[llvm-bugs] [Bug 45073] New: false positive: MD_misexpect is not updated after the branch condition is flipped

via llvm-bugs llvm-bugs at lists.llvm.org
Sun Mar 1 21:51:47 PST 2020


https://bugs.llvm.org/show_bug.cgi?id=45073

            Bug ID: 45073
           Summary: false positive: MD_misexpect is not updated after the
                    branch condition is flipped
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Transformation Utilities
          Assignee: unassignedbugs at nondot.org
          Reporter: wmi at google.com
                CC: llvm-bugs at lists.llvm.org

We turned on -Wmisexpect and found some false positive warning. Here is a
testcase to show the problem. The cause is that MD_misexpect is not updated
after the branch condition is flipped. This looks like a general issue of
MD_misexpect. 

---------------- 1.ll --------------------
declare void @goo()
declare i64 @llvm.expect.i64(i64, i64)

define void @foo(i1 %t0) {
  %t1 = xor i1 %t0, true
  %t2 = zext i1 %t1 to i64
  %t4 = call i64 @llvm.expect.i64(i64 %t2, i64 0)
  %t3 = icmp ne i64 %t2, 0
  br i1 %t3, label %true, label %false, !prof !0, !misexpect !1

true:
  call void @goo()
  br label %false

false:
  ret void
}

!0 = !{!"branch_weights", i32 1, i32 2000}
!1 = !{!"misexpect", i64 1, i64 2000, i64 1} ====> misexpect predicts else
branch is hot, which matches with !prof !0.
------------------------------------------

~/workarea/llvm/build/d20200121/bin/opt -instcombine -S 1.ll

Note that the branch condition is flipped after instcombine.

; ModuleID = '1.ll'
source_filename = "1.ll"
declare void @goo()
; Function Attrs: nounwind readnone willreturn
declare i64 @llvm.expect.i64(i64, i64) #0
define void @foo(i1 %t0) {
  br i1 %t0, label %false, label %true, !prof !0, !misexpect !1

true:                                             ; preds = %0
  call void @goo()
  br label %false

false:                                            ; preds = %0, %true
  ret void
}

attributes #0 = { nounwind readnone willreturn }

!0 = !{!"branch_weights", i32 2000, i32 1}
!1 = !{!"misexpect", i64 1, i64 2000, i64 1} ====> misexpect still predicts
else branch is hot, which contradicts with !prof !0.

-- 
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/20200302/9778086a/attachment-0001.html>


More information about the llvm-bugs mailing list