cmd/go-post: "unresolved: erroneous handling of global variables"

[mewmew]

From @karlek on May 29, 2015 23:13

// a.go
package main

func main() {
    // go-post wants to change this to:
    // dx := 0
    // which shadows the global variable dx.
    dx = 0
}

// b.go
package main

var dx = 1

diff a.go fixed/a.go
--- /tmp/go-fix673596466    2015-05-30 00:41:51.888890332 +0200
+++ /tmp/go-fix613483497    2015-05-30 00:41:51.888890332 +0200
@@ -4,5 +4,5 @@
    // go-post wants to change this to:
    // dx := 0
    // which shadows the global variable dx.
-   dx = 0
+   dx := 0
 }

Copied from original issue: #2

[mewmew]

This issue will be solved through proper data-flow analysis, taking part in the middle-end module of the decompiler pipeline. The data flow analysis will be developed, similar to the control flow analysis component, as a stand-alone component which analyses the data dependencies between instructions, and outputs these dependencies in JSON format, for the back-end to make use of. As such, the back-end will have all information required to declare variables in their correct scope, and this will make is possible to remove the unresolved source code rewrite rule of go-post.

Several of the go-post rewrite rules are currently there as a poor man's data flow analysis substitute. Once proper data flow analysis is in place, the following rewrite rules will be removed from go-post:

• deadassign
  • Remove "x = x" assignments
• localid
  • Replace the use of local variable IDs with their definition.
• unresolved
  • Replace assignment statements with declare and initialize statements at the first occurance of an unresolved identifier.

More specifically, the following rewrite rules are likely to remain, after both control flow analysis and data flow analysis reach maturity.

• assignbinop
  • Replace "x = x + z" with "x += z".
• forloop
  • Add initialization and post-statements to for-loops.
• mainret
  • Replace return statements with calls to os.Exit in the "main" function.

Proper control flow analysis will remove the need for the deadlabel (i,e. Remove unused labels.) rewrite rule.

Updating the milestone of this issue to 0.5, as that release is intended to focus on data flow analysis. For a rough time frame, the intention is to start looking into data flow analysis for next year (2018). The autumn of 2017 is intended focus on control flow analysis (i.e. release 0.4). Release 0.2 is basically what is in the master branch right now. It will be tagged as such after some more testing (and potentially support for DWARF metadata in llir/llvm). Release 0.3 may be retargetted to happen after 0.5, depending on what is most fun to play with at the moment :)