aboutsummaryrefslogtreecommitdiffhomepage
path: root/src
Commit message (Collapse)AuthorAgeFilesLines
* support extern inline "definitions" lemon2026-04-173-8/+12
|
* cpp: #include_next lemon2026-04-172-29/+46
|
* c: fix using array as condition value lemon2026-04-171-2/+2
| | | | Decays to pointer
* x86-64 backend: fix some edgecases not emitting GOT relocations lemon2026-04-151-5/+13
| | | | | Logic for this was duplicated. Updated in c4d9087746 but only in one of the places, so broke on some PIC code.
* 0.2.6 lemon2026-04-141-2/+2
|
* revert cbf1018b21e and actual fix lemon2026-04-141-3/+10
| | | | | Actually that check was OK. The issue was in isnullpo(), not detecting the folded `(void*)0`
* c: allow nested aggregate with flexible array member lemon2026-04-141-2/+4
|
* c: fix type of `<unsigned> /= <float>` lemon2026-04-141-2/+5
| | | | | | Would wrongly generate `udiv` and fail later in the backend. Should add a IR sanity checks pass to catch e.g. unsigned operations being used with floating point classes.
* backend: allow zero-size stack allocations lemon2026-04-141-1/+1
| | | | Hm.
* c: make type of conditional expr with pointers match non void pointer lemon2026-04-141-1/+1
| | | | | | | | | | | | | | | | Here's what the C standard[1] says about the conditional operator: ``` if one operand is a null pointer constant, the result has the type of the other operand; otherwise, one operand is a pointer to void or a qualified version of void, in which case the result type is a pointer to an appropriately qualified version of void. `` But actually, some codebases rely on things like 'x ? &foo.memb : NULL' to have the type of `&foo.memb`, and not `void *`... [1]: ISO/IEC 9899:TC3 §6.5.15
* cpp: concat 123 ## . lemon2026-04-142-4/+19
|
* c: implement explicit cast complex -> scalar lemon2026-04-131-3/+7
|
* c: match class, not typetag, for libgcc stubs lemon2026-04-131-1/+1
|
* C99 complex types MVP lemon2026-04-1311-68/+436
| | | | | | | | | Missing: static eval of complex values, Silly inefficient implementation of equality comparisons between them The whole thing is pretty inefficient without proper aggregate mem2reg anyway
* ir: remove unused field MCTarg.bpr lemon2026-04-133-4/+1
|
* frontend: accept zero size arrays and empty aggregates lemon2026-04-133-31/+39
| | | | | | GNU extension. Previously was overloading arraylength==0 to mean unsized array, now they are distinct with the type flag TFUNKNOWN marking unsized arrays & aggregates.
* get rid of other warnings lemon2026-04-122-5/+4
|
* ir_simpl: remove unused variable lemon2026-04-121-3/+2
|
* aarch64: align leaf func stacks too lemon2026-04-121-2/+2
|
* aarch64: fix stack frame layout, again lemon2026-04-122-35/+37
| | | | | Turns out hardware enforces SP 16-bit alignment in every SP relative load/store, but QEMU doesn't emulate this.
* frontend: always narrow assignment expression used as value lemon2026-04-111-2/+2
|
* 0.2.5 lemon2026-04-111-2/+2
|
* irdump: dont print instr numbers by default lemon2026-04-111-6/+21
|
* backend: fix memory leak from not calling deluses() after replcuses in some ↵ lemon2026-04-114-12/+11
| | | | places
* backend: run mem2reg after inlining lemon2026-04-112-1/+7
|
* x86-64/emit: fix stack arg offset for edgecases lemon2026-04-111-1/+2
| | | | | It was wrong for functions that didn't use the frame pointer but took arguments in the stack
* c: fix bitfields in unions lemon2026-04-101-4/+6
|
* x86-64/isel: always test for ZF after shift lemon2026-04-091-3/+3
| | | | | The shift instructions do not update flags if the shift amount was zero, so can't realiably do this micro optimization
* backend: inlining improvements lemon2026-04-093-13/+29
|
* mkintcon: bugfix clamp i32 to i32 range lemon2026-04-092-3/+5
|
* x86-64/isel: fix stack addr regression from 8b79b61ae7 lemon2026-04-091-2/+4
| | | | Checks for RADDR should also look for RSTACK when fusing addresses
* x86-64/isel: use integer stores for storing float constants lemon2026-04-091-2/+16
| | | | | Because the integer mov can have immediate operands, avoids a round trip through an XMM register.
* aarch64 handle more cases of stack relative addressing lemon2026-04-092-3/+21
|
* 0.2.4 lemon2026-04-081-2/+2
|
* Implement basic aarch64 struct arg passing ABI lemon2026-04-0810-139/+204
| | | | | | | | | | - Missing: vaargs, >2 member HFAs - Reworked the way stack allocation references are lowered. Now RSTACK persists throughout all passes until emit. This allows deferring stack frame layouting until the end in a less messy way than before, which was emiting frame-pointer relative addresses @ isel time and patching them up later in emit to account for actual stack frame layout.
* x86-64: redzone optimization for leaf functions lemon2026-03-271-7/+10
| | | | Don't use frame pointer or explicitly modify stack pointer for these
* aarch64: GOT relocations lemon2026-03-263-18/+49
|
* driver: try to autodetect crosscc includs lemon2026-03-251-0/+21
|
* inliner: fix UB when ninstrtab==0 lemon2026-03-251-1/+1
|
* abi: fix use bug preventing mem2reg optimization of struct return lemon2026-03-252-4/+5
|
* aarch64 struct arg passing ABI wip lemon2026-03-255-11/+127
|
* *skip lemon2026-03-241-1/+1
|
* cpp: support directives within macro argument list lemon2026-03-242-55/+82
| | | | Undefined behavior by the standard but a GNU extension.
* driver: -pthread -> -D_REENTRANT & -lpthread lemon2026-03-241-1/+4
|
* regalloc: remove dead code for liveness fixup lemon2026-03-241-100/+22
| | | | | | I tested it and it seems this wasn't doing anything. Maybe at some point in the past it was needed to fix up a worse implementation of mem2reg? So turn it into a check. Maybe it's wrong and it can be necessary.
* ir_inliner: redundant check lemon2026-03-241-1/+0
|
* get rid of imap (was only used for mem2reg, no longer) lemon2026-03-242-111/+0
|
* mem2reg rewrite lemon2026-03-242-108/+74
| | | | | | | Do it in 2 passes: 1. gather variables, 2. step thru RPO transforming loads/stores to each variable. More elegant solution to f9e3a52eaa. More efficient since all variable usages are visited in CFG order, which matters due to block sealing.
* 0.2.3 lemon2026-03-231-2/+2
|
* IR: emit inline function standalone bodies lazily lemon2026-03-2310-41/+139
| | | | | | | If a function is stashed for inlining and inlined in all of its callsites or unused, it never ends up in the object file. If any symbol reference to it is emitted, then it must be de-inlined (rematerialized), and this is done near the end before emitting the actual object file.
generated by cgit v1.2.3 (git 2.39.1) at 2026-05-02 10:31:47 +0000