[llvm-bugs] [Bug 33084] New: SimplifyCFG sinks instructions too aggressively

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=33084

            Bug ID: 33084
           Summary: SimplifyCFG sinks instructions too aggressively
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Transformation Utilities
          Assignee: unassignedbugs at nondot.org
          Reporter: efriedma at codeaurora.org
                CC: james.molloy at arm.com, llvm-bugs at lists.llvm.org

Testcase:

int f(int a, int b, int g(void)) {
  int c;
  if (g) {
    g(); c = (a + 31) & -32;
  } else {
    c = (b + 62) & -64;
  }
  return c;
}

Pass this to clang -O2, and you get IR something like this:

define i32 @f(i32 %a, i32 %b, i32 ()* %g) local_unnamed_addr #0 {
entry:
  %tobool = icmp eq i32 ()* %g, null
  br i1 %tobool, label %if.else, label %if.then

if.then:                                          ; preds = %entry
  %call = tail call i32 %g() #1
  %add = add nsw i32 %a, 31
  br label %if.end

if.else:                                          ; preds = %entry
  %add1 = add nsw i32 %b, 62
  br label %if.end

if.end:                                           ; preds = %if.else, %if.then
  %.sink = phi i32 [ -64, %if.else ], [ -32, %if.then ]
  %add1.sink = phi i32 [ %add1, %if.else ], [ %add, %if.then ]
  %and2 = and i32 %add1.sink, %.sink
  ret i32 %and2
}

Sinking the "and" here might not make sense: we're generating two phi nodes to
remove one "and" instruction.

We started sinking the "and" in r280351.

I'm planning to continue looking into this, but suggestions welcome.

-- 
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/20170518/7f6bb2d1/attachment.html>