[Bug c++/82888] New: terrible code generation for initialization of POD array members vs. clang

[froydnj at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82888

            Bug ID: 82888
           Summary: terrible code generation for initialization of POD
                    array members vs. clang
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: froydnj at gcc dot gnu.org
                CC: jseward at acm dot org, mh+gcc at glandium dot org
  Target Milestone: ---

Consider the testcase:

class A
{
public:
  A();

private:
  unsigned char mStorage[4096];
};

A::A()
  : mStorage()
{}

gcc -O2 generates a loop that looks like:

.L2:
        movb    $0, (%rdi)
        addq    $1, %rdi
        cmpq    %rdi, %rax
        jne     .L2

which is terribly slow.  (The original motivation for this bug report came from
an -Og compilation, where you get:

        movl    $4095, %eax
.L3:
        testq   %rax, %rax
        js      .L1
        movb    $0, (%rdi)
        addq    $1, %rdi
        subq    $1, %rax
        jmp     .L3

which is even worse.)

clang -O2 (or -O1), on the other hand, generates:

        xorl    %esi, %esi
        movl    $4096, %edx             # imm = 0x1000
        jmp     memset                  # TAILCALL

which is ideal, and opens up opportunities for the backend to lower the memset
to something intelligent (e.g. rep stos)