[Bug middle-end/109766] New: Passing doubles through the stack generates a stack adjustment pear each such argument at -Os/-Oz.

[Bug middle-end/109766] New: Passing doubles through the stack generates a stack adjustment pear each such argument at -Os/-Oz.

[pskocik at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109766

            Bug ID: 109766
           Summary: Passing doubles through the stack generates a stack
                    adjustment pear each such argument at -Os/-Oz.
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pskocik at gmail dot com
  Target Milestone: ---

/*
 Passing doubles through the stack generates a stack adjustment pear each such
argument at -Os/-Oz.
 These stack adjustments are only coalesced at -O1/-O2/-O3, leaving -Os/-Oz
with larger code.
*/
#define $expr(...) (__extension__({__VA_ARGS__;}))
#define $regF0 $expr(register double x __asm("xmm0"); x)
#define $regF1 $expr(register double x __asm("xmm1"); x)
#define $regF2 $expr(register double x __asm("xmm2"); x)
#define $regF3 $expr(register double x __asm("xmm3"); x)
#define $regF4 $expr(register double x __asm("xmm4"); x)
#define $regF5 $expr(register double x __asm("xmm5"); x)
#define $regF6 $expr(register double x __asm("xmm6"); x)
#define $regF7 $expr(register double x __asm("xmm7"); x)

void func(char const*Fmt, ...);
void callfunc(char const*Fmt, double D0, double D1, double D2, double D3,
double D4, double D5, double D6, double D7){
    func(Fmt,$regF0,$regF1,$regF2,$regF3,$regF4,$regF5,$regF6,$regF7,
            D0,D1,D2,D3,D4,D5,D6,D7);
/*
//gcc @ -Os/-Oz
0000000000000000 <callfunc>:
   0:   50                      push   %rax
   1:   b0 08                   mov    $0x8,%al
   3:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
   8:   66 0f d6 3c 24          movq   %xmm7,(%rsp)
   d:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  12:   66 0f d6 34 24          movq   %xmm6,(%rsp)
  17:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  1c:   66 0f d6 2c 24          movq   %xmm5,(%rsp)
  21:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  26:   66 0f d6 24 24          movq   %xmm4,(%rsp)
  2b:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  30:   66 0f d6 1c 24          movq   %xmm3,(%rsp)
  35:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  3a:   66 0f d6 14 24          movq   %xmm2,(%rsp)
  3f:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  44:   66 0f d6 0c 24          movq   %xmm1,(%rsp)
  49:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
  4e:   66 0f d6 04 24          movq   %xmm0,(%rsp)
  53:   e8 00 00 00 00          callq  58 <callfunc+0x58>
                        54: R_X86_64_PLT32      func-0x4
  58:   48 83 c4 48             add    $0x48,%rsp
  5c:   c3                      retq
$sz(callfunc)=93

//clang @ -Os/-Oz
0000000000000000 <callfunc>:
   0:   48 83 ec 48             sub    $0x48,%rsp
   4:   f2 0f 11 7c 24 38       movsd  %xmm7,0x38(%rsp)
   a:   f2 0f 11 74 24 30       movsd  %xmm6,0x30(%rsp)
  10:   f2 0f 11 6c 24 28       movsd  %xmm5,0x28(%rsp)
  16:   f2 0f 11 64 24 20       movsd  %xmm4,0x20(%rsp)
  1c:   f2 0f 11 5c 24 18       movsd  %xmm3,0x18(%rsp)
  22:   f2 0f 11 54 24 10       movsd  %xmm2,0x10(%rsp)
  28:   f2 0f 11 4c 24 08       movsd  %xmm1,0x8(%rsp)
  2e:   f2 0f 11 04 24          movsd  %xmm0,(%rsp)
  33:   b0 08                   mov    $0x8,%al
  35:   e8 00 00 00 00          callq  3a <callfunc+0x3a>
                        36: R_X86_64_PLT32      func-0x4
  3a:   48 83 c4 48             add    $0x48,%rsp
  3e:   c3                      retq   
$sz(callfunc)=63


