[PATCH v3 04/19] Add umul_ppmm to gmp-arch.hdoc

[Maciej W. Rozycki]

On Fri, 31 Oct 2025, Adhemerval Zanella wrote:

> Most of the architecture uses the generic implementation, which is
> expanded from a macro, except for alpha, arm, hppa, x86, m68k, mips,
> powerpc, and sparc.  I kept only x86 optimization, where there is no
> easy way to emit mul{q}. For the rest, the compiler generates good
> enough code.

 Hmm, this reference code works for me:

$ cat mulq.c
typedef unsigned int __attribute__ ((mode (DI))) uint64_t;
typedef unsigned int __attribute__ ((mode (TI))) uint128_t;

uint128_t
mulq (uint64_t x, uint64_t y)
{
  return (uint128_t) x * (uint128_t) y;
}
$ x86_64-linux-gnu-gcc -O2 -S mulq.c -o x86_64-linux-gnu-mulq.s
$ cat x86_64-linux-gnu-mulq.s
	.file	"mulq.c"
	.text
	.p2align 4
	.globl	mulq
	.type	mulq, @function
mulq:
.LFB0:
	.cfi_startproc
	movq	%rdi, %rax
	mulq	%rsi
	ret
	.cfi_endproc
.LFE0:
	.size	mulq, .-mulq
	.ident	"GCC: (GNU) 11.0.0 20200919 (experimental)"
	.section	.note.GNU-stack,"",@progbits
$ 

Is this not what you need?  It should work for any 64-bit platform that 
has a widening multiply operation defined in the compiler backend, e.g. 
for MIPS64:

$ mips64-linux-gnu-gcc -O2 -S mulq.c -o mips64-linux-gnu-mulq.s
$ cat mips64-linux-gnu-mulq.s
	.file	1 "mulq.c"
	.section .mdebug.abi64
	.previous
	.nan	legacy
	.module	fp=64
	.module	oddspreg
	.module	arch=mips3
	.abicalls
	.text
	.align	2
	.align	3
	.globl	mulq
	.set	nomips16
	.set	nomicromips
	.ent	mulq
	.type	mulq, @function
mulq:
	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
	.mask	0x00000000,0
	.fmask	0x00000000,0
	.set	noreorder
	.set	nomacro
	dmultu	$4,$5
	mflo	$3
	mfhi	$2
	jr	$31
	nop

	.set	macro
	.set	reorder
	.end	mulq
	.size	mulq, .-mulq
	.ident	"GCC: (GNU) 12.0.1 20220404 (experimental)"
	.section	.note.GNU-stack,"",@progbits
$

  Maciej