GitLab

There is only an fstp variant for storing 80bit float values. Introduce
a special fstp node so we can already create the the fixup copies in the
spillprepare phase.

There is only the fistp variant for 64bit integer stores. Introduce a
special fistp node so we can already create the fixup copies in the
spillprepare phase and avoid the fixup code in the x87 simulator which
was also wrong for the fist case.

The kills value constraint indicated that the value in a register is
modified after it is read. A copy is inserted if the value is still
required after a kills_value use.

It is like new_r_Proj(), but the mode is determined automatically from the out requirement of the predecessor.
This probably resolves some mode mismatches in untested code paths -- AMD64 backend and SPARC hard float, I'm looking at you!

- Rename other_same field to should_be_same.
- Remove arch_register_req_type_other_same, it's enough to see if the
  should_be_same field is not zero.

In non-debug builds the simulator now only runs when it was requested, instead of running with an uninitialized state.

Do this by always running the x87 simulator in debug builds and asserting on push to the simulated x87 stack that the request flag is set.

This way get_irn_Proj_for_mode() can be removed.
Use get_Proj_for_pn() instead.

This properly models the data dependencies and avoids keep edges.

Determine the pos from the node to push.

It only worked, if a single value, which was at TOS, should be popped.

The variable was initialised to this just a few lines above.

All other structs holding a register class call it cls, too.

ia32: Directly put the bitfields in ia32_attr_t instead of having another struct needlessly wrapping them.

Do not insert instructions, which need to be simulated afterwards.
The only case, which did this, (80bit float/64bit int store with full stack) handles the inserted node locally now.