The encoding field specifies the register number used when encoding an
instruction as machine code (and not as assembly)

ia32-fpunit     -> ia32-fpmath
ia32-opt        -> ia32-tune
ia32-soft-float -> new
Also fixes a too early invocation of set_arch_costs

Usually '-' was used as delimiter for hierarchically nested option
groups. We now try the completely remaining string as option if no
subgroup can be found, so options like "soft-float" work now (as long as
there is not subgroup named "soft")

This especially removes unused complicated error reporting mechanisms

In the previous buggy form it was creating an overly conservative
approximation.

(C1 - x) + C2 = (C1 + C2) - x
C1 - (x + C2) = (C1 - C2) - x
~(x + C) = (~C) - x (extended from ~(x + -1) = -x)
~(C - x) = x + (~C)
-(x + C) = (-C) - x

Restructure storing the nodes so that memory unchaining a pure call does not access the memory proj of another pure call, which was deleted before.

Ingmar Wirths authored Jan 21, 2013 and yb9976 committed Apr 04, 2013

This implements the "round down" algorithm by ridiculous_fish (see
Faster Unsigned Division by Constants,
http://ridiculousfish.com/blog/posts/labor-of-division-episode-iii.html),
which improves the generated code for some constants (like 42).

The builtin can be used to generate fast code for unsigned division by constant.

Code generation is supported for the IA32 and the SPARC backend.
Since our ARM backend currently has no Add with Carry instruction,
the builtin is currently not supported on Arm.
The same holds for the AMD64 backend, which does not support a division yet.

This fixes 176.gcc.

The code requires assigned registers but no kept Proj.

This fixes backend/udiv.c on SPARC.

With single quotes all except the first line of the emit string are
not indented, which makes the assembler illegible.

This fixes opt/fehler211.c and opt/fehler212.c.

For newly created nodes without any user the function now correctly
returns false. Thus, these nodes are no longer ignored when performing
(global?) common subexpression elimination.

Since we have two fp-vrp.o files, the -u argument of ar sometimes (on
rebuilds) missed one. As -u is not necessary anyways, we now include both.

This is safer for resources the user requires itself, since the
information must live longer than the analysis then.

Now we have a simple optimization actually using DCA.

warning: no previous prototype for 'dca_analyze' [-Wmissing-prototypes]

Factor out the following function:

/* Compute VRP fixpoint.  The Result is available via links to
 * vrp_bitinfo*, allocated on client_obst. */
void fp_vrp_analyze(ir_graph* const irg, struct obstack *client_obst);