[SR-4497] Memory issue causes crash when application built with optimization turned on.

[swift-ci]

Previous ID	SR-4497
Radar	rdar://problem/31459425
Original Reporter	anaglik (JIRA User)
Type	Bug
Status	Resolved
Resolution	Duplicate

Attachment: Download

Additional Detail from JIRA

Votes	0
Component/s	Compiler
Labels	Bug, 3.1Regression, OptimizedOnly, RunTimeCrash
Assignee	@eeckstein
Priority	Medium

md5: baaae3d56f01cbf35295f4840f3516f5

duplicates:

• SR-4393 Miscompilation: object modified after being freed (Swift 3.1 regression)

Issue Description:

I've noticed crashes after I upgraded Swift to version 3.1. It happens only when optimization is set to "-O" or "-O whole-module-optimization"
Here is an example code (Example app crashes when following method is called a few times (usually two is enough).

fileprivate func funWithWords(_ words: [String]) -> [String] {        
        var resultArray = [String]()
        
        for index in 0..<words.count {
            // This will crash
            let pre = words[0...index].reduce("", +)
            resultArray.append(pre)
            
            // This version is ok
//            let pre = words[0...index].reduce("", +)
//            resultArray = resultArray + [pre]
            
            // This version is ok too
//            let pre = words[0...index].joined()
//            resultArray.append(pre)
            
            // This version is ok too
//            let pre = words[0...index]
//            let reduced = pre.reduce("", +)
//            resultArray.append(reduced)
        }
        return resultArray
}

Instruments also detects leak in that code. I am attaching screenshot of stack trace from Instruments and Xcode.
I will verify if issue exists on swift-master as well.

[belkadan]

I tried it and got a miscompile—for a particular call it returned the wrong results. It seems to be working in master, though.

@eeckstein, does this sound familiar? Was there a 'semantics' attribute in the wrong place in 3.1?

[belkadan]

@swift-ci create

[eeckstein]

It's not working on master either. Although the behavior is different.
I'll look at it.

[bob-wilson]

This looks like the same thing as SR-4393.

[atrick]

I verified this is fixed by

commit 2fce87c
Author: Erik Eckstein <eeckstein@apple.com>
Date: Wed Apr 5 17:02:29 2017

EscapeAnalysis: fix a wrong use-point detection.

This caused redundant load elimination to remove a load although the value is overwritten in a called function.
Most likely this could only occur if the load address is a block argument.

fixes SR-4393