GitLab

[r27014]

[r27012]

[r27011]

[r27008]

In mode_b lowering move the assertion for mode_b below the check, whether the node was replaced already. Some nodes are their own replacement and therefore do not have mode_b anymore after lowering.

[r27007]

Consistently record the replacement nodes and removal candidates when doing mode_b lowering. This was not done in some cases, most importantly the former was not done for certain Projs, which are their own replacement (just their mode gets changed). Also add a FIXME that the algorithm is subtly broken and may replace the same node twice due to recursion over Phis.

[r27006]

[r27005]

[r27004]

[r27003]

[r27002]

[r27001]

[r27000]

[r26997]

[r26996]

[r26995]

[r26994]

[r26991]

[r26990]

[r26988]

  interface now.
- Fix several bugs in the ia32 version of it
- Unfortunately this leads to the obscure Doz construct being always created
  (instead of only when cmovs are allowed) and firm generates broken code
  for that (testcase coming in 5 minutes)

[r26985]

[r26983]

CSE neutral to its users, i.e. op(..., x, ...) ~ op(..., y, ...) if x and y are both neutral.
Setting this property to ia32_ProduceVal nodes allow to CSE nodes that differ in ProduceVals only (which themself NEVER are cse'd).
This "fixes" divmod.

[r26981]

Add a rule to transform x .op. (x .op. y) into y .op. (x .op. x) for commutative nodes. This finally "fixes" codegen/lea.c and might be usefull for and,or,xor as well.

[r26980]

[r26979]

Partly undo r26950: there seems be be subtile difference between get_flags_node() and determine_final_pnc() that are triggered by that change.
This should fix 175.vpr.

[r26977]

[r26974]

[r26973]

[r26972]

[r26970]

[r26969]

[r26968]

[r26967]

[r26966]

[r26965]

[r26964]

As a reminiscence to the famous MAC/65 assembler changed modifier + into > and - into <, so >R is the hight part of a register, <R the low part!

[r26963]

[r26962]

[r26961]

[r26960]

[r26959]