1 $NetBSD: README.NetBSD,v 1.2 2002/03/24 23:04:03 bjh21 Exp $
4 Changes at initial import...ross@NetBSD.org
7 o source reorganized with arch/
8 o <sys/endian.h> conversion
9 o <inttypes.h> conversion
10 o <ieeefp.h> conversion
12 o the previously target-specific softfloat.h and milieu.h made mostly
13 generic. Still some work to do with default NaN bitpatterns, endian,
14 and arm/fpa-DEMANGLE issues.
15 o arch/i386/systfloat.S extended to handle:
16 int32_t syst_floatx80_to_int32_round_to_zero(floatx80)
17 int64_t syst_floatx80_to_int64_round_to_zero(floatx80)
18 o LONG_DOUBLE_IS_FLOATX80 was used to enable both the C module system
19 access ops and the FLOATX80 tests, all of which also required FLOATX80.
20 Besides being rundundant, this made it impossible to use the asm module
21 for i386 which the package actually comes with and also made it
22 impossible to test FLOATX80 ops without compiler support even if the
23 machine actually does them. While this is arguably OK for a regression
24 test, the two cases are obviously different. Now, the tests (which
25 don't actually require any compiler-understood extended type) are
26 always run just by defining FLOATX80. If LONG_DOUBLE_IS_FLOATX80 is
27 also defined, then the C system ops are also enabled. To switch back
28 and forth, you modify only the arch/${MACHINE_ARCH}/Makefile.inc to
29 do the cpp def and add or remove systfloat.S. For regression testing,
30 it's better in C, but for testing the test itself or verifying a
31 regression, the .S module is a better reference case.
33 A similar change should probably be made for LONG_DOUBLE_IS_FLOAT128,
34 but we can't test that yet.
37 LONG_DOUBLE_IS_FLOAT{X80,128}
38 define if you do not have a .S file and do have compiler
39 support, or have both and want to use the compiled version
40 (but then it may be necessary to remove the .S file or add
43 define if you have either HW or SW support and want it tested
45 And to confuse things further: the .S files for i386 appear to have been
46 originally created by compiling C code with -S, at least as a start.
48 You can see some of the price of ANSI+IEEE in these files: a simple
49 conversion from floatxx to intxx requires storing, modifying, loading,
50 and restoring the modes on each conversion in order to get a specific
51 round, thanks to the four IEEE-mandated options. A pipeline/peephole
52 pass can no doubt fix this up within one function, but not if other
53 arithmetic is done in between.