//gcc @ -O1
0000000000000000 <callfunc>:
   0:   48 83 ec 48             sub    $0x48,%rsp
   4:   f2 0f 11 7c 24 38       movsd  %xmm7,0x38(%rsp)
   a:   f2 0f 11 74 24 30       movsd  %xmm6,0x30(%rsp)
  10:   f2 0f 11 6c 24 28       movsd  %xmm5,0x28(%rsp)
  16:   f2 0f 11 64 24 20       movsd  %xmm4,0x20(%rsp)
  1c:   f2 0f 11 5c 24 18       movsd  %xmm3,0x18(%rsp)
  22:   f2 0f 11 54 24 10       movsd  %xmm2,0x10(%rsp)
  28:   f2 0f 11 4c 24 08       movsd  %xmm1,0x8(%rsp)
  2e:   f2 0f 11 04 24          movsd  %xmm0,(%rsp)
  33:   b8 08 00 00 00          mov    $0x8,%eax
  38:   e8 00 00 00 00          callq  3d <callfunc+0x3d>
                        39: R_X86_64_PLT32      func-0x4
  3d:   48 83 c4 48             add    $0x48,%rsp
  41:   c3                      retq   
$sz(callfunc)=66
*/
}

https://godbolt.org/z/d8T3hxqWK

[Bug middle-end/109766] Passing doubles through the stack generates a stack adjustment per each such argument at -Os/-Oz.

[sjames at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109766

Sam James <sjames at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |roger at nextmovesoftware dot com

--- Comment #1 from Sam James <sjames at gcc dot gnu.org> ---
I'll cc Roger on the off chance he's interested, given he added -Oz.

[Bug middle-end/109766] Passing doubles through the stack generates a stack adjustment per each such argument at -Os/-Oz.

[roger at nextmovesoftware dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109766

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-05-08
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW

--- Comment #2 from Roger Sayle <roger at nextmovesoftware dot com> ---
I believe the problem is in the cprop_hardreg pass, which undoes reload's
register assignments (to use DImode GPR registers with -Os), by propagating DF
mode values into *pushdi2_rex64, which then get split during the split3 pass
into lea/movq pairs, that are each larger than a DImode push.  The work around,
for this test case, is to use -Os -fno-cprop-registers which produces code
that's shorter than -O2.

0000000000000000 <callfunc>:
   0:   66 48 0f 7e ca          movq   %xmm1,%rdx
   5:   66 48 0f 7e d1          movq   %xmm2,%rcx
   a:   66 48 0f 7e de          movq   %xmm3,%rsi
   f:   50                      push   %rax
  10:   66 49 0f 7e e0          movq   %xmm4,%r8
  15:   66 48 0f 7e c0          movq   %xmm0,%rax
  1a:   66 49 0f 7e e9          movq   %xmm5,%r9
  1f:   66 49 0f 7e f2          movq   %xmm6,%r10
  24:   66 49 0f 7e fb          movq   %xmm7,%r11
  29:   41 53                   push   %r11
  2b:   41 52                   push   %r10
  2d:   41 51                   push   %r9
  2f:   41 50                   push   %r8
  31:   56                      push   %rsi
  32:   51                      push   %rcx
  33:   52                      push   %rdx
  34:   50                      push   %rax
  35:   b0 08                   mov    $0x8,%al
  37:   e8 00 00 00 00          callq  3c <callfunc+0x3c>
  3c:   48 83 c4 48             add    $0x48,%rsp
  40:   c3                      retq

Now to figure out if there's a way, using target rtx_costs or pushdi2_rex64's
constraints/predicates, to prevent hardreg cprop performing this substitution.
Plan B might be to investigate reload's choice of DFmode SSE vs DImode GPR, but
this is within one or two bytes of optimal (for four arguments I believe GCC
would produce shorter code than clang).

[Bug middle-end/109766] Passing doubles through the stack generates a stack adjustment per each such argument at -Os/-Oz.

[roger at nextmovesoftware dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109766

--- Comment #3 from Roger Sayle <roger at nextmovesoftware dot com> ---
For the record a solution was proposed at
https://gcc.gnu.org/pipermail/gcc-patches/2023-May/618197.html
but this approach failed review at
https://gcc.gnu.org/pipermail/gcc-patches/2023-May/618278.html