GitLab

this is part of the initiative to move object-orientation support stuff
out of libfirm into liboo

We missed the case where transform_node created a node that
computed_value or equivalent_node would catch

It was a strange feature only inspired by the lazyness of some frontend
developer (=me)

they were in lower_mode_b but just generically apply to all Mux(c, 0, 1)

Add op(op(x, c0), c1) => op(x, op(c0, c1)) as localopt rule. This is an
very usual and often encountered case so it makes sense to have this
working without explicitely invoking the associativity optimisation phase.

when bits are flipped for a bittest we can just as well negate the pnc
and don't flip the bits

This is the new way of handling switch-jumps. The node contains a table
which maps (ranges of) input values to proj numbers. Compared to a
Cond-node this results in a clean consecutive sequence of Proj numbers
(no searching for a free number for the default_pn anymore) and allows
factoring multiple cases jumping to the same block in a single Proj
(though we still need the optimisation in cfopt for that).

The main change here is splitting new_ir_mode into new_int_mode,
new_reference_mode and new_float_mode. You can now specify
mantissa+exponent size in new_float_mode. This also changes:

- x86 80bit-FP mode is NOT a ieee754 don't put "ieee754" into functions
  names that can also handle x86 80bit fps
- Move ieee_descriptor_t from tarval module into ir_mode struct
  (and rename to float_descriptor_t)
- Introduce mode_Q which represents binary128 from ieee754
- You can ask float modes for mantissa/exponent sizes now
- Fix endianess when emitting big float values in begnuas
- A bunch of long double fixes in ia32: the mode there has 10bytes
  (80bit) but the variables typically are 12 or 16 byte big
- This fixes some problems of sparc binary128 handling

If vrp can prove that both operand of an Or/Eor/Add never a bit set in
common then we can match them as each of these opcodes.

They're adding more noise than they provide gain in readability. They
also make extra work to keep up to date.

The new interface uses no complicated callbacks anymore, but is a simple
ir_node* -> void* nodemap. Adapted all users of ir_phase for this.

The transformation (a << c1) | c2 => (a | (c2 >> c1)) << c1
is only valid if the lower bits of c2 are clean.

This fixes opt/pr37102.c.

backends can use this to allow all mux nodes which get optimized away by
the middleend anyway.
This fixes abs handling on sparc.

- Make API private as it should only be called by backends
- Let ia32 backend create special ia32_Set nodes instead of relying
  on muxes which must not be touched anymore
- Does not create ConvB nodes anymore but instead produces the Cmp
  directly. (All backends did this anyway during code-selection so I was
  able to leave this case out code-selection)
- First collect nodes to lower, then lower them. This avoids robustness
  problems when transforming the graph while at the same time walking
  it.

We have to convert the values the same way the program does.

This fixes opt/fehler194.c.