Mail Archives: djgpp-workers/1997/11/05/12:45:36
I got tired to have to look into half a dozen different references when I
need to set x87 to a certain state. So I added the necessary info to the
libc docs. The patches are below.
Btw, is there any real reason why we shouldn't mask the Invalid Operation
exception on startup? It generates SIGFPE when you compute e.g.
sqrt(-1), which seems to defeat the ANSI behavior of setting errno (when
the functions set it, but at least libm.a does if you ask it nicely).
*** src/libc/pc_hw/fpu/cntrl87.t~0 Wed Nov 5 13:08:04 1997
--- src/libc/pc_hw/fpu/cntrl87.txh Wed Nov 5 18:16:08 1997
***************
*** 10,17 ****
@subheading Description
! Sets or resets bits in the FPU's control register. The bits are
! defined by macros in float.h, and by this table:
@example
---- ---- --XX XXXX = MCW_EM - exception masks (1=handle exception internally, 0=fault)
--- 10,22 ----
@subheading Description
! This function sets and retrieves the FPU's @dfn{control word}.
!
! The control word is a special 16-bit register maintained by the math
! coprocessor. By setting and clearing bit fields in the control word,
! you can exercise control of certain aspects of coprocessor operation.
! The individual bits of the x87 control word are defined by macros in
! float.h, and shown in this table:
@example
---- ---- --XX XXXX = MCW_EM - exception masks (1=handle exception internally, 0=fault)
***************
*** 35,40 ****
--- 40,105 ----
---1 ---- ---- ---- = IC_PROJECTIVE - -Inf == +Inf
@end example
+ @code{_control87} uses the value of @var{newcw} and @var{mask} variables
+ together to determine which bits of the FPU's control word should be
+ set, and to what values. For each bit in @var{mask} that is set (equals
+ to 1), the corresponding bit in @var{newcw} specifies the new value of
+ the same bit in the FPU's control word, which @code{_control87} should
+ set. Bits which correspond to reset (zero) bits in @var{mask} are not
+ changed in the FPU's control word. Thus, using a zero value for
+ @var{mask} retrieves the current value of the control word without
+ changing it.
+
+ The exception bits @code{MCW_EM} (the low-order 6 bits) of the control
+ word define the exception @emph{mask}. That is, if a certain bit is
+ set, the corresponding exception will be masked, i.e., it will not
+ generate an FP exception (which normally causes signal @code{SIGFPE} to
+ be delivered). A masked exception will be handled internally by the
+ coprocessor. In general, that means that it will generate special
+ results, such as @dfn{NaN}, Not-a-Number (e.g., when you attempt to
+ compute a square root of a negative number), denormalized result (in
+ case of underflow), or infinity (e.g., in the case of division by zero,
+ or when the result overflows).
+
+ By default, DJGPP startup code masks all exceptions except Invalid
+ Operation. This setting will cause your program to crash with
+ @code{SIGFPE} if it tries to compute e.g. @code{sqrt (-1)}. If you want
+ your program to set @code{errno} and continue, you will need to make
+ sure the @code{EM_INVALID} bit in the control word is set (you will also
+ need to link with the alternate math library, @samp{-lm}).
+
+ The precision-control field @code{MCW_PC} (bits 8 and 9) controls the
+ internal precision of the coprocessor by selecting the number of
+ precision bits in the mantissa of the FP numbers. The values
+ @code{PC_24}, @code{PC_53}, and @code{PC_64} set the precision to 24,
+ 53, and 64-bit mantissa, respectively. This feature of the coprocessor
+ is for compatibility with the @dfn{IEEE 745 standard} and only affect
+ the @code{FADD}, @code{FSUB} @code{FSUBR}, @code{FMUL}, @code{FDIV},
+ @code{FDIVR}, and @code{FSQRT} instructions. Lowering the precision
+ will @strong{not} decrease the execution time of FP instructions.
+
+ The @code{MCW_PC} field is set to use the full-precision 64-bit
+ mantissa by the DJGPP startup code.
+
+ The rounding-control field @code{MCW_RC} (bits 10 and 11) controls the
+ type (round or chop) and direction (-Inf or +Inf) of the rounding. It
+ only affects arithmetic instructions. Set to round-to-nearest state by
+ the DJGPP startup code.
+
+ The infinity-control bit @code{MCW_IC} has no effect on 80387 and later
+ coprocessors.
+
@subheading Return Value
The previous control word.
+
+ (Note that this is different from what @code{_control87} from the
+ Borland C library which returns the @emph{new} control word.)
+
+ @subheading Example
+
+ @example
+ /* mask all exceptions, except invalid operation */
+ _control87 (0x033e, 0xffff);
+ @end example
+
- Raw text -