 | mnt.x0 >> 60); + else if (exp > 17406) + x = 0x7ff0000000000000 | (uint64_t)sgn << 63; + else if (exp < 15308) + x = (uint64_t)sgn << 63; + else { + exp -= 15361; + x = mnt.x1 << 6 | mnt.x0 >> 58 | !!(mnt.x0 << 6); + if (exp < 0) { + x = x >> -exp | !!(x << (64 + exp)); + exp = 0; + } + if ((x & 3) == 3 || (x & 7) == 6) + x += 4; + x = ((x >> 2) + ((uint64_t)exp << 52)) | (uint64_t)sgn << 63; + } + memcpy(&fx, &x, 8); + return fx; +} + +int32_t __fixtfsi(long double fa) +{ + u128_t a; + int32_t a_exp; + int a_sgn; + int32_t x; + f3_unpack(&a_sgn, &a_exp, &a, fa); + if (a_exp < 16369) + return 0; + if (a_exp > 16413) + return a_sgn ? -0x80000000 : 0x7fffffff; + x = a.x1 >> (16431 - a_exp); + return a_sgn ? -x : x; +} + +int64_t __fixtfdi(long double fa) +{ + u128_t a; + int32_t a_exp; + int a_sgn; + int64_t x; + f3_unpack(&a_sgn, &a_exp, &a, fa); + if (a_exp < 16383) + return 0; + if (a_exp > 16445) + return a_sgn ? -0x8000000000000000 : 0x7fffffffffffffff; + x = (a.x1 << 15 | a.x0 >> 49) >> (16446 - a_exp); + return a_sgn ? -x : x; +} + +uint32_t __fixunstfsi(long double fa) +{ + u128_t a; + int32_t a_exp; + int a_sgn; + f3_unpack(&a_sgn, &a_exp, &a, fa); + if (a_sgn || a_exp < 16369) + return 0; + if (a_exp > 16414) + return -1; + return a.x1 >> (16431 - a_exp); +} + +uint64_t __fixunstfdi(long double fa) +{ + u128_t a; + int32_t a_exp; + int a_sgn; + f3_unpack(&a_sgn, &a_exp, &a, fa); + if (a_sgn || a_exp < 16383) + return 0; + if (a_exp > 16446) + return -1; + return (a.x1 << 15 | a.x0 >> 49) >> (16446 - a_exp); +} + +long double __floatsitf(int32_t a) +{ + int sgn = 0; + int exp = 16414; + uint32_t mnt = a; + u128_t x = { 0, 0 }; + long double f; + int i; + if (a) { + if (a < 0) { + sgn = 1; + mnt = -mnt; + } + for (i = 16; i; i >>= 1) + if (!(mnt >> (32 - i))) { + mnt <<= i; + exp -= i; + } + x.x1 = ((uint64_t)sgn << 63 | (uint64_t)exp << 48 | + (uint64_t)(mnt << 1) << 16); + } + memcpy(&f, &x, 16); + return f; +} + +long double __floatditf(int64_t a) +{ + int sgn = 0; + int exp = 16446; + uint64_t mnt = a; + u128_t x = { 0, 0 }; + long double f; + int i; + if (a) { + if (a < 0) { + sgn = 1; + mnt = -mnt; + } + for (i = 32; i; i >>= 1) + if (!(mnt >> (64 - i))) { + mnt <<= i; + exp -= i; + } + x.x0 = mnt << 49; + x.x1 = (uint64_t)sgn << 63 | (uint64_t)exp << 48 | mnt << 1 >> 16; + } + memcpy(&f, &x, 16); + return f; +} + +long double __floatunsitf(uint32_t a) +{ + int exp = 16414; + uint32_t mnt = a; + u128_t x = { 0, 0 }; + long double f; + int i; + if (a) { + for (i = 16; i; i >>= 1) + if (!