codegen fixes, --no-gc-check / --fast-math, raise caps, DATE$/TIME$ shift

Four roadmap items:

- codegen: fix parenthesized string comparison.  emit_atom didn't
  consume the body of a string-literal token (`"`), so for
  PRINT (A$+B$ < "ZZZ") it emitted a 0 placeholder, advanced one byte,
  and left "ZZZ" to be reparsed as a variable + extra trailing tokens
  -- the binary then failed to link with `var_ZZ_sng` undeclared.
  emit_atom now skips to the closing quote.  Separately, the
  left_type tracking in emit_num_prec dropped VT_STR after a string +
  string concat (becoming VT_SNG), so the string-comparison codepath
  skipped when the relational operator arrived.  Preserve VT_STR
  through TOK_PLUS when both operands are strings.  Verified: paren
  string-cmp now compiles and produces the same -1 / 0 result as the
  interpreter.

- compiler: --no-gc-check and --fast-math optimization flags.
  --no-gc-check skips the per-line gwrt_check_line() (no string-pool
  GC, no Ctrl+Break trap).  --fast-math drops the divide-by-zero
  guard on `/`; the divisor still goes through (double) so 10/0
  produces inf rather than SIGFPE.  Both threaded through
  codegen_opts_t and exposed in --help.  --inline-arrays from the
  roadmap deferred -- larger refactor.

- interp: raise static caps on 32-bit / Linux builds.  vars 256
  -> 1024, arrays 64 -> 256, MAX_FOR_DEPTH 16 -> 64, MAX_GOSUB_DEPTH
  24 -> 128, MAX_WHILE_DEPTH 16 -> 64.  Codegen FOR_STACK_MAX 16
  -> 64.  Analysis-pass caps: MAX_LINES 4096 -> 8192, MAX_VARS 256
  -> 1024, MAX_GOTOS 256 -> 1024, MAX_DATA 1024 -> 4096,
  MAX_GOSUB_RET 256 -> 1024.  16-bit DOS keeps the original modest
  caps via #ifdef _M_I86 -- the MEDIUM model has a single 64KB
  DGROUP for all static data and the bumped sizes broke runtime
  startup under DOSBox-X.  16-bit binary grew from 128KB to 132KB
  from the offset_secs field plus DATE$/TIME$ shift code, well
  within the FreeDOS budget.

- interp + codegen: DATE$ / TIME$ assignment via process-local
  clock offset.  Was a no-op accept-and-ignore.  Now sets
  gw.time_offset_secs (long), and DATE$ / TIME$ / TIMER readers
  apply it to time(NULL) before formatting.  The OS clock is
  unaffected (would need root).  Compiled-binary readers also
  reference gw.time_offset_secs since libgwrt shares the gw
  struct.  Verified: PRINT DATE$; DATE$="12-31-1999"; PRINT DATE$
  shows the expected before/after in both interpreter and AOT
  paths.

After these changes: 72/72 interpreter tests, 68/68 compat, 63/63
compiler tests, DOS smoke under DOSBox-X all pass.  Build clean on
both Linux (cmake) and 16-bit DOS (build_dos.sh 16).
This commit is contained in:
Eremey Valetov
2026-05-04 18:56:58 -04:00
parent da1e6cebf1
commit f207d74aec
9 changed files with 162 additions and 46 deletions
+23 -7
View File
@@ -113,16 +113,32 @@ line numbers are preserved. Direct-mode scratchpad scripts and classic
```
Usage: gwbasic-compile [options] input.bas
Options:
-o FILE Output C source file (default: stdout)
-c Compile to executable (invoke gcc)
-O LEVEL GCC optimization level (default: 2)
--keep-c Keep generated C file (with -c)
--runtime DIR Path to runtime headers/library
--warn Static analysis warnings
--safe Runtime safety checks (implies --warn)
-o FILE Output C source file (default: stdout)
-c Compile to executable (invoke gcc)
-O LEVEL GCC optimization level (default: 2)
--keep-c Keep generated C file (with -c)
--runtime DIR Path to runtime headers/library
--warn Static analysis warnings
--safe Runtime safety checks (implies --warn)
--safe=sanitize Above + address/UB sanitizers (with -c)
--no-gc-check Skip per-line gwrt_check_line() (no GC, no Break)
--fast-math Skip division-by-zero checks
```
### Performance Flags (`--no-gc-check` / `--fast-math`)
`--no-gc-check` skips the `gwrt_check_line()` call emitted at the start of
every non-REM line. That call drives the string-pool compacting GC and
the Ctrl+Break trap. Removing it gives a small per-line speedup for
programs that don't allocate strings or need responsive interruption.
String reassignment can still trigger compaction lazily, but the
guaranteed periodic check is gone.
`--fast-math` removes the explicit divide-by-zero check around the `/`
operator. The result of `X = 10 / 0` becomes `inf` rather than raising
"Division by zero". Useful for compute-bound code that already validates
inputs.
### Memory Safety (`--warn` / `--safe`)
The `--warn` flag enables compile-time static analysis warnings:
+7 -3
View File
@@ -112,8 +112,12 @@ Tested on FreeDOS 1.4 via QEMU.
## Known Limitations
- Maximum 256 variables, 64 arrays, 16 FOR nesting, 24 GOSUB nesting,
16 WHILE nesting
- Static caps -- 32-bit / Linux builds: 1024 variables, 256 arrays,
64 FOR nesting, 128 GOSUB nesting, 64 WHILE nesting. 16-bit real-mode
DOS keeps the original modest caps (256 / 64 / 16 / 24 / 16) because
the MEDIUM model has a single 64KB DGROUP for all static data.
- `CALL`/`CALLS` (machine code execution) raises Illegal function call
- `DATE$`/`TIME$` assignment accepted but does not modify the system clock
- `DATE$`/`TIME$` assignment shifts the program's view of the clock via
a process-local offset; the OS time is unaffected (setting the OS
clock would require root)
- Device stubs (`ERDEV`, `IOCTL`, `COM`, `LCOPY`) return defaults