aboutsummaryrefslogtreecommitdiffhomepage
Commit message (Collapse)AuthorAgeFilesLines
* hm lemon2023-08-071-4/+4
|
* regalloc fixes and rpo lemon2023-07-094-21/+78
|
* amd64/emit bugfix lemon2023-07-091-3/+3
|
* amd64 codegen fixes lemon2023-07-072-13/+20
|
* mem2reg; style lemon2023-07-071-7/+12
|
* regalloc: update preds during simplify pass lemon2023-07-071-1/+7
|
* fix copyopt lemon2023-07-061-1/+1
|
* fix emit() setcc and copy lemon2023-07-061-1/+2
|
* fix mem2reg ?? lemon2023-07-061-10/+43
|
* some frontend fixes lemon2023-06-301-3/+24
|
* misc bugfixs lemon2023-06-304-29/+98
|
* add initializers (only static for initialier list rn) lemon2023-06-2916-255/+948
| | | | and other fixes
* also C23 static_assert lemon2023-06-261-3/+3
|
* c: move codegen code after decl parser lemon2023-06-261-748/+748
|
* frontend: add static assert lemon2023-06-262-2/+38
|
* regalloc: remove unused variable lemon2023-06-261-5/+0
|
* backend: fix mem2reg & regalloc lemon2023-06-267-61/+206
| | | | | | | | | | | | | | | | | | | | | | | | | | | | they were broken, especially for unstructured control flow. most significant fix is to register allocator for temporaries that are used before the first definition in the source order, e.g.: @1: %x = add %y, 1 b @3 @2 %y = ... b @1 it's legal for %x to use %y there (assuming @2 dominates @1) but from the point of view of the register allocator %y is defined and freed and then used again, which broke things. the fix is to introduce phis for this situation: @1: %y.1 = phi @2 %y %x = add %y.1, 1 b @3 @2 %y = ... b @1 then regalloc phi handling code makes it work
* frontend: add labels and goto statement lemon2023-06-262-8/+120
|
* fix some warnings lemon2023-06-253-11/+2
|
* lex: use a circular buffer cache for pre-pre-processed chars lemon2023-06-252-56/+67
|
* regalloc: fix temporary rename clobbering return register with multiple returns lemon2023-06-251-2/+13
|
* regalloc add sources lemon2023-06-251-1/+3
|
* abi0: fix agg return by pointer not updating arg count lemon2023-06-241-0/+1
|
* fix regalloc thinking some ins are dead by adding ins.keep lemon2023-06-243-3/+8
|
* update todo.txt lemon2023-06-241-1/+1
|
* ir: dattab shouldn't reset for every function lemon2023-06-241-1/+1
| | | | this was causing later functions' data to override earlier data
* amd64/emit: more float fixes lemon2023-06-241-21/+24
| | | | | optimize loading 0.0 in phis as well as regular copies and also don't use inc or addr for float addition
* backend: don't mixup float and int temps lemon2023-06-244-11/+29
| | | | | | copy propagation only happens when dataclasses match, register allocator ignores hints if hint register class and instruction class differ, also add mov between int and float regs in amd64/emit
* fix replcuses lemon2023-06-243-19/+19
|
* backend: fix regalloc to work with more complex dataflow lemon2023-06-2412-225/+748
| | | | | | | | basically an allocation map at the beginning (in) and end (out) of each block is kept and after the first allocation pass another pass is ran to resolve allocation conflicts between each edge, plus another pass to finish lowering phi functions. also introduced `regset` and plenty of other miscellaneous fixes
* change RMORE -> RADDR; use RXXX (RNONE) for special args,also undef lemon2023-06-229-74/+73
|
* Add .editorconfig lemon2023-06-221-0/+14
|
* misc fix lemon2023-06-223-18/+55
|
* mem2reg: implement ssa construction; this breaks regalloc right now lemon2023-06-219-159/+300
|
* explicitly store predecessors in each block lemon2023-06-218-115/+135
|
* amd64: fix aggregate return in regs lemon2023-06-201-3/+3
|
* amd64: conform to ABI for varargs func calls lemon2023-06-202-1/+9
|
* elf: fix emitting small dats lemon2023-06-201-3/+3
|
* fix type of undeclared functions lemon2023-06-201-0/+1
|
* improve codegen for div a little (bikeshedding) lemon2023-06-202-1/+7
|
* amd64/emit: ensure stack is 16-byte aligned lemon2023-06-203-10/+34
|
* add basic mem2reg lemon2023-06-208-14/+221
| | | | | | | promotes uniform stack slots to temporaries currently only for immutable variables, next thing to implement is ssa construction
* fix regression wiht ret args not being fixed lemon2023-06-201-2/+7
|
* fix out of bounds read lemon2023-06-201-0/+1
|
* regalloc: fix bug in spilling code lemon2023-06-201-7/+24
|
* another emit() errata lemon2023-06-201-1/+1
|
* fix regression lemon2023-06-201-4/+4
|
* don't lower params and args to registers in abi0 lemon2023-06-206-57/+101
| | | | | it's better to do it later, currently in isel(), but perhaps in a pre-isel abi1 pass
* fix cls logic for comparison instrs lemon2023-06-206-17/+23
| | | | | | | | previously instr.cls always represented the output dataclass. this doesn't work for comparisons because we know the output is always a boolean integer and we care about the actual comparison dataclass. so now .cls represents the operation dataclass, which matches the result class except for comparisons where the result is always KI4V
* errata lemon2023-06-201-2/+2
|