(mnt >> (32 - i))) { + mnt <<= i; + exp -= i; + } + x.x1 = (uint64_t)exp << 48 | (uint64_t)(mnt << 1) << 16; + } + memcpy(&f, &x, 16); + return f; +} + +long double __floatunditf(uint64_t a) +{ + int exp = 16446; + uint64_t mnt = a; + u128_t x = { 0, 0 }; + long double f; + int i; + if (a) { + for (i = 32; i; i >>= 1) + if (!(mnt >> (64 - i))) { + mnt <<= i; + exp -= i; + } + x.x0 = mnt << 49; + x.x1 = (uint64_t)exp << 48 | mnt << 1 >> 16; + } + memcpy(&f, &x, 16); + return f; +} + +static int f3_cmp(long double fa, long double fb) +{ + u128_t a, b; + memcpy(&a, &fa, 16); + memcpy(&b, &fb, 16); + return (!(a.x0 | a.x1 << 1 | b.x0 | b.x1 << 1) ? 0 : + ((a.x1 << 1 >> 49 == 0x7fff && (a.x0 | a.x1 << 16)) || + (b.x1 << 1 >> 49 == 0x7fff && (b.x0 | b.x1 << 16))) ? 2 : + a.x1 >> 63 != b.x1 >> 63 ? (int)(b.x1 >> 63) - (int)(a.x1 >> 63) : + a.x1 < b.x1 ? (int)(a.x1 >> 63 << 1) - 1 : + a.x1 > b.x1 ? 1 - (int)(a.x1 >> 63 << 1) : + a.x0 < b.x0 ? (int)(a.x1 >> 63 << 1) - 1 : + b.x0 < a.x0 ? 1 - (int)(a.x1 >> 63 << 1) : 0); +} + +int __eqtf2(long double a, long double b) +{ + return !!f3_cmp(a, b); +} + +int __netf2(long double a, long double b) +{ + return !!f3_cmp(a, b); +} + +int __lttf2(long double a, long double b) +{ + return f3_cmp(a, b); +} + +int __letf2(long double a, long double b) +{ + return f3_cmp(a, b); +} + +int __gttf2(long double a, long double b) +{ + return -f3_cmp(b, a); +} + +int __getf2(long double a, long double b) +{ + return -f3_cmp(b, a); +} diff --git a/lib/testfp.c b/lib/testfp.c new file mode 100644 index 0000000..63342b4 --- /dev/null +++ b/lib/testfp.c @@ -0,0 +1,510 @@ +/* + * Test 128-bit floating-point arithmetic on arm64: + * build with two different compilers and compare the output. + * + * Copyright (c) 2015 Edmund Grimley Evans + * + * Copying and distribution of this file, with or without modification, + * are permitted in any medium without royalty provided the copyright + * notice and this notice are preserved. This file is offered as-is, + * without any warranty. + */ + +#include +#include +#include +#include + +#define check(x) ((x) ? (void)0 : check_fail(#x, __FILE__, __LINE__)) + +void check_fail(const char *assertion, const char *file, unsigned int line) +{ + printf("%s:%d: Check (%s) failed.", file, line, assertion); + exit(1); +} + +typedef struct { + unsigned long long x0, x1; +} u128_t; + +float copy_fi(uint32_t x) +{ + float f; + memcpy(&f, &x, 4); + return f; +} + +double copy_di(uint64_t x) +{ + double f; + memcpy(&f, &x, 8); + return f; +} + +long double copy_ldi(u128_t x) +{ + long double f; + memcpy(&f, &x, 16); + return f; +} + +uint32_t copy_if(float f) +{ + uint32_t x; + memcpy(&x, &f, 4); + return x; +} + +uint64_t copy_id(double f) +{ + uint64_t x; + memcpy(&x, &f, 8); + return x; +} + +u128_t copy_ild(long double