Compiler — 80cc

Binary: z88dk-80cc. Source: src/80cc/. Rules file: lib/80cc_rules.1.

Agent facts

1. Select via zcc -compiler=80cc when the target allows it.
2. -compiler=multi also runs 80cc. It always compiles 80cc-sp. It compiles 80cc-fp when 80cc treats the CPU as having IX. Spec: src/zcc-multi/zcc-compiler-multi.md.
3. Do not assume sccz80 or sdcc runtime helpers match 80cc output.
4. Prefer reading src/80cc notes for current status before large work.
5. Rebuild: make -C src/80cc PREFIX=$(pwd) from the z88dk root, then make -C src/80cc PREFIX=$(pwd) install.
6. The --version / -h banner comes from src/config.h Z88DK_VERSION. That string can be stale after a rebuild. Prove a hunk with strings bin/z88dk-80cc, not the banner.
7. kc160 mul de,hl is integer 16×16. _Float16 / accum16 stay on l_f16_mul / l_fix16_*. Do not emit IR_MUL for those (math16 80cc kc160 suite: 1.5*1.5 and 0*Inf).

Frame pointer (Z80 benches)

Flag	Meaning
-fframe-pointer	IX is the frame pointer (ix+d locals). Use on Z80 80cc TIMER / PRINTF recipes, including --math-mbf32.
(default)	Omit the frame pointer. Locals via sp. IX is a spare index register.

8085 has no IX. Do not pass -fframe-pointer on -clib=8085.

Integer benches (sieve, fannkuch) are usually smaller and faster with -fframe-pointer. math32 TIMER is usually slower and larger: extra time is in 80cc C glue (ix+d, byte add/adc a,(ix+d), spills around l_f32_*). math32 library cycle counts stay the same. Prove with ticks hotspots rolled up to the C function (advance, selectRandom), not to CLIB_* const symbols.

Recipe comment used in z88dk-classic/readme.txt:

# Z80 80cc: -fframe-pointer (IX).

TIMER / correctness

• Classic: +test -compiler=80cc -DSTATIC -DTIMER -D__Z88DK (no PRINTF for published ticks).
• Some 80cc sources need __asm__("TIMER_START:"); (same as vanilla SDCC).
• Dhrystone: -DNOSTRUCTASSIGN. TIMER is not published (Run_Index loop exits in ~800 cycles).
• 80cc+8085 sorting: qsort does not link. Do not invent 8085 80cc sort rows.
• Mandelbrot image: z88dk-ticks … -output verify.bin then z88dk-appmake +extract -b verify.bin -s 0xc000 -l 480 -o image.bin. Compare to z88dk-classic/image-golden.bin. One edge pixel can differ (float). Many last-bytes-of-row (≈15) means missing variable-count byte << (PR #3066).

PR #3066 (mandelbrot leftover pack)

Not on master until merged. Branch 80cc_mandelbrot_pr.

Hunk	Role
gen_shl width-1 + op->src[1] >= 0	Count in B, add a,a loop. Else imm is 0 and the leftover byte is << 0.
gen_ld_mem ld a,(bc) / (de)	Gate on bc_has / de_has (live), not vreg_in_pr_bc / de.
compound_assign.sh ca_char_shl_var	Expect 00c0. A lone zcc +test of that snippet can print c0 even without the hunk. Trust the suite or the full mandelbrot image.

Use emit_sp for the BC push/pop. Three-way merge onto current master is clean. Do not cherry-pick only the last commit.

Prove the binary has the fix: strings bin/z88dk-80cc | rg bshl_loop.

Condensed wiki

Tool — z88dk-80cc

Alternate small-C style compiler. Select with -compiler=80cc on the zcc line.

Capability summary

Role	A/B codegen vs sccz80; some float benchmarks
Not	Full day-to-day replacement for sccz80/zsdcc
CPUs	See help (-m8080, -m8085, -mz80, …)

Help (live)

80cc - z80 family Crosscompiler   Version: 25141-6b30e0885e-20260716
(C) 1980-2026 Cain, Van Zandt, Hendrix, Yorston, z88dk
Usage: z88dk-80cc [flags] [file]
   -v -verbose                  Be verbose
   -h -help                     Show this help page
   -o -output                   Set the output filename

CPU Targetting:
      -m8080                    Generate output for the i8080
      -m8085                    Generate output for the i8085
      -mez80_z80                Generate output for the ez80 in z80 mode
      -mz80                     Generate output for the z80
      -mz80n                    Generate output for the z80n
      -mz180                    Generate output for the z180
      -mr2ka                    Generate output for the Rabbit 2000A
      -mr3k                     Generate output for the Rabbit 3000
      -mr4k                     Generate output for the Rabbit 4000
      -mr6k                     Generate output for the Rabbit 4000
      -mgbz80                   Generate output for the Gameboy CPU
      -mkc160                   Generate output for the KC160

Code generation options
      -unsigned                 Make all types unsigned
      --disable-builtins        Disable builtin functions
      -params-offset            =<num> Base offset for the function parameters (default: 2)
      -math-z88                 (deprecated) Make FP constants match z88
      -banked-style=regular     Use regular banked calling style
      -banked-style=ti          Use ticalc banked calling style
      -fp-exponent-bias         =<num> FP exponent bias (default: 128)
      -fp-mantissa-size         =<num> FP mantissa size (default: 5 bytes)
      -fp-mode=z80              Use 48 bit doubles
      -fp-mode=ieee             Use 32 bit IEEE doubles
      -fp-mode=mbf32            Use 32 bit Microsoft Binary format
      -fp-mode=mbf40            Use 40 bit Microsoft binary format
      -fp-mode=mbf64            Use 64 bit Microsoft binary format
      -fp-mode=z88              Use 40 bit z88 doubles
      -fp-mode=am9511           Use 32 bit AM9511 doubles
      -standard-escape-chars    Use standard mappings for \r and \n (inert — escapes are always standard)
      -set-r2l-by-default       Use r->l calling convention by default
      --constseg                =<name> Set the const section name
      --codeseg                 =<name> Set the code section name
      --bssseg                  =<name> Set the bss section name
      --dataseg                 =<name> Set the data section name
      --initseg                 =<name> Set the initialisation section name
      -gcline                   Generate C_L

… (more in wiki page)


## Related

- `tool-zcc`, `tool-copt`, `methodology-measure`, `methodology-sdcc-vanilla` (not 80cc)
- Multi selector: `src/zcc-multi/zcc-compiler-multi.md`