Changeset c67aff2 in mainline for uspace/lib/softfloat/include/sftypes.h

Timestamp:

2011-08-06T07:04:50Z (13 years ago)

Author:

Petr Koupy <petr.koupy@…>

Branches:

lfn, master, serial, ticket/834-toolchain-update, topic/msim-upgrade, topic/simplify-dev-export

Children:

d3e241a, e0e922d

Parents:

9a6034a

Message:

Quadruple-precision softfloat, coding style improvements. Details below…

Highlights:

• completed double-precision support
• added quadruple-precision support
• added SPARC quadruple-precision wrappers
• added doxygen comments
• corrected and unified coding style

Current state of the softfloat library:

Support for single, double and quadruple precision is currently almost complete (apart from power, square root, complex multiplication and complex division) and provides the same set of features (i.e. the support for all three precisions is now aligned). In order to extend softfloat library consistently, addition of quadruple precision was done in the same spirit as already existing single and double precision written by Josef Cejka in 2006 - that is relaxed standard-compliance for corner cases while mission-critical code sections heavily inspired by the widely used softfloat library written by John R. Hauser (although I personally think it would be more appropriate for HelenOS to use something less optimized, shorter and more readable).

Most of the quadruple-precision code is just an adapted double-precision code to work on 128-bit variables. That means if there is TODO, FIXME or some defect in single or double-precision code, it is most likely also in the quadruple-precision code. Please note that quadruple-precision functions are currently not tested - it is challenging task for itself, especially when the ports that use them are either not finished (mips64) or badly supported by simulators (sparc64). To test whole softfloat library, one would probably have to either write very non-trivial native tester, or use some existing one (e.g. TestFloat from J. R. Hauser) and port it to HelenOS (or rip the softfloat library out of HelenOS and test it on a host system). At the time of writing this, the code dependent on quadruple-precision functions (on mips64 and sparc64) is just a libposix strtold() function (and its callers, most notably scanf backend).

File:

• uspace/lib/softfloat/include/sftypes.h (modified) (4 diffs)

uspace/lib/softfloat/include/sftypes.h

@@ -1 +1 @@
 /*
  * Copyright (c) 2005 Josef Cejka
+ * Copyright (c) 2011 Petr Koupy
  * All rights reserved.
  *
@@ -30 +31 @@
  * @{
  */
-/** @file
+/** @file Floating point types and constants.
  */
 
@@ -77 +78 @@
 } float64;
 
-#define FLOAT32_MAX  0x7f800000
-#define FLOAT32_MIN  0xff800000
-#define FLOAT64_MAX
-#define FLOAT64_MIN
+typedef union {
+        long double ld;
+        struct {
+#if defined(__BE__)
+                uint64_t hi;
+                uint64_t lo;
+#elif defined(__LE__)
+                uint64_t lo;
+                uint64_t hi;
+#else
+        #error Unknown endianess
+#endif
+        } binary;
+
+        struct {
+#if defined(__BE__)
+                uint64_t sign : 1;
+                uint64_t exp : 15;
+                uint64_t frac_hi : 48;
+                uint64_t frac_lo : 64;
+#elif defined(__LE__)
+                uint64_t frac_lo : 64;
+                uint64_t frac_hi : 48;
+                uint64_t exp : 15;
+                uint64_t sign : 1;
+#else
+        #error Unknown endianess
+#endif
+        } parts __attribute__ ((packed));
+} float128;
 
 /*
- * For recognizing NaNs or infinity use isFloat32NaN and is Float32Inf,
+ * For recognizing NaNs or infinity use specialized comparison functions,
  * comparing with these constants is not sufficient.
  */
@@ -95 +122 @@
 #define FLOAT64_INF     0x7FF0000000000000ll
 
-#define FLOAT32_FRACTION_SIZE  23
-#define FLOAT64_FRACTION_SIZE  52
+#define FLOAT128_NAN_HI     0x7FFF800000000000ll
+#define FLOAT128_NAN_LO     0x0000000000000001ll
+#define FLOAT128_SIGNAN_HI  0x7FFF000000000000ll
+#define FLOAT128_SIGNAN_LO  0x0000000000000001ll
+#define FLOAT128_INF_HI     0x7FFF000000000000ll
+#define FLOAT128_INF_LO     0x0000000000000000ll
 
-#define FLOAT32_HIDDEN_BIT_MASK  0x800000
-#define FLOAT64_HIDDEN_BIT_MASK  0x10000000000000ll
+#define FLOAT32_FRACTION_SIZE   23
+#define FLOAT64_FRACTION_SIZE   52
+#define FLOAT128_FRACTION_SIZE 112
+#define FLOAT128_FRAC_HI_SIZE   48
+#define FLOAT128_FRAC_LO_SIZE   64
+
+#define FLOAT32_HIDDEN_BIT_MASK      0x800000
+#define FLOAT64_HIDDEN_BIT_MASK      0x10000000000000ll
+#define FLOAT128_HIDDEN_BIT_MASK_HI  0x1000000000000ll
+#define FLOAT128_HIDDEN_BIT_MASK_LO  0x0000000000000000ll
 
 #define FLOAT32_MAX_EXPONENT  0xFF
 #define FLOAT64_MAX_EXPONENT  0x7FF
+#define FLOAT128_MAX_EXPONENT 0x7FFF
 
 #define FLOAT32_BIAS  0x7F
 #define FLOAT64_BIAS  0x3FF
 #define FLOAT80_BIAS  0x3FFF
+#define FLOAT128_BIAS 0x3FFF
 
 #endif