f) +{ + u128_t x; + memcpy(&x, &f, 16); + return x; +} + +long double make(int sgn, int exp, uint64_t high, uint64_t low) +{ + u128_t x = { low, + (0x0000ffffffffffff & high) | + (0x7fff000000000000 & (uint64_t)exp << 48) | + (0x8000000000000000 & (uint64_t)sgn << 63) }; + return copy_ldi(x); +} + +void cmp(long double a, long double b) +{ + u128_t ax = copy_ild(a); + u128_t bx = copy_ild(b); + int eq = (a == b); + int ne = (a != b); + int lt = (a < b); + int le = (a <= b); + int gt = (a > b); + int ge = (a >= b); + + check(eq == 0 || eq == 1); + check(lt == 0 || lt == 1); + check(gt == 0 || gt == 1); + check(ne == !eq && le == (lt | eq) && ge == (gt | eq)); + check(eq + lt + gt < 2); + + printf("cmp %016llx%016llx %016llx%016llx %d %d %d\n", + ax.x1, ax.x0, bx.x1, bx.x0, lt, eq, gt); +} + +void cmps(void) +{ + int i, j; + + for (i = 0; i < 2; i++) + for (j = 0; j < 2; j++) + cmp(make(i, 0, 0, 0), make(j, 0, 0, 0)); + + for (i = 0; i < 2; i++) { + for (j = 0; j < 64; j++) { + long double f1 = make(i, 32767, (uint64_t)1 << j, 0); + long double f2 = make(i, 32767, 0, (uint64_t)1 << j); + cmp(f1, 0); + cmp(f2, 0); + cmp(0, f1); + cmp(0, f2); + } + } + + for (i = 0; i < 6; i++) + for (j = 0; j < 6; j++) + cmp(make(i & 1, i >> 1, 0, 0), + make(j & 1, j >> 1, 0, 0)); + + for (i = 0; i < 2; i++) { + for (j = 0; j < 2; j++) { + int a, b; + for (a = 0; a < 2; a++) { + for (b = 0; b < 2; b++) { + cmp(make(i, j, a, b), make(i, j, 0, 0)); + cmp(make(i, j, 0, 0), make(i, j, a, b)); + } + } + } + } +} + +void xop(const char *name, long double a, long double b, long double c) +{ + u128_t ax = copy_ild(a); + u128_t bx = copy_ild(b); + u128_t cx = copy_ild(c); + printf("%s %016llx%016llx %016llx%016llx %016llx%016llx\n", + name, ax.x1, ax.x0, bx.x1, bx.x0, cx.x1, cx.x0); +} + +void fadd(long double a, long double b) +{ + xop("add", a, b, a + b); +} + +void fsub(long double a, long double b) +{ + xop("sub", a, b, a - b); +} + +void fmul(long double a, long double b) +{ + xop("mul", a, b, a * b); +} + +void fdiv(long double a, long double b) +{ + xop("div", a, b, a / b); +} + +void nanz(void) +{ + // Check NaNs: + { + long double x[7]; + int i, j, n = 0; + x[n++] = make(0, 32000, 0x95132b76effc, 0xd79035214b4f8d53); + x[n++] = make(1, 32001, 0xbe71d7a51587, 0x30601c6815d6c3ac); + x[n++] = make(0, 32767, 0, 1); + x[n++] = make(0, 32767, (uint64_t)1 << 46, 0); + x[n++] = make(1, 32767, (uint64_t)1 << 47, 0); + x[n++] = make(1, 32767, 0x7596c7099ad5, 0xe25fed2c58f73fc9); + x[n++] = make(0, 32767, 0x835d143360f9, 0x5e315efb35630666); + check(n == sizeof(x) / sizeof(*x)); + for (i = 0; i < n; i++) { + for (j = 0; j < n; j++) { + fadd(x[i], x[j]); + fsub(x[i], x[j]); + fmul(x[i], x[j]); + fdiv(x[i], x[j]); + } + } + } + + // Check infinities and zeroes: + { + long double x[6]; + int i, j, n = 0; + x[n++] = make(1, 32000, 0x62acda85f700, 0x47b6c9f35edc4044); + x[n++] = make(0, 32001, 0x94b7abf55af7, 0x9f425fe354428e19); + x[n++] = make(0, 32767, 0, 0); + x[n++] = make(1, 32767, 0, 0); + x[n++] = make(0, 0, 0, 0); + x[n++] = make(1, 0, 0, 0); + check(n == sizeof(x) / sizeof(*x)); + for (i = 0; i < n; i++) { + for (j = 0; j < n; j++) { + fadd(x[i], x[j]); + fsub(x[i], x[j]); + fmul(x[i], x[j]); + fdiv(x[i], x[j]); + } + } + } +} + +void adds(void) +{ + // Check shifting and add/sub: + { + int i; + for (i = -130; i <= 130; i++) { + int s1 = (uint32_t)i % 3 < 1; + int s2 = (uint32_t)i % 5 < 2; + fadd(make(s1, 16384 , 0x502c065e4f71a65d, 0xd2f9bdb031f4f031), + make(s2, 16384 + i, 0xae267395a9bc1033, 0xb56b5800da1ba448)); + } + } + + // Check normalisation: + { + uint64_t a0 = 0xc6bab0a6afbef5ed; + uint64_t a1 = 0x4f84136c4a2e9b52; + int ee[] = { 0, 1, 10000 }; + int e, i; + for (e = 0; e < sizeof(ee) / sizeof(*ee); e++) { + int exp = ee[e]; + fsub(make(0, exp, a1, a0), make(0, 0, 0, 0)); + for (i = 63; i >= 0; i--) + fsub(make(0, exp, a1 | (uint64_t)1 << i >> 1, a0), + make(0, exp, a1 >> i << i, 0)); + for (i = 63; i >=0; i--) + fsub(make(0, exp, a1, a0 | (uint64_t)1 << i >> 1), + make(0, exp, a1, a0 >> i << i)); + } + } + + // Carry/overflow from rounding: + { + fadd(make(0, 114, -1, -1), make(0, 1, 0, 0)); + fadd(make(0, 32766, -1, -1), make(0, 32653, 0, 0)); + fsub(make(1, 32766, -1, -1), make(0, 32653, 0, 0)); + } +} + +void muls(void) +{ + int i, j; + + { + long double max = make(0, 32766, -1, -1); + long double min = make(0, 0, 0, 1); + fmul(max, max); + fmul(max, min); + fmul(min, min); + } + + for (i = 117; i > 0; i--) + fmul(make(0, 16268, 0x643dcea76edc, 0xe0877a598403627a), + make(i & 1, i, 0, 0)); + + fmul(make(0, 16383, -1, -3), make(0, 16383, 0, 1)); + // Round to next exponent: + fmul(make(0, 16383, -1, -2), make(0, 16383, 0, 1)); + // Round from subnormal to normal: + fmul(make(0, 1, -1, -1), make(0, 16382, 0, 0)); + + for (i = 0; i < 2; i++) + for (j = 0; j < 112; j++) + fmul(make(0, 16383, (uint64_t)1 << i, 0), + make(0, 16383, + j < 64 ? 0 : (uint64_t)1 << (j - 64), + j < 64 ? (uint64_t)1 << j : 0)); +} + +void divs(void) +{ + int i; + + { + long double max = make(0, 32766, -1, -1); + long double min = make(0, 0, 0, 1); + fdiv(max, max); + fdiv(max, min); + fdiv(min, max); + fdiv(min, min); + } + + for (i = 0; i < 64; i++) + fdiv(make(0, 16383, -1, -1), make(0, 16383, -1, -(uint64_t)1 << i)); + for (i = 0; i < 48; i++) + fdiv(make(0, 16383, -1, -1), make(0, 16383, -(uint64_t)1 << i, 0)); +} + +void cvtlsw(int32_t a) +{ + long double f = a; + u128_t x = copy_ild(f); + printf("cvtlsw %08lx %016llx%016llx\n", (long)(uint32_t)a, x.x1, x.x0); +} + +void cvtlsx(int64_t a) +{ + long double f = a; + u128_t x = copy_ild(f); + printf("cvtlsx %016llx %016llx%016llx\n", + (long long)(uint64_t)a, x.x1, x.x0); +} + +void cvtluw(uint32_t a) +{ + long double f = a; + u128_t x = copy_ild(f); + printf("cvtluw %08lx %016llx%016llx\n", (long)a, x.x1, x.x0); +} + +void cvtlux(uint64_t a) +{ + long double f = a; + u128_t x = copy_ild(f); + printf("cvtlux %016llx %016llx%016llx\n", (long long)a, x.x1, x.x0); +} + +void cvtil(long double a) +{ + u128_t x = copy_ild(a); + int32_t b1 = a; + int64_t b2 = a; + uint32_t b3 = a; + uint64_t b4 = a; + printf("cvtswl %016llx%016llx %08lx\n", + x.x1, x.x0, (long)(uint32_t)b1); + printf("cvtsxl %016llx%016llx %016llx\n", + x.x1, x.x0, (long long)(uint64_t)b2); + printf("cvtuwl %016llx%016llx %08lx\n", + x.x1, x.x0, (long)b3); + printf("cvtuxl %016llx%016llx %016llx\n", + x.x1, x.x0, (long long)b4); +} + +void cvtlf(float a) +{ + uint32_t ax = copy_if(a); + long double b = a; + u128_t bx = copy_ild(b); + printf("cvtlf %08lx %016llx%016llx\n", (long)ax, bx.x1, bx.x0); +} + +void cvtld(double a) +{ + uint64_t ax = copy_id(a); + long double b = a; + u128_t bx = copy_ild(b); + printf("cvtld %016llx %016llx%016llx\n", (long long)ax, bx.x1, bx.x0); +} + +void cvtfl(long double a) +{ + u128_t ax = copy_ild(a); + float b = a; + uint32_t bx = copy_if(b); + printf("cvtfl %016llx%016llx %08lx\n", ax.x1, ax.x0, (long)bx); +} + +void cvtdl(long double a) +{ + u128_t ax = copy_ild(a); + double b = a; + uint64_t bx = copy_id(b); + printf("cvtdl %016llx%016llx %016llx\n", ax.x1, ax.x0, (long long)bx); +} + +void cvts(void) +{ + int i, j; + + { + uint32_t x = 0xad040c5b; + cvtlsw(0); + for (i = 0; i < 31; i++) + cvtlsw(x >> (31 - i)); + for (i = 0; i < 31; i++) + cvtlsw(-(x >> (31 - i))); + cvtlsw(0x80000000); + } + { + uint64_t x = 0xb630a248cad9afd2; + cvtlsx(0); + for (i = 0; i < 63; i++) + cvtlsx(x >> (63 - i)); + for (i = 0; i < 63; i++) + cvtlsx(-(x >> (63 - i))); + cvtlsx(0x8000000000000000); + } + { + uint32_t x = 0xad040c5b; + cvtluw(0); + for (i = 0; i < 32; i++) + cvtluw(x >> (31 - i)); + } + { + uint64_t x = 0xb630a248cad9afd2; + cvtlux(0); + for (i = 0; i < 64; i++) + cvtlux(x >> (63 - i)); + } + + for (i = 0; i < 2; i++) { + cvtil(make(i, 32767, 0, 1)); + cvtil(make(i, 32767, (uint64_t)1 << 47, 0)); + cvtil(make(i, 32767, 123, 456)); + cvtil(make(i, 32767, 0, 0)); + cvtil(make(i, 16382, -1, -1)); + cvtil(make(i, 16383, -1, -1)); + cvtil(make(i, 16384, 0x7fffffffffff, -1)); + cvtil(make(i, 