For next release
Need help with
c128 target: various machine specific things (free zp locations, how banking works, getting the floating point routines working, …)
atari target: more details details about the machine, fixing library routines. I have no clue whatsoever.
see the Porting Guide for details on what information is needed.
Future Things and Ideas
add McCarthy evaluation to shortcircuit and/or expressions. First do ifs by splitting them up? Then do expressions that compute a value?
Inliner: also inline function call expressions, and remove it from the StatementOptimizer
vm: implement remaining sin/cos functions in math.p8
vm: somehow deal with asmsubs otherwise the vm IR can’t fully encode all of prog8
vm: don’t store symbol names in instructions to make optimizing the IR easier? but what about jumps to labels. And it’s no longer readable by humans.
vm: how to remove all unused subroutines? (in the 6502 assembly codegen, we let 64tass solve this for us)
vm: rather than being able to jump to any ‘address’ (IPTR), use ‘blocks’ that have entry and exit points -> even better dead code elimination possible too
when the vm is stable and if its language can get promoted to prog8 IL, the variable allocation should be changed. It’s now done before the vm code generation, but the IL should probably not depend on the allocations already performed. So the CodeGen doesn’t do VariableAlloc before the codegen, but as a last step.
createAssemblyAndAssemble(): make it possible to actually get rid of the VarDecl nodes by fixing the rest of the code mentioned there. but probably better to rewrite the 6502 codegen on top of the new Ast.
simplifyConditionalExpression() should not split expression if it still results in stack-based evaluation, but how does it know?
simplifyConditionalExpression() sometimes introduces needless assignment to r9 tempvar (what scenarios?)
make it possible to use cpu opcodes such as ‘nop’ as variable names by prefixing all asm vars with something such as
p8v_? Or not worth it (most 3 letter opcodes as variables are nonsensical anyway) then we can get rid of the instruction lists in the machinedefinitions as well?
[problematic due to using 64tass:] add a compiler option to not remove unused subroutines. this allows for building library programs. But this won’t work with 64tass’s .proc … Perhaps replace all uses of .proc/.pend by .block/.bend will fix that? (but we lose the optimizing aspect of the assembler where it strips out unused code. There’s not really a dynamic switch possible as all assembly lib code is static and uses one or the other)
Zig-like try-based error handling where the V flag could indicate error condition? and/or BRK to jump into monitor on failure? (has to set BRK vector for that)
add special (u)word array type (or modifier?) that puts the array into memory as 2 separate byte-arrays 1 for LSB 1 for MSB -> allows for word arrays of length 256
fix the problems in c128 target, and flesh out its libraries.
fix the problems in atari target, and flesh out its libraries.
c64: make the graphics.BITMAP_ADDRESS configurable (VIC banking)
optimize several inner loops in gfx2 even further?
add modes 2 and 3 to gfx2 (lowres 4 color and 16 color)?
add a flood fill routine to gfx2?
diskio: use cx16 MACPTR() in f_read() to load stuff faster? (see its use in X16edit to fast load blocks) note that it might fail on non sdcard files so have to make graceful degradation
pipe operator: (targets other than ‘Virtual’): allow non-unary function calls in the pipe that specify the other argument(s) in the calls.
rethink the whole “isAugmentable” business. Because the way this is determined, should always also be exactly mirrorred in the AugmentableAssignmentAsmGen or you’ll get a crash at code gen time. note: new ast PtAssignment already has no knowledge about this anymore.
can we get rid of pieces of asmgen.AssignmentAsmGen by just reusing the AugmentableAssignment ? generated code should not suffer
rewrite expression tree evaluation suchthat it doesn’t use an eval stack but flatten the tree into linear code that uses a fixed number of predetermined value ‘variables’? “Three address code” was mentioned. https://en.wikipedia.org/wiki/Three-address_code these variables have to be unique for each subroutine because they could otherwise be interfered with from irq routines etc.
this removes the need for the BinExprSplitter? (which is problematic and very limited now) and perhaps as well the assignment splitting in BeforeAsmAstChanger too
various optimizers skip stuff if compTarget.name==VMTarget.NAME. Once (if?) 6502-codegen is no longer done from the old CompilerAst, those checks should probably be removed, or be made permanent
VariableAllocator: can we think of a smarter strategy for allocating variables into zeropage, rather than first-come-first-served
- translateUnaryFunctioncall() in BuiltinFunctionsAsmGen: should be able to assign parameters to a builtin function directly from register(s), this will make the use of a builtin function in a pipe expression more efficient without using a temporary variable
aa = startvalue(1) |> sin8u() |> cos8u() |> sin8u() |> cos8u()versus:
aa = cos8u(sin8u(cos8u(sin8u(startvalue(1)))))the second one contains no sta cx16.r9L in between.
AssignmentAsmGen.assignExpression() -> better code gen for assigning boolean comparison expressions
when a for loop’s loopvariable isn’t referenced in the body, and the iterations are known, replace the loop by a repeatloop but we have no efficient way right now to see if the body references a variable.
introduce byte-index operator to avoid index multiplications in loops over arrays? see github issue #4