[phix33]

An ELF32 N32 image fails automatic loading even though its decoder and compiler model are bundled.

python3 - <<'PY'
import struct
code = struct.pack('>III', 0x64820001, 0x03e00008, 0)
header = b'\x7fELF\x01\x02\x01' + bytes(9)
header += struct.pack('>HHIIIIIHHHHHH', 2, 8, 1, 0x1000, 52, 0,
                      0x80000020, 52, 32, 1, 40, 0, 0)
segment = struct.pack('>8I', 1, 0x100, 0x1000, 0x1000, 12, 12, 5, 4)
open('n32.elf', 'wb').write((header + segment).ljust(0x100, b'\0') + code)
PY
kuna decompile n32.elf 0x1000 --mode fast 2>&1 | sed 's/.*unsupported machine/unsupported machine/'

unsupported machine Mips64_N32 (no kuna SLEIGH language; pass an explicit --target language id)

Fix

• Select generic MIPS register width, instruction release and ABI from ELF declarations; preserve explicit overrides and leave processor-specific selection separate.
• Read FPR width independently and preserve O32 word slots and byte-order-specific aggregate placement in wider registers.
• Accept R6 with microMIPS while rejecting contradictory mode declarations.

Tests

Loader and CLI cases in both byte orders cover selection, overrides, segment-only metadata, mixed modes, FPR widths and aggregate register/stack placement. Stage cases pin O32 word-slot reads and big-endian one-byte aggregates; each fails without its correction.

[mahaloz]

[AUTOMATED] The N32 repro works, and so do microMIPS, ELF32 O64, EABI32 and R6 -mfp64, all of which main mis-decodes or cannot load. Non-MIPS output and every O32 image main already loads are byte-identical across 516 images. Three problems remain. All use Debian/Ubuntu's gcc-mipsel-linux-gnu (GCC 10).

1. Two kinds of O32 image that main decompiles no longer load.

printf 'int neg(int a) { return a < 0; }\nvoid __start(void) { for (;;); }\n' > t.c
F='-O1 -nostdlib -static -fno-pic -mno-abicalls'
mipsel-linux-gnu-gcc -EB -mabi=32 -march=mips32r6 $F -o r6.elf t.c                    # FP ABI FPXX, CPR1 size 32
mipsel-linux-gnu-gcc -EB -mabi=32 -march=mips32r2 -mfp64 -mhard-float $F -o fp64.elf t.c
kuna decompile r6.elf neg
kuna decompile fp64.elf neg

Both fail with no bundled MIPS language for the declared ISA, ABI and FPR width; pass an explicit --target language id.

• R6: This is plain GCC output for -march=mips32r6. --target MIPS:BE:32:R6:default decodes it correctly (return a0 >> 0x1f;). Under FPXX, the CPR1 size of 32 is the minimum the code needs, not the hardware width, and R6 FPRs are always 64-bit. So registers = Some(4) should not trip r6 && fpr != 8.
• Pre-R6 FP64: main prints return a0 >> 0x1f;. Rejecting these images turns them into hard failures. Falling back to the MIPS32 selection main uses, with a [kuna target] note like the O64 one, would keep them loading.

2. The wide-register O32 models drop every integer argument at call sites.

printf 'int __attribute__((noinline)) add2(int a, int b) { return a + b; }\nint call(void) { return add2(4, 5) + 1; }\nvoid __start(void) { for (;;); }\n' > c.c
mipsel-linux-gnu-gcc -EB -mabi=32 -march=mips64r2 -fno-inline -O1 -nostdlib -static -fno-pic -mno-abicalls -o c.elf c.c
kuna decompile-all c.elf --target MIPS:BE:64:64-32addr:o32

main prints long long add2(int a0,int a1) and return (int)add2(4,5) + 1;. This PR prints int add2(int a0,int a1), which is better, but return add2() + 1;.

• The same happens with MIPS:BE:64:64-32addr-fp32:o32.
• It also happens without --target once EF_MIPS_32BITMODE is cleared (python3 -c "b=bytearray(open('c.elf','rb').read()); b[38]&=~1; open('c2.elf','wb').write(b)"). main prints add2(4,5) there.
• In a caller with ten calls, every call loses its arguments (mkp(); setg(); first_b1(); ...).

The new stage and CLI tests check callee inputs and call results, but not caller-side arguments.

3. Overlap with #1394. #1394 adds the identical kuna_mipsabi.rs and a second mod kuna_mipsabi; in loader/mod.rs. That line auto-merges into a duplicate-module compile error. The two PRs also conflict in kuna_floatabi.rs, mips_markers.rs, loadimage_object.rs and spec chapter 01. Whichever lands second needs a rebase.

[phix33]

Thanks, all three reproduce here.

1. Fixed. Under FPXX the CPR1 size is now a minimum, so r6.elf selects MIPS:BE:32:R6:default. Pre-R6 O32 -mfp64 images fall back to the MIPS32 language main uses, with a [kuna target] note. Both print return a0 >> 0x1f;.
2. Fixed. A call now passes the declared four-byte slot of each argument register the caller writes. The incoming-argument extension is only added for registers the function reads, so untouched registers are not passed. c.elf prints int add2(int a0,int a1) and return add2(4,5) + 1; with both wide O32 targets and with EF_MIPS_32BITMODE cleared. A ten-call caller matches main's calls. A new stage case and CLI test check caller-side arguments.
3. kuna_mipsabi.rs is byte-identical to [AUTOMATED] Recover Octeon ELF targets and N32/N64 ABI values #1394's. Its mod line now sits next to [AUTOMATED] Recover Octeon ELF targets and N32/N64 ABI values #1394's, so merging both conflicts there instead of compiling a duplicate module. Whichever lands second will be rebased.