GCC+ARM Assembly: more than 30 operands in ‘asm’

[mratsim]

See https://github.com/mratsim/constantine/actions/runs/17549567238/job/49838582594?pr=581

This is impacting Constantine integration in Nimbus-eth1, which is in particular targeting the growing SBC market (Single Board Computer):

  Executing task make_lib in /home/runner/work/constantine/constantine/constantine/constantine.nimble
Compiling static library:  lib/libconstantine.a
  compiler config:  --tlsEmulation=off  -d:danger  --panics:on -d:noSignalHandler  --mm:arc -d:useMalloc  --verbosity:0 --hints:off --warnings:off  --passC:-fno-semantic-interposition  --passC:-falign-functions=64  --passC:-fmerge-all-constants
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c: In function ‘mulMont_CIOS_sparebit_asm__OOZconstantineZnamedZconstantsZbls129537795sqrt_u2351’:
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c:388:1: error: more than 30 operands in ‘asm’
  388 | asm volatile(
      | ^~~
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c: In function ‘mulMont_CIOS_sparebit_asm__OOZconstantineZnamedZconstantsZbls129537795sqrt_u1159’:
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c:1080:1: error: more than 30 operands in ‘asm’
 1080 | asm volatile(
      | ^~~
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c: In function ‘sumprodMont_CIOS_spare2bits_asm__OOZconstantineZethereum95bls95signatures_u2214’:
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c:2073:1: error: more than 30 operands in ‘asm’
 2073 | asm volatile(
      | ^~~
Error: execution of an external compiler program 'gcc -c  -w -fmax-errors=3 -pthread -fno-semantic-interposition -falign-functions=64 -fmerge-all-constants -O3 -fno-strict-aliasing -fno-ident -fno-math-errno   -I/home/runner/work/constantine/constantine/nim-linux-arm64-version-2-2-8ea8755cc0252eb52a3f1d68d1c91f8f5f7d449f/lib -I/home/runner/work/constantine/constantine/constantine/bindings -o /home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c.o /home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_mul_mont_arm64.nim.c' failed with exit code: 1


/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_redc_mont_arm64.nim.c: In function ‘redcMont_asm__OOZconstantineZnamedZconstantsZbls129538195subgroups_u2115’:
/home/runner/work/constantine/constantine/constantine/nimcache/libconstantine_static/@pconstantine@smath@sarithmetic@sassembly@slimbs_asm_redc_mont_arm64.nim.c:53:1: error: more than 30 operands in ‘asm’
   53 | asm volatile(
      | ^~~
stack trace: (most recent call last)
/tmp/nimblecache-1194919871666104896/nimscriptapi_2151298592259230405.nim(219, 16)
/home/runner/work/constantine/constantine/constantine/constantine.nimble(310, 15) make_libTask
/home/runner/work/constantine/constantine/constantine/constantine.nimble(298, 5) genStaticLib
/home/runner/work/constantine/constantine/constantine/constantine.nimble(278, 5) compile
/home/runner/work/constantine/constantine/nim-linux-arm64-version-2-2-8ea8755cc0252eb52a3f1d68d1c91f8f5f7d449f/lib/system/nimscript.nim(264, 7) exec
/home/runner/work/constantine/constantine/nim-linux-arm64-version-2-2-8ea8755cc0252eb52a3f1d68d1c91f8f5f7d449f/lib/system/nimscript.nim(264, 7) Error: unhandled exception: FAILED: nim c  --tlsEmulation=off  -d:danger  --panics:on -d:noSignalHandler  --mm:arc -d:useMalloc  --verbosity:0 --hints:off --warnings:off  --passC:-fno-semantic-interposition  --passC:-falign-functions=64  --passC:-fmerge-all-constants --threads:on  --noMain --app:staticlib  --nimMainPrefix:ctt_init_  --out:libconstantine.a --outdir:lib  --nimcache:nimcache/libconstantine_static bindings/lib_constantine.nim [OSError]

The assembly does work with Clang on MacOS.

The code generator should be under the limit of 30 operands:

constantine/constantine/math/arithmetic/assembly/limbs_asm_mul_mont_arm64.nim

Lines 40 to 63 in b3f4ebd

	let
	r = asmArray(r_PIR, N, PointerInReg, asmInput, memIndirect = memWrite)
	b = asmArray(b_PIR, N, PointerInReg, asmInput, memIndirect = memRead)
	M = asmArray(M_REG, N, ElemsInReg, asmInput)
	tSym = ident"t"
	t = asmArray(tSym, N, ElemsInReg, asmOutputEarlyClobber)
	m0ninv = asmValue(m0ninv_REG, Reg, asmInput)
	aSym = ident"A"
	A = asmValue(aSym, Reg, asmOutputEarlyClobber)
	biSym = ident"bi"
	bi = asmValue(biSym, Reg, asmOutputEarlyClobber)
	aaSym = ident"aa"
	aa = asmArray(aaSym, N, ElemsInReg, asmInputOutput) # used as buffer for final substraction
	mSym = ident"m"
	m = asmValue(mSym, Reg, asmOutputEarlyClobber)
	uSym = ident"u"
	vSym = ident"v"
	var # Break dependencies chain
	u = asmValue(uSym, Reg, asmOutputEarlyClobber)
	v = asmValue(vSym, Reg, asmOutputEarlyClobber)

• r -> 1
• b -> 1
• M -> variable, maximum 6 for BLS12-381
• t -> variable, maximum 6 for BLS12-381
• m0ninv -> 1
• A -> 1
• bᵢ -> 1
• aa -> variable, maximum 6 for BLS12-381, counts double as tagged input and output
• m -> 1
• u -> 1
• v -> 1

For a total of 32 operands.

[mratsim]

I have an old branch (Jan 2025) to properly generates assembly files for ARM to sidestep inline assembly limitations here: https://github.com/mratsim/constantine/commits/llvm-arm/

It works currently but instead of one big inline ASM statements it makes one per operation. The performance is the same on MacOS but stack usage is 368 bytes instead of 64 for the optimal single inline assembly: d1e6001#diff-fb4cbc94434a294026d904d27bc3f9f471bc4fc159a14bd534ff510e1dcc7dd8R143

[mratsim]

On MacOS, Nim still uses Clang for some reason even with --cc:gcc (and gcc is aliased to Homebrew GCC)