optimization/6880: Inlining inefficiencies

[Dan Nicolaescu <dann@godzilla.ics.uci.edu>]

>Number:         6880
>Category:       optimization
>Synopsis:       Inlining inefficiencies
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          pessimizes-code
>Submitter-Id:   net
>Arrival-Date:   Fri May 31 00:36:01 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     Dan Nicolaescu <dann@godzilla.ics.uci.edu>
>Release:        gcc-3.1
>Organization:
>Environment:
sparc-sun-solars2.7
>Description:
There are some problems with inlining as shown by the code below 
(derived from oopack)

class Complex_d {
public:
  double re, im;
  Complex_d (double r, double i) : re(r), im(i) {}
  Complex_d () {}
};

inline Complex_d operator+ (Complex_d   a, Complex_d   b)
{
    return Complex_d (a.re+b.re, a.im+b.im);
}

Complex_d Zd;

void
foo (void)
{
    Complex_d factor (123.2374, 428.234);
    // You'd think this function would be optimized to just do a couple of
    // stores? Nope.
    Zd = factor + factor;
}

class Complex_i {
public:
  int re, im;
  Complex_i (int r, int i) : re(r), im(i) {}
  Complex_i () {}
};

inline Complex_i operator+ (Complex_i   a, Complex_i   b)
{
    return Complex_i ( a.re+b.re, a.im+b.im);
}

Complex_i Zi;

void
bar (void)
{
    Complex_i factor (123, 428);
    Zi = factor + factor;
}

void
foobar (void)
{
    Complex_i factor (123, 428);
    factor = factor + factor;
}


The SPARC assembly generated by gcc-3.1 -O3 is: 

_Z3foov:
.LLFB2:
        !#PROLOGUE# 0
        save    %sp, -176, %sp
.LLCFI0:
        !#PROLOGUE# 1
        sethi   %hi(.LLC0), %i0
        ldd     [%i0+%lo(.LLC0)], %f14
        sethi   %hi(.LLC1), %g1
        std     %f14, [%fp-32]
        ldd     [%fp-32], %i2
        sethi   %hi(Zd), %i4
        ldd     [%g1+%lo(.LLC1)], %f2
        std     %i2, [%fp-80]
        std     %i2, [%fp-64]
        std     %f2, [%fp-24]
        ldd     [%fp-64], %f12
        or      %i4, %lo(Zd), %g1
        ldd     [%fp-80], %f4
        ldd     [%fp-24], %i0
        faddd   %f12, %f4, %f8
        std     %i0, [%fp-72]
        std     %i0, [%fp-56]
        ldd     [%fp-56], %f10
        ldd     [%fp-72], %f6
        std     %f8, [%fp-48]
        faddd   %f10, %f6, %f0
        ldd     [%fp-48], %i0
        std     %i0, [%i4+%lo(Zd)]
        std     %f0, [%fp-40]
        ldd     [%fp-40], %i0
        std     %i0, [%g1+8]
        nop
        ret
        restore

A couple of things are wrong here: 
  1. the reserved stack space is too big
  2. most of the stores are dead stores but they are not eliminated.
     Isn't there enough information to determine that?

  3. it looks like arguments to the inlined functions are moved from
     memory to integer registers then back to memory then to fp
     registers. (side note: the SPARC v8 ABI specifies that fp values
     are passed in integer registers, there no mov INTREG, FPREG
     instruction, but still all this should not be done for an inlined
     function)


The code looks a little better for "bar", but still there's a lot of

_Z3barv:
.LLFB3:
        !#PROLOGUE# 0
        add     %sp, -144, %sp
.LLCFI1:
        !#PROLOGUE# 1
        mov     123, %o3
        mov     428, %g1
        st      %g1, [%sp+124]
        st      %o3, [%sp+120]
        ldd     [%sp+120], %o2
        sethi   %hi(Zi), %o5
        std     %o2, [%sp+96]
        std     %o2, [%sp+104]
        ld      [%sp+104], %o2
        or      %o5, %lo(Zi), %g1
        ld      [%sp+96], %o0
        ld      [%sp+108], %o3
        add     %o2, %o0, %o4
        ld      [%sp+100], %o1
        st      %o4, [%o5+%lo(Zi)]
        add     %o3, %o1, %o2
        st      %o2, [%g1+4]
        nop
        retl
        sub     %sp, -144, %sp


And finally the stack adjustment code is useless for "foobar":

_Z6foobarv:
.LLFB4:
        !#PROLOGUE# 0
        add     %sp, -144, %sp
.LLCFI2:
        !#PROLOGUE# 1
        nop
        retl
        sub     %sp, -144, %sp

I don't know how much of a performance impact all these have on real
programs, it seems that any inline function that is passed objects by
value would be affected.

Also see the thread starting at:
http://gcc.gnu.org/ml/gcc/2002-05/msg02821.html

>How-To-Repeat:

>Fix:

>Release-Note:
>Audit-Trail:
>Unformatted: