Enhance soft-fp for TFmode<->[DI,SI,TI,XF]mode conversions

[Uros Bizjak]

Hello!

This patch enhances soft-fp for TFmode<->[DI,SI,TI,XF]mode conversions. 
It was tested for x86_64 target with gcc testsuite.

Regarding attached patch:

In addition to TFmode conversions to [DI,SI,TI]mode, attached patch
fixes a couple of inconsistencies in XFmode handling:

a) It is wrong to increase exponent for XFmode denormal operands during
RAW_UNPACK operation, as this causes other code to wrongly detect
denormals as normal operands. Implicit" bit (which  is in fact explicit 
for XFmode) is correctly handled through the code due to _FP_FRACBITS_E 
definition. Also, there is no need to generate DENORMAL exception for 
XFmode operands during unpack.

b) Some compile time errors were fixed on the way.

Regarding the truncation:

The TFmode-XFmode truncation is performed in semi-raw mode, where we 
keep 3 LSB bits  in addition to FP_FRACBITS_E. In case where EXPBIAS of 
target and source operands are the same, the mantissa is shifted right 
with sticky LSB. When the result is packed, UNDERFLOW and INEXACT 
exceptions are generated by generic code. DENORM exception is always set 
as we have denormal input operand (this is also the case without my patch).

2007-04-14  Uros Bizjak  <ubizjak@gmail.com>

         * soft-fp/floattisf.c: New file.
         * soft-fp/floattidf.c: New file.
         * soft-fp/floattitf.c: New file.
         * soft-fp/floatuntisf.c: New file.
         * soft-fp/floatuntidf.c: New file.
         * soft-fp/floatuntitf.c: New file.
         * soft-fp/fixsfti.c: New file.
         * soft-fp/fixdfti.c: New file.
         * soft-fp/fixtfti.c: New file.
         * soft-fp/fixunssfti.c: New file.
         * soft-fp/fixunsdfti.c: New file.
         * soft-fp/fixunstfti.c: New file.
         * soft-fp/extendxftf.c: New file.
         * soft-fp/trunctfxf.c: New file.

         * soft-fp/floatunsidf.c (__floatunsidf): Use DFtype instead of
         double in the function declaration.
         * soft-fp/floatundidf.c (__floatundidf): Use DFtype instead of
         double in the function declaration.
         * soft-fp/floatunsisf.c (__floatunsisf): Use SFtype instead of
         float in the function declaration.
         * soft-fp/floatunsisf.c (__floatunsisf): Use SFtype instead of
         float in the function declaration.

         * soft-fp/extended.h (FP_UNPACK_RAW_E): Do not increase X##_e for
         denormal operands.  Do not generate FP_EX_DENORM exception.
         (FP_UNPACK_RAW_EP): Ditto.
         (FP_UNPACK_SEMIRAW_E): Use FP_UNPACK_RAW_E instead of
         undefined _FP_UNPACK_RAW_E.
         (FP_UNPACK_SEMIRAW_EP): Use FP_UNPACK_RAW_EP instead of
         undefined _FP_UNPACK_RAW_EP.
         (FP_PACK_SEMIRAW_E): Use FP_PACK_RAW_E instead of
         undefined _FP_PACK_RAW_E.
         (FP_PACK_SEMIRAW_EP): Use FP_PACK_RAW_EP instead of
         undefined _FP_PACK_RAW_EP.

         * op-2.h (_FP_FRAC_COPY_2_2): Define as alias to _FP_FRAC_COPY_2.
         * op-4.h (_FP_FRAC_COPY_2_2): Define as alias to _FP_FRAC_COPY_4.

2007-04-16  Uros Bizjak  <ubizjak@gmail.com>
             Jakub Jelinek  <jakub@redhat.com>

         * soft-fp/op-common.h (FP_EXTEND): Do not abort when
         _FP_EXPBIAS_##dfs == _FP_EXPBIAS_##sfs.  Handle denormals for
         this case.
         * soft-fp/op-common.h (FP_EXTEND): Ditto.

Uros.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: soft-fp.diff
Type: text/x-patch
Size: 35290 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20070422/619c42e1/attachment.bin>