[llvm-bugs] [Bug 24818] New: hoisting expensive ops like FP division is bad for perf

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=24818

            Bug ID: 24818
           Summary: hoisting expensive ops like FP division is bad for
                    perf
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Transformation Utilities
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Test case based on bug 24807:

double foo(double a, double b) {
    return (a > 0.0) ? b/a : 0.0;
}

If 'a' is always less than zero, then we should never clog up an FPU with a
potentially very expensive, non-pipelined division instruction. But we're
hoisting that op with 'opt' built from r247615:

target triple = "x86_64-apple-macosx10.10.0"

define double @foo(double %a, double %b) #0 {
entry:
  %cmp = fcmp ogt double %a, 0.000000e+00
  %div = fdiv double %b, %a
  %cond = select i1 %cmp, double %div, double 0.000000e+00
  ret double %cond
}


This is a relatively recent change; it doesn't happen with the clang shipped
with Xcode based on 3.6.0:

define double @foo(double %a, double %b) #0 {
  %1 = fcmp ogt double %a, 0.000000e+00
  br i1 %1, label %2, label %4

; <label>:2                                       ; preds = %0
  %3 = fdiv double %b, %a
  br label %4

; <label>:4                                       ; preds = %0, %2
  %5 = phi double [ %3, %2 ], [ 0.000000e+00, %0 ]
  ret double %5
}

-- 
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/20150914/97df1c17/attachment.html>