[llvm-bugs] [Bug 44062] New: instcombine: doesn't remove redundant allocation / memory access with aligned_alloc

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=44062

            Bug ID: 44062
           Summary: instcombine: doesn't remove redundant allocation /
                    memory access with aligned_alloc
           Product: tools
           Version: 8.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: opt
          Assignee: unassignedbugs at nondot.org
          Reporter: uday at polymagelabs.com
                CC: llvm-bugs at lists.llvm.org

In the snippet below, the allocated storage is optimized away when a malloc is
used, but not when an aligned_alloc is used. (aligned_alloc has effectively
similar semantics for the purpose of relevant opts here, and it has been in
glibc since 2.16 and in the C11 standard.)  When a malloc is used, once the
load/stores are hoisted out by LICM, -instcombine gets rid of the heap
allocation and *s is effectively in a virtual register. But the same thing
doesn't happen with aligned_alloc; the allocation and the store to it remain. 

https://godbolt.org/z/WCUF9n
------------------------------------------------
#include <stdlib.h>
#define N 100

void foo(int Y[N], int A[N][N], int X[N]) {
  for (int i = 0; i < N; ++i) {
    // Allocated storage optimized away when malloc is used.
    // int *s = (int*)malloc((sizeof(int)));
    int *s = (int *)aligned_alloc(sizeof(int), sizeof(int));
    *s = Y[i];
    for (int j = 0; j < N; ++j) {
      *s += A[i][j] * X[j];
    }
    Y[i] += *s;
    free(s);
  }
}
--------------

On this note, this issue exists with posix_memalign as well. With
posix_memalign, in fact, the load/stores to *s in the loop aren't replaced by a
scalar. Although this is an artificial C/C++ example to reproduce this issue,
the more interesting/real use case is for example when a pass/utility has to
generate an allocation for a buffer of vector element types -- as such, an
alignment to vector size boundaries is desired, which malloc doesn't provide
for > 16 bytes.

-- 
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/20191119/006fa5c3/attachment.html>