Interesting regression in parameter passing (x86_64)

Interesting regression in parameter passing (x86_64)

[Stefan Ring]

I recently noticed that gcc 9 introduced a strange push/pop pair in a
function that does nothing other than shift all arguments by one
position and transfer control to another function:

int func(int, int, int, int, int, int);
int caller(int a, int b, int c, int d, int e) { return func(0, a, b, c, d, e); }

pushq %r12
movl %r8d, %r9d
popq %r12
movl %ecx, %r8d
movl %edx, %ecx
movl %esi, %edx
movl %edi, %esi
xorl %edi, %edi
jmp func

Obviously, pushing and popping r12 serves no useful purpose, and gcc 8
does not produce it. It also disappears when a is used instead of the
constant 0 as the first argument. Where does this come from?

Re: Interesting regression in parameter passing (x86_64)

[Segher Boessenkool]

Hi!

On Sun, Feb 14, 2021 at 12:38:26PM +0100, Stefan Ring via Gcc-help wrote:
> I recently noticed that gcc 9 introduced a strange push/pop pair in a
> function that does nothing other than shift all arguments by one
> position and transfer control to another function:
> 
> int func(int, int, int, int, int, int);
> int caller(int a, int b, int c, int d, int e) { return func(0, a, b, c, d, e); }
> 
> pushq %r12
> movl %r8d, %r9d
> popq %r12
> movl %ecx, %r8d
> movl %edx, %ecx
> movl %esi, %edx
> movl %edi, %esi
> xorl %edi, %edi
> jmp func
> 
> Obviously, pushing and popping r12 serves no useful purpose, and gcc 8
> does not produce it. It also disappears when a is used instead of the
> constant 0 as the first argument. Where does this come from?

The pop was emitted right before the jump, but it was moved to earlier
by the instruction scheduled (sched2).

The prologue/epilogue push and pop r12 because that is a non-volatile
("callee-saved") register.  At the point the prologue and expilogue code
is generated r12 is used in the code, to shuffle these registers through.
It is essentially

  r10 := edi
  r11 := esi
  r12 := edx
  r9  := r8
  r8  := ecx
  ecx := r12
  edx := r11
  esi := r10
  edi := 0

which then by cprop_hardreg is simplified to

  r9  := r8
  r8  := ecx
  ecx := edx
  edx := esi
  esi := edi
  edi := 0

but by then it is too late to omit the push and pop.

Please open a PR (see https://gcc.gnu.org/bugs.html for how).  Thanks!


Segher

Re: Interesting regression in parameter passing (x86_64)

[Stefan Ring]

On Mon, Feb 15, 2021 at 11:45 PM Segher Boessenkool
<segher@kernel.crashing.org> wrote:
>
> Please open a PR (see https://gcc.gnu.org/bugs.html for how).  Thanks!

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99531

Sorry for the long delay!