ir/TODO
2022-04-14 23:03:30 +03:00

98 lines
2.6 KiB
Plaintext

- type casting nodes
- overflow detection nodes
- va_arg nodes
- BSTART, BEND nodes (to free data allocated by ALLOCA)
- ENTRY node for multy-entry units
- IJMP
- guards
- variable name binding
- VLOAD, VSTORE -> SSA
? reassociation folding rules
- folding engine improvement (one rule for few patterns)
- irreducable loops detection
- irreducable loops support
- range inference and PI node
- SCCP edge cases
- Folding after SCCP (see combo4.ir)
- local scheduling according to data dependencies, register presure and pipeline stalls
- basic block trace scheduling
? C code generation
- VLOAD, VSTORE
? instruction selection
- xor, btsl=INCL, btrl=EXCL, btl=IN, bsr, maxss, maxsd, minss, minsd
- MOVZX to avoid a SHIFT and AND instruction
- Using CMOVcc to remove branches
- mov rax, 123456789abcdef0h ; 10 bytes (64-bit constant)
- mov rax, -100 ; 7 bytes (32-bit sign-extended)
- mov eax, 100 ; 5 bytes (32-bit zero-extended)
? register allocation
+ linear scan
- allocate scratch registers first
- separate INT and FP allocation phases
+ use positions
? fixed registers constraints
+ fixed input
+ fixed output
+ fixed tmp
+ fixed intervals for parameters
- accurate fixed interval construction
- constraints
- kill
- restricted regset
- temporary registers
? spills
+ spill slot allocation ( uniqe for each VREG)
- uniqe for each VREG
- SpillRange ???
- splitting
- SplitAt
- splinting
- spill only at cold path if possible
? hints
+ hints for fixed input/outpu registers
+ hints for parameter nodes
- hints for call arguments
- hints propagation
- hints and low priority
? code generation
- POW, NEG (fp), ABS, OVERFLOW, MIN, MAX, COND
- TAILCALL
- ALLOCA, VLOAD, VSTORE, LOAD, STORE
+ SWITCH
- ir_last_use
- 64-bit constant (ir_emit_ref, ir_emit_dssa_move)
- binop_int $imm, mem
- commutative insns and swap (binop_int, mul, binop_sse, binop_avx, cmp_int, cmp_fp, cmp_br_int)
- dessa_move (push, pop)
- param_move
- temporary register (e.g. for unsupported immediate operand in mul, div)
+ temporary register for swap (dessa3.ir)
- temporary register for spill loads and stores
- parallel parameter loading
+ parallel argument passing
- return merge/split
? binary code emission
+ DynAsm
- BinAsm
- 32-bit x86 code
? disassembler
- .rodata section and relative data labels
- modules (functions, data objecs, import, export, prototypes, forward declarations, memory segments, ref data, expr data)
- C compiler
- interpreter
- alias analyzes
- PHP support