411843 - ESC line numbers incorrect

Status: RESOLVED FIXED

(Tamarin Graveyard :: Self-hosting compiler (ESC), defect)

Description

[Mark Hammond [:markh] [:mhammond]]

ESC sometimes generates OP_debugline instructions in the wrong place.

To reproduce, create a sourcefile containing 'x=y\n'.  Compile it using esc.sh and encode it using abcencode.sh.  The 'code' in the resulting .asm file shows:

                    , 'code': [ [ 'getlocal0' ]
                              , [ 'pushscope' ]
                              , [ 'debugfile', 5 ]
                              , [ 'debugline', 1 ]
                              , [ 'findproperty', 4 ]
                              , [ 'debugline', 2 ]
                              , [ 'findpropstrict', 5 ]
                              , [ 'getproperty', 5 ]

Note the second 'debugline' instruction - it is indicating we are on line 2 way too early - that first 'findproperty' is locating the 'x' - assignment doesn't finish until 6 or 7 instructions later.

Comment 1

[Erik Tierney]

Attachment: Fix the bug

ESC was using the line number for the wrong token.  And then it was mistakenly adding 1 to it.  The first problem ended up with the correct line number sometimes, since often the next token was on the same line.  The second problem made everything off by one.

Comment 2

[Erik Tierney]

Pushed fix to tc.

Comment 3

[Mark Hammond [:markh] [:mhammond]]

works for me - thanks!

Comment 4

[Mark Hammond [:markh] [:mhammond]]

There appears to have been a regression here - the same sample above appears to generate identical asm as shown above - ie, a premature debugline instruction.

Comment 5

[Lars T Hansen]

Attachment: Patches for the regression (and for similar, unfixed bugs)

The problem was that the token was consumed before its source position was recorded.  This happened all over the place and I've gone over the parser and fixed the obvious ones.

Comment 6

[Mark Hammond [:markh] [:mhammond]]

Comment on attachment 302524 [details] [diff] [review]
Patches for the regression (and for similar, unfixed bugs)

looks good and works for me.

Comment 7

[Lars T Hansen]

378:e5dc4ee7b1ff

Comment 8

[Mark Hammond [:markh] [:mhammond]]

This seems to have come back (at rev 548).  In the example above, the runtime error seen is:

  ReferenceError: Error #1065: Variable y is not defined.
          at global$init()[hello.as:0]

You can add as many lines at the start as you like, and the line will continue to be reported as 0.  If you put a successful line first, the error will then be reported as being at the line which worked.  Sadly, the abc encoder is currently broken so I can't paste in the assembler.

Comment 9

[Mark Hammond [:markh] [:mhammond]]

I may have misled - rev 548 is what I happen to be at and is only a couple of days old.  There is nothing to suggest that rev was responsible for the problem though.

Comment 10

[Lars T Hansen]

Attachment: Patches the new regression by recording sourcepos more often

Comment 11

[Lars T Hansen]

Pushed.