[Bug target/71475] [6/7 Regression] Optimization of copying into long double looses bytes

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71475

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Target|                            |i?86-*-*
             Status|ASSIGNED                    |NEW
          Component|c                           |target
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org
   Target Milestone|---                         |6.2
            Summary|Optimization of copying     |[6/7 Regression]
                   |into long double looses     |Optimization of copying
                   |bytes                       |into long double looses
                   |                            |bytes

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so we arrive (slowly) at

  MEM[(char * {ref-all})&d0] = 0x41414141414141414141414141414141;
  d0.3_6 = d0;
  d = d0.3_6;
  d.4_7 = d;
  printf ("after memcpy: %Lf\n", d.4_7);

which is then optimized by CSE to

  MEM[(char * {ref-all})&d0] = 0x41414141414141414141414141414141;
  d = 4.35573826932891467758901725805789285479666446831741854231e+96;
  printf ("after memcpy: %Lf\n",
4.35573826932891467758901725805789285479666446831741854231e+96);

and thus the issue must be that FP literal passing to printf is broken somehow.

The first call is expanded to:

;; printf ("after memset: %Lf\n", d0.0_1);

(insn 28 27 29 (set (mem:XF (pre_dec:DI (reg/f:DI 7 sp)) [1  S16 A128])
        (reg:XF 99 [ d0.0_1 ])) t.c:11 -1
     (expr_list:REG_ARGS_SIZE (const_int 16 [0x10])
        (nil)))

(insn 29 28 30 (set (reg:DI 5 di)
        (symbol_ref/f:DI ("*.LC0") [flags 0x2]  <var_decl 0x7fa583c676c0
*.LC0>)) t.c:11 -1
     (nil))

(insn 30 29 31 (set (reg:QI 0 ax)
        (const_int 0 [0])) t.c:11 -1
     (nil))

(call_insn 31 30 0 (set (reg:SI 0 ax)
        (call (mem:QI (symbol_ref:DI ("printf") [flags 0x41]  <function_decl
0x7fa583b25000 printf>) [0 __builtin_printf S1 A8])
            (const_int 16 [0x10]))) t.c:11 -1
     (expr_list:REG_CALL_DECL (symbol_ref:DI ("printf") [flags 0x41] 
<function_decl 0x7fa583b25000 printf>)
        (nil))

while the second:

;; printf ("after memcpy: %Lf\n",
4.35573826932891467758901725805789285479666446831741854231e+96);

(insn 53 52 54 (set (reg:XF 116)
        (mem/u/c:XF (symbol_ref/u:DI ("*.LC2") [flags 0x2]) [1  S16 A128]))
t.c:20 -1
     (nil))

(insn 54 53 55 (set (mem:XF (pre_dec:DI (reg/f:DI 7 sp)) [1  S16 A128])
        (reg:XF 116)) t.c:20 -1
     (expr_list:REG_ARGS_SIZE (const_int 16 [0x10])
        (nil)))

(insn 55 54 56 (set (reg:DI 5 di)
        (symbol_ref/f:DI ("*.LC3") [flags 0x2]  <var_decl 0x7fa583c677e0
*.LC3>)) t.c:20 -1
     (nil))

(insn 56 55 57 (set (reg:QI 0 ax)
        (const_int 0 [0])) t.c:20 -1
     (nil))
(call_insn 57 56 0 (set (reg:SI 0 ax)
        (call (mem:QI (symbol_ref:DI ("printf") [flags 0x41]  <function_decl
0x7fa583b25000 printf>) [0 __builtin_printf S1 A8])
            (const_int 16 [0x10]))) t.c:20 -1
     (expr_list:REG_CALL_DECL (symbol_ref:DI ("printf") [flags 0x41] 
<function_decl 0x7fa583b25000 printf>)
        (nil))

with assembly:

        movabsq $4702111234474983745, %rax
...
        movq    %rax, 16(%rsp)
        movq    %rax, 24(%rsp)
        xorl    %eax, %eax
        fldt    16(%rsp)
        fld     %st(0)
        fstpt   32(%rsp)
        fstpt   (%rsp)
        call    printf

vs.

        fldt    .LC2(%rip)
        movabsq $4702111234474983745, %rdx
        movabsq $4702111234474983745, %rax
        subq    $16, %rsp
        .cfi_def_cfa_offset 80
        movq    %rax, 16(%rsp)
        movq    %rdx, 24(%rsp)
        movl    $.LC3, %edi
        xorl    %eax, %eax
        fld     %st(0)
        fstpt   32(%rsp)
        fstpt   (%rsp)
        call    printf

So confirmed, but it looks like an RTL optimization / target (regstack related)
issue.  The testcase itself is a GCC 6 regression but I guess that using
a FP literal (if one can express that value...) would produce the same issue?