[llvm-bugs] [Bug 38554] New: Loop optimization makes code bloat when moving conditionally called pure functions outside of loop

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=38554

            Bug ID: 38554
           Summary: Loop optimization makes code bloat when moving
                    conditionally called pure functions outside of loop
           Product: libraries
           Version: 6.0
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: o_kniemeyer at maxon.de
                CC: llvm-bugs at lists.llvm.org

If you compile this code snippet at https://gcc.godbolt.org with -O3

int Func(const void*) __attribute__ ((pure));

int Test(int n, const void* ptr2)
{
    const void* ptr = ptr2;
    int result = 0;
    for (int i = 0; i < n; ++i)
    {
#if 1
        result += Func(ptr);
#else
        result += ptr ? Func(ptr) : 0;
#endif
    }
    return result;
}

you get the expected result:

Test(int, void const*): # @Test(int, void const*)
  pushq %rbx
  movl %edi, %ebx
  testl %ebx, %ebx
  jle .LBB0_1
  movq %rsi, %rdi
  callq Func(void const*)
  imull %ebx, %eax
  popq %rbx
  retq
.LBB0_1:
  xorl %eax, %eax
  popq %rbx
  retq

So the optimizer has moved the call to Func outside of the loop, and there's
just a multiplication with the loop count. If you use the code with nullptr
check (#if 0) the call is also moved to the outside, but the multiplication is
replaced by code bloat:

Test(int, void const*): # @Test(int, void const*)
  pushq %rbx
  movl %edi, %ebx
  xorl %ecx, %ecx
  testl %ebx, %ebx
  jle .LBB0_13
  testq %rsi, %rsi
  je .LBB0_13
  movq %rsi, %rid
  callq Func(void const*)
  xorl %edx, %edx
  movl $0, %ecx
  cmpl $7, %ebx
  jbe .LBB0_3
  movl %ebx, %edx
  andl $-8, %edx
  movd %eax, %xmm0
  pshufd $0, %xmm0, %xmm0 # xmm0 = xmm0[0,0,0,0]
  leal -8(%rdx), %esi
  movl %esi, %edi
  shrl $3, %edi
  leal 1(%rdi), %ecx
  andl $7, %ecx
  cmpl $56, %esi
  jae .LBB0_7
  pxor %xmm1, %xmm1
  pxor %xmm2, %xmm2
  testl %ecx, %ecx
  jne .LBB0_10
  jmp .LBB0_12
.LBB0_7:
  leal -1(%rcx), %esi
  subl %edi, %esi
  pxor %xmm1, %xmm1
  pxor %xmm2, %xmm2
.LBB0_8: # =>This Inner Loop Header: Depth=1
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  addl $8, %esi
  jne .LBB0_8
  testl %ecx, %ecx
  je .LBB0_12
.LBB0_10:
  negl %ecx
.LBB0_11: # =>This Inner Loop Header: Depth=1
  paddd %xmm0, %xmm1
  paddd %xmm0, %xmm2
  addl $1, %ecx
  jne .LBB0_11
.LBB0_12:
  paddd %xmm2, %xmm1
  pshufd $78, %xmm1, %xmm0 # xmm0 = xmm1[2,3,0,1]
  paddd %xmm1, %xmm0
  pshufd $229, %xmm0, %xmm1 # xmm1 = xmm0[1,1,2,3]
  paddd %xmm0, %xmm1
  movd %xmm1, %ecx
  cmpl %ebx, %edx
  je .LBB0_13
.LBB0_3:
  subl %edx, %ebx
.LBB0_4: # =>This Inner Loop Header: Depth=1
  addl %eax, %ecx
  addl $-1, %ebx
  jne .LBB0_4
.LBB0_13:
  movl %ecx, %eax
  popq %rbx
  retq

Adding -fno-vectorize helps a bit, but I doubt that the code is faster than
imull.

-- 
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/20180814/55f64aaf/attachment.html>