[llvm-bugs] [Bug 25261] New: [x86] hoist complex address generation common subexpressions?

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=25261

            Bug ID: 25261
           Summary: [x86] hoist complex address generation common
                    subexpressions?
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

There's a potential optimization remaining from the test case in bug 20134:

void foo(int *a, int i) {
    a[i] = a[i+1] + a[i+2];
}

After:
http://llvm.org/viewvc/llvm-project?view=revision&revision=250560

We generate:
00    movslq   %esi, %rax
03    movl     0x4(%rdi,%rax,4), %ecx
07    addl     0x8(%rdi,%rax,4), %ecx
0b    movl     %ecx, (%rdi,%rax,4)
0e    retq

We could hoist the common factor (%rdi,%rsi,4) out of the load and store
addresses:

00    movslq   %esi, %rax
03    leaq     (%rdi,%rax,4), %rax
07    movl     0x4(%rax), %ecx
0a    addl     0x8(%rax), %ecx
0d    movl     %ecx, (%rax)
0f    retq

This probably requires simulation and micro-benchmarking across
micro-architectures to decide if it's worthwhile. The LEA option appears to
increase code size by 1 byte (hand-hacked asm - didn't verify that it's
correct) for this case where the CSE is used 3 times. In the general case,
using an LEA may also require an additional register. Also, whether the simpler
address calcs actually provide a time-savings or not may be
implementation-dependent.

The solution to this problem may be related to whatever is implemented for bug
24447.

-- 
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/20151020/185a18d6/attachment-0001.html>