[llvm-bugs] [Bug 49955] New: Unrolled loop generates worse code

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=49955

            Bug ID: 49955
           Summary: Unrolled loop generates worse code
           Product: clang
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: -New Bugs
          Assignee: unassignedclangbugs at nondot.org
          Reporter: davidfromonline at gmail.com
                CC: htmldeveloper at gmail.com, llvm-bugs at lists.llvm.org,
                    neeilans at live.com, richard-llvm at metafoo.co.uk

The following code

```
#include <vector>

void f() {
        auto v = std::vector<int>();
        v.reserve(2);
#if defined LOOP
        for (int n = 0; n != 2; ++n) {
                v.push_back(0);
        }
#else
        v.push_back(0);
        v.push_back(0);
#endif
}
```

When compiled with `clang++ -stdlib=libc++ -O3 -DLOOP` (the loop version)
compiles down to just

```
f():                                  # @f()
        ret
```

But when compiled with `clang++ -stdlib=libc++ -O3` (the unrolled version)
generates

```
f():                                  # @f()
        push    rbx
        sub     rsp, 32
        xorps   xmm0, xmm0
        movaps  xmmword ptr [rsp], xmm0
        mov     qword ptr [rsp + 16], 0
        mov     edi, 8
        call    operator new(unsigned long)
        mov     rcx, rax
        add     rcx, 8
        mov     qword ptr [rsp], rax
        mov     qword ptr [rsp + 8], rax
        mov     qword ptr [rsp + 16], rcx
        mov     qword ptr [rax], 0
        mov     rdi, rax
        add     rsp, 32
        pop     rbx
        jmp     operator delete(void*)                          # TAILCALL
        mov     rbx, rax
        mov     rdi, qword ptr [rsp]
        test    rdi, rdi
        je      .LBB0_4
        mov     qword ptr [rsp + 8], rdi
        call    operator delete(void*)
.LBB0_4:                                #
%_ZNSt3__16vectorIiNS_9allocatorIiEEED2Ev.exit69
        mov     rdi, rbx
        call    _Unwind_Resume at PLT
```

See it live: https://godbolt.org/z/3MY3YMME6


A similar (but less dramatic) example can be seen in the following much reduced
code:

```
int buffer[3];

struct vector {
        int * ptr = buffer;
        long size = 0;
        long capacity = 0;
};

auto a() {
        auto v = vector();

#if defined LOOP
        for (int n = 0; n != 3; ++n) {
                v.ptr[n] = 0;
        }
        v.size = 3;
#elif defined LOOP_ASSIGN_INSIDE
        for (int n = 0; n != 3; ++n) {
                v.ptr[n] = 0;
                v.size = 3;
        }
#else
        v.ptr[0] = 0;
        v.ptr[1] = 0;
        v.ptr[2] = 0;
        v.size = 3;
#endif
        return v;
}
```

When compiled with `clang++ -O3 -DLOOP` this gives

```
a():                                  # @a()
        mov     rax, rdi
        mov     qword ptr [rdi], offset buffer
        mov     qword ptr [rdi + 16], 0
        mov     dword ptr [rip + buffer+8], 0
        mov     qword ptr [rip + buffer], 0
        mov     qword ptr [rdi + 8], 3
        ret
buffer:
        .zero   12
```

When compiled with `clang++ -O3 -DLOOP_ASSIGN_INSIDE` gives

```
.LCPI0_0:
        .long   3                               # 0x3
        .long   0                               # 0x0
        .long   0                               # 0x0
        .long   0                               # 0x0
a():                                  # @a()
        mov     rax, rdi
        mov     qword ptr [rdi], offset buffer
        movaps  xmm0, xmmword ptr [rip + .LCPI0_0] # xmm0 = [3,0,0,0]
        movups  xmmword ptr [rdi + 8], xmm0
        mov     dword ptr [rip + buffer+8], 0
        mov     qword ptr [rip + buffer], 0
        ret
buffer:
        .zero   12
```

When compiled with `clang++ -O3` gives

```
.LCPI0_0:
        .long   3                               # 0x3
        .long   0                               # 0x0
        .long   0                               # 0x0
        .long   0                               # 0x0
a():                                  # @a()
        mov     rax, rdi
        mov     qword ptr [rdi], offset buffer
        mov     qword ptr [rip + buffer], 0
        mov     dword ptr [rip + buffer+8], 0
        movaps  xmm0, xmmword ptr [rip + .LCPI0_0] # xmm0 = [3,0,0,0]
        movups  xmmword ptr [rdi + 8], xmm0
        ret
buffer:
        .zero   12
```

See it live https://godbolt.org/z/8EG5db4fc

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210413/b1153187/attachment.html>