[llvm-bugs] [Bug 37573] New: AArch64 MachineOutliner incorrectly outlines use of r16/r17

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=37573

            Bug ID: 37573
           Summary: AArch64 MachineOutliner incorrectly outlines use of
                    r16/r17
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AArch64
          Assignee: unassignedbugs at nondot.org
          Reporter: efriedma at codeaurora.org
                CC: jpaquette at apple.com, llvm-bugs at lists.llvm.org

Take the following contrived testcase; compile it with "-O2 -mllvm
-enable-machine-outliner -ffunction-sections", for the target
aarch64-pc-linux-gnu.  The resulting binary prints 19454, when it should print
5.  The problem is that calling the outlined function requires a veneer, so r16
gets clobbered.

It's probably unlikely this will come up in practice... usually the outlined
function will be close enough to its callers that we don't need a veneer.  But
it would be painful to debug if someone did manage to trip over it somehow.

In theory, the linker could also clobber r17, but that doesn't seem to happen
in practice; not sure if it's a good idea to try to take advantage of that.

(Without -ffunction-sections, the testcase fails to compile; that's also an
issue, but it's very unlikely to come up in practice, and it prints a readable
error message.  There are four separate padding functions to work around a bug
in binutils ld: the program fails to link with a single, larger padding
function.)

#include <stdio.h>
int a(int x) {
  register int y asm("r16") = x;
  asm("nop":"+r"(y));
  y = (y*33333+44444)/55555+2;
  asm("nop":"+r"(y));
  return y+1;
}

int b(int x) {
  register int y asm("r16") = x;
  asm("nop":"+r"(y));
  y = (y*33333+44444)/55555+2;
  asm("nop":"+r"(y));
  return y+2;
}
void c() {
  asm(".fill 0x2000000");
}

void c2() {
  asm(".fill 0x2000000");
}

void c3() {
  asm(".fill 0x2000000");
}

void c4() {
  asm(".fill 0x2000000");
}

int main() {
  void * volatile aa = a, * volatile bb = b;
  printf("%d\n", ((int(*)(int))aa)(3));
}

-- 
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/20180524/c1bf028d/attachment.html>