code for decomposing float/double/long double, proper FX2DEC NaN support.

toolbox/CMakeLists.txt

@@ -28,6 +28,7 @@ set(TOOLBOX_SRC
 	fs_spec.cpp
 	realpath.c
 	dispatch.cpp
+	fpinfo.cpp
 )
 
 

toolbox/fpinfo.cpp

@@ -0,0 +1,106 @@
+#include "fpinfo.h"
+
+#include <machine/endian.h>
+
+
+static_assert(sizeof(float) == 4, "Unexpected float size");
+static_assert(sizeof(double) == 8, "Unexpected double size");
+static_assert(sizeof(long double) == 8 || sizeof(long double) == 16, "Unexpected long double size");
+
+void fpinfo::init(float *f)
+{
+	union split {
+		float f;
+		uint32_t i;
+	};
+
+	uint32_t i = ((split *)f)->i;
+
+	sign = i >> 31;
+	one = 1;
+	exp = (i >> 23) & ((1 << 8) - 1);
+	sig = i & ((1 << 24) - 1);
+
+	if (exp == 255) {
+		if (sig == 0) inf = true;
+		else nan = true;
+		return;
+	}
+
+	if (exp == 0) {
+		// 0 or denormalized.
+		one = 0;
+		exp = -126;
+		return;
+	}
+
+	exp -= 127; // bias
+}
+
+void fpinfo::init(double *d)
+{
+
+	union split {
+		double d;
+		uint64_t i;
+	};
+
+	uint64_t i = ((split *)d)->i;
+
+	sign = i >> 63;
+	one = 1;
+	exp = (i >> 52) & ((1 << 11) - 1);
+	sig = i & ((UINT64_C(1) << 53) - 1);
+
+
+	if (exp == 2047) {
+		if (sig == 0) inf = true;
+		else nan = true;
+		return;
+	}
+	if (exp == 0) {
+		// 0 or denormalized.
+		one = 0;
+		exp = -1022;
+		return;
+	}
+
+	exp -= 1023; // bias
+}
+
+void fpinfo::init(long double *ld)
+{
+
+	union split {
+		long double ld;
+		uint64_t i[2];
+	};
+
+	uint64_t i;
+	uint32_t sexp;
+
+
+	// this needs to be verified.
+	#if BYTE_ORDER == LITTLE_ENDIAN
+	i = ((split *)ld)->i[0];
+	sexp = ((split *)ld)->i[1];
+	#else
+	i = ((split *)ld)->i[1];
+	sexp = ((split *)ld)->i[0];
+	#endif
+
+	sign = sexp >> 15;
+	sig = sexp & ((1 << 15) - 1);
+
+	one = i >> 63;
+	sig = i & ((UINT64_C(1) << 63) - 1);
+
+	if (exp == 32767) {
+		if (sig == 0) inf = true;
+		else nan = true;
+		return;
+	}
+	// 
+
+	exp -= 16383;
+}

toolbox/fpinfo.h

@@ -0,0 +1,29 @@
+#ifndef __fpinfo_h__
+#define __fpinfo_h__
+
+#include <cstdint>
+
+struct fpinfo {
+	bool sign = false;
+	bool one = false;
+	int exp = 0;
+	uint64_t sig = 0;
+
+	bool nan = false;
+	bool inf = false;
+
+	fpinfo(float f) { init(&f); }
+	fpinfo(double d) { init(&d); }
+	fpinfo(long double ld) {
+		if (sizeof(long double) == 16) init(&ld);
+		if (sizeof(long double) == 8) init((double *)&ld);
+	}
+
+private:
+	void init(float *);
+	void init(double *);
+	void init(long double *);
+};
+
+
+#endif

toolbox/sane.cpp

@@ -39,6 +39,7 @@
 
 #include "stackframe.h"
 #include "complex.h"
+#include "fpinfo.h"
 
 using ToolBox::Log;
 
@@ -104,6 +105,8 @@ using its_complicated::signbit;
 
 
 
+
+
 	}
 
 	// long double is an 80-bit extended with an extra 48-bits of 0 padding.
@@ -354,6 +357,9 @@ using its_complicated::signbit;
 
 		if (df.digits < 0) df.digits = 0;
 
+		fpinfo fpi(s);
+		//fprintf(stderr, "%02x %02x %08x %016llx\n", fpi.sign, fpi.one, fpi.exp, fpi.sig);
+
 		// handle infinity, nan as a special case.
 		switch (fpclassify(s))
 		{
@@ -364,11 +370,14 @@ using its_complicated::signbit;
 				return 0;
 
 			case FP_NAN:
-				// NAN type should be encoded in the sig.
-				// 3.2 CSANELib.o nan() function is broken.
-				d.sgn = signbit(s);
-				d.sig = "N40xx000000000000";
-				d.write(d_adr);
+				// NAN type encoded in the sig.
+				{
+					char buffer[20]; // 16 + 2 needed
+					snprintf(buffer, 20, "N%016llx", fpi.sig);
+					d.sgn = signbit(s);
+					d.sig = buffer;
+					d.write(d_adr);
+				}
 				return 0;				
 
 			case FP_INFINITE:
@@ -383,6 +392,7 @@ using its_complicated::signbit;
 
 		}
 
+		// normal and subnormal handled here....
 
 		#if 0
 		if (df.style == decform::FIXEDDECIMAL)