NOTES

BONES                                                                 
=====


* TODO Convert makefile.scm to normal makefile

* TODO "string->number": return NaN if integer result doesn't fit

* TODO Properly integrate use as Xen microkernel
  - or any OS-less platform.
  - code used for test:
    #+BEGIN_SRC scheme
(program
 (provide ieee754 nolibc embedded)
 (code

  (define-syntax define-syntax-rule
    (syntax-rules
	___ ()
	((_ (name args ___) rule)
	 (define-syntax name
	   (syntax-rules ()
	     ((_ args ___) rule))))))

  (define-syntax-rule (%exit code)
    (%terminate code))
  
  (define-syntax %_exit %exit))
 
 (files "x86_64/intrinsics.scm"
	"r5rs.scm"
	"nonstd.scm")
 (code

  (define-syntax-rule (printk str)
    ($inline "CALL copy_to_buffer; LIBCALL1 printk1, buffer; INT2FIX rax" str))

  (printk "BONES!\n")
  (printk (number->string (* 7 4)))))
    #+END_SRC

* TODO loading "megalet" and "match" into si hangs when using match
  - strangely, this doesn't happen in grass.
  - may be caused by different handling of =letrec*= (CPS-transform)

* TODO document bones-autocompile

* TODO work around nasm-limitation in =__float64__= literals
  - nasm doesn't accept an integer, but we could output the raw ieee
    data by extracting the values as two words.
    - would be bones-specific, via =cond-expand=.

* TODO test file:copy.scm

* TODO add checks for copying of slots and bytes (bounds, types)

* TODO add exact development system versions (xcode, mingw64, msvc)

* TODO we need floating-point num/str conversion

* TODO figure out more simplifications

* TODO port embedded test with shared library to Windows
   - and add to file:scripts/check.bat

* add f64vectors
   - FLOATVECTOR_TYPE (#x12)
   - would need proper alignment on 32-bit systems.
   - put API into file:fastmath.scm
   - adapt printer + reader and pp.

* configuration form "(export VAR ...)"
  - makes global externally visible (+ PREFIX)

* potential optimizations
** we could coalesce allocations inside a lambda
  - but will probably gain very little.
** implement basic known procedure optimizations
  - beta-contraction (needs knowledge about FVs).
** directly invoke case-lambda bodies
   - known case-lambda target at call-site could jump to matching case
** handle redundant moves
   - keep list of equivalence-sets for registers, and avoid move if
     dest is known to be equal to src.
** unboxing
  - "($the TYPE EXP)" annotates with type.
  - "let" can check for "$the" and assign FP-register.
  - accessing variables in FP-registers needs to move them to normal registers.
  - requires special intrinsics for FP arguments.
  - implement primitives as "($typed-case-lambda (((VAR TYPE) ...) BODY ...) ...)"
    - specially handled in operator position during canonicalization.
  - have local FA pass ("lfa") that expands $typed-case-lambda according to inferred
    types and convert "let"-bound flonum vars using a custom binding form
    - e.g. "($let-float ...)" or something like this.

* issues
** "inexact->exact" uses "fisttp", so needs SSE3
** i don't fully trust that "integer?" handles all cases correctly
   - is the exponent overflow-test correct?
** align flonums so that block-data is at 16-byte boundary
   - this is supposed to be faster.
   - but needs to be handled in GC (skip padding bytes.)
   - also needed for 32-bit variant with flonum support
     - when allocating flonums, align ALLOC first.
     - GC: when copying, insert marker word before flonum; during scan,
       check for marker word and skip.
** there is no way to access =locals= from other modules or host programs
** something is wrong with musl's "sprintf" (or varargs) implementation
  - "(number->string (expt 2.1 3))" returns "nan".
  - ok with glibc.
** precision of 16 in num2str produces sometimes garbage digits
   - see file:r5rs_pitfalls.scm, where test 8.1 is printed as "8.1..."
   - but precision 15 does not match result of =integer?=, which
     might be inaccurate or needs to be less accurate. This shows
     up when num->str generates a string with ".", but the =integer?=
     test returns true (thus appending another ".0")
** gcc/clang prints warning on Mac when linking compiled code ("PIE disabled")
   - executable still runs, warning currently disabled with "-Wl,-no_pie".
** do popping of the FPU stack in =flonum_expt= in file:x86_64/boneslib.s right
   - currently uses =fstp=, but probably should do a combination of
     =ffree= and =fincstp=.
** SIGINT-handling not working on Windows
   - according to msdn this is just normal.
** embedded check doesn't work on Mac OS
   - linker complains when linking embedded2.o with "-bundle".
   - needs probably some terrible magic specific to shared libraries.

* referencess
  - X86_64:
    - x87: http://www.website.masmforum.com/tutorials/fptute/index.html
    - x86 reference: http://www.fermimn.gov.it/linux/quarta/x86/index.htm
    - SSE: http://softpixel.com/~cwright/programming/simd/sse.php
    - General information: file:/archive/optimizing_assembly.pdf
    - x86-64 ABI: file:/archive/x86-64-abi.pdf
    - Syscalls: http://blog.rchapman.org/post/36801038863/linux-system-call-table-for-x86-64
  - "long mode", bootloader: http://wiki.osdev.org/Entering_Long_Mode_Directly

* release steps:
   1. update NEWS
   2. update version in version.scm and MANUAL.org
   3. merge /develop/ into /master/
   4. "make check"
   5. "make dist"
   6. tag
   7. push, push tags
   8. "make upload"
      (ev. index.html)
   9. (announce)