[Bug middle-end/37233] 64-bit product of 32-bit value loses optimisation when inlined

[ajrobb at bigfoot dot com]

------- Comment #2 from ajrobb at bigfoot dot com  2008-08-30 07:21 -------
Thanks for that.

I notice that the 64-bit value is now being accumulated in memory rather than
in registers. To be fair, a full 64-bit accumulator need not be maintained -
only the high 32 bits. Could you compile the following version with the fixed
4.4, please?

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
uint64_t mul2(uint32_t a, uint32_t b) {
  return a * (uint64_t)b;
}

void mul32(uint32_t * a, const size_t n, const uint32_t factor) {
  ssize_t i = n;
  uint32_t hi = 0;

  while (--i >= 0) {
    uint64_t p = mul2(*a, factor) + hi;
    *a++ = p;
    hi = p >> 32;
  }

  *a = p;
}

For what it's worth, I tweaked the assembler by hand to avoid local stack
variables:

.globl _mul32
        .def    _mul32; .scl    2;      .type   32;     .endef
_mul32:
        pushl   %ebp
        pushl   %esi
        xorl    %esi, %esi
        pushl   %ebx
        movl    20(%esp), %ecx
        movl    16(%esp), %ebp
        testl   %ecx, %ecx
        movl    24(%esp), %ebx
        je      L6
        .p2align 4,,7
L7:
        movl    (%ebp), %eax
        mull    %ebx
        addl    %esi, %eax
        movl    %eax, (%ebp)
        adcl    $0, %edx
        decl    %ecx
        leal    4(%ebp), %ebp
        movl    %edx, %esi
        jne     L7
L6:
        movl    %esi, (%ebp)
        popl    %ebx
        popl    %esi
        popl    %ebp
        ret


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37233