16384, 0x800000000000, 0)); + for (j = 0; j < 68; j++) + cvtil(make(i, 16381 + j, 0xd4822c0a10ec, 0x1fe2f8b2669f5c9d)); + } + + cvtlf(copy_fi(0x00000000)); + cvtlf(copy_fi(0x456789ab)); + cvtlf(copy_fi(0x7f800000)); + cvtlf(copy_fi(0x7f923456)); + cvtlf(copy_fi(0x7fdbcdef)); + cvtlf(copy_fi(0x80000000)); + cvtlf(copy_fi(0xabcdef12)); + cvtlf(copy_fi(0xff800000)); + cvtlf(copy_fi(0xff923456)); + cvtlf(copy_fi(0xffdbcdef)); + + cvtld(copy_di(0x0000000000000000)); + cvtld(copy_di(0x456789abcdef0123)); + cvtld(copy_di(0x7ff0000000000000)); + cvtld(copy_di(0x7ff123456789abcd)); + cvtld(copy_di(0x7ffabcdef1234567)); + cvtld(copy_di(0x8000000000000000)); + cvtld(copy_di(0xcdef123456789abc)); + cvtld(copy_di(0xfff0000000000000)); + cvtld(copy_di(0xfff123456789abcd)); + cvtld(copy_di(0xfffabcdef1234567)); + + for (i = 0; i < 2; i++) { \ + cvtfl(make(i, 0, 0, 0)); + cvtfl(make(i, 16232, -1, -1)); + cvtfl(make(i, 16233, 0, 0)); + cvtfl(make(i, 16233, 0, 1)); + cvtfl(make(i, 16383, 0xab0ffd000000, 0)); + cvtfl(make(i, 16383, 0xab0ffd000001, 0)); + cvtfl(make(i, 16383, 0xab0ffeffffff, 0)); + cvtfl(make(i, 16383, 0xab0fff000000, 0)); + cvtfl(make(i, 16383, 0xab0fff000001, 0)); + cvtfl(make(i, 16510, 0xfffffeffffff, -1)); + cvtfl(make(i, 16510, 0xffffff000000, 0)); + cvtfl(make(i, 16511, 0, 0)); + cvtfl(make(i, 32767, 0, 0)); + cvtfl(make(i, 32767, 0, 1)); + cvtfl(make(i, 32767, 0x4cbe01ac5f40, 0x75cee3c6afbb00b5)); + cvtfl(make(i, 32767, 0x800000000000, 1)); + cvtfl(make(i, 32767, 0xa11caaaf6a52, 0x696033e871eab099)); + } + + for (i = 0; i < 2; i++) { + cvtdl(make(i, 0, 0, 0)); + cvtdl(make(i, 15307, -1, -1)); + cvtdl(make(i, 15308, 0, 0)); + cvtdl(make(i, 15308, 0, 1)); + cvtdl(make(i, 16383, 0xabc123abc0ff, 0xe800000000000000)); + cvtdl(make(i, 16383, 0xabc123abc0ff, 0xe800000000000001)); + cvtdl(make(i, 16383, 0xabc123abc0ff, 0xf7ffffffffffffff)); + cvtdl(make(i, 16383, 0xabc123abc0ff, 0xf800000000000000)); + cvtdl(make(i, 16383, 0xabc123abc0ff, 0xf800000000000001)); + cvtdl(make(i, 17406, 0xffffffffffff, 0xf7ffffffffffffff)); + cvtdl(make(i, 17406, 0xffffffffffff, 0xf800000000000000)); + cvtdl(make(i, 17407, 0, 0)); + cvtdl(make(i, 32767, 0, 0)); + cvtdl(make(i, 32767, 0, 1)); + cvtdl(make(i, 32767, 0x4cbe01ac5f40, 0x75cee3c6afbb00b5)); + cvtdl(make(i, 32767, 0x800000000000, 1)); + cvtdl(make(i, 32767, 0xa11caaaf6a52, 0x696033e871eab099)); + } +} + +void tests(void) +{ + cmps(); + nanz(); + adds(); + muls(); + divs(); + cvts(); +} + +int main() +{ +#ifdef __aarch64__ + tests(); +#else + printf("This test program is intended for a little-endian architecture\n" + "with an IEEE-standard 128-bit long double.\n"); +#endif + return 0; +} -- cgit v1.3.1