[Bug rtl-optimization/70703] [6/7 regression] Regression in register usage on x86

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70703

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 41094
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41094&action=edit
gcc7-pr70703-widen.patch

The widening_mult change.  We get tiny bit better code with it with the #c0
testcase:
-       movl    $6700417, %ecx
-       movl    %ecx, %eax
+       movl    $6700417, %edx
+       movl    %edx, %eax
        mull    4(%esp)
-       movl    %edx, %ecx
-       movl    %ecx, %eax
+       movl    %edx, %eax
but still not ideal.  On the other side, we regress on -m64:
unsigned long
foo (unsigned long x)
{
  return ((__uint128_t) x * 0x663d811234567ULL) >> 64;
}

-       movabsq $1798629511873895, %rax
-       mulq    %rdi
+       movq    %rdi, %rax
+       movabsq $1798629511873895, %rdx
+       mulq    %rdx

Another option is to deal with this at combine time, I see on the unpatched
compiler:
Failed to match this instruction:
(set (reg:SI 95)
    (subreg:SI (mult:DI (zero_extend:DI (mem/c:SI (reg/f:SI 16 argp) [1 x+0 S4
A32]))
            (const_int 6700417 [0x663d81])) 4))

Maybe we could add some define_insn_and_split that would deal with this and
make sure the constant is forced into a register (if the constant has depending
on <s> all upper bits zero or set) and transform it into the highpart insns?
Though, I'm worried about the regression above we got with the TImode highpart.