[llvm-bugs] [Bug 26532] New: don't unroll a loop that has no hope for improvement

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=26532

            Bug ID: 26532
           Summary: don't unroll a loop that has no hope for improvement
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

I'm not sure where this should be solved, but the "unroll-runtime" pass should
be reined in / less aggressive.

I was looking at the example in bug 25108 and wondered what happened with
simpler cases:

int foo(int x, int n) {
  for (int i=0; i<n; i++)
    x = x * x;
  return x;
}

With:
$ ./clang -v
clang version 3.9.0 (trunk 260122)
Target: x86_64-apple-darwin15.3.0

$ ./clang -O2 25108.c -S -emit-llvm -o -

This gets unrolled 8x + prologue code. The net effect is that code which should
be simple like this (-Os or gcc 5.3 with -O2 is similar):
00    testl    %esi, %esi
02    jle    0xb
04    imull    %edi, %edi
07    decl    %esi
09    jne    0x4
0b    movl    %edi, %eax
0d    retq

...is bloated out to 64 bytes with almost no chance that this could be better
for performance on any recent x86:
00    testl    %esi, %esi
02    jle    0x3d
04    leal    -0x1(%rsi), %ecx
07    xorl    %eax, %eax
09    movl    %esi, %edx
0b    andl    $0x7, %edx
0e    je    0x19
10    imull    %edi, %edi
13    incl    %eax
15    cmpl    %eax, %edx
17    jne    0x10
19    cmpl    $0x7, %ecx
1c    jb    0x3d
1e    subl    %eax, %esi
20    imull    %edi, %edi
23    imull    %edi, %edi
26    imull    %edi, %edi
29    imull    %edi, %edi
2c    imull    %edi, %edi
2f    imull    %edi, %edi
32    imull    %edi, %edi
35    imull    %edi, %edi
38    addl    $-0x8, %esi
3b    jne    0x20
3d    movl    %edi, %eax
3f    retq

-- 
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/20160208/4298de9f/attachment.html>