Optimizations appear to malfunction when encountering setjmp

[VWoeltjen]

First, some background: I am using XMLVM (http://xmlvm.org) to cross-compile from Java to C, and Emscripten to cross-compile from C to JavaScript. XMLVM converts Java's try-catch blocks to if(setjmp(...))-else blocks, and converts throw to longjmp(...).

When cross-compiling with Emscripten without optimizations, everything works as expected (including in cases where Java exceptions are thrown - these are handled as if by appropriate catch blocks).

However, once optimizations (-O1 or -O2) are enabled, problems emerge. These have varied depending on the specific sources being cross-compiled; results include infinite loops and crashes due the execution of unexpected code paths. Notably, this occurs without any Java exceptions being thrown (that is, before any invocations of longjmp occur)

A workaround is to replace all "setjmp(...)" calls with explicit 0 values (this turns out to be easy, as XMLVM uses an XMLVM_SETJMP(...) macro). In this case, code behaves as expected even under -O1 or -O2, so long as no Java exceptions (in C, longjmp) occur. However, this effectively upgrades all exceptions to critical failures

I have prepared a minimal test case; this is a cross-compiled minimal Java application (the original Java program does nothing but create an object; I have added a printf("Hello!") to the resulting C sources to provide some output). Due to size, I have hosted this as a git repository:

https://github.com/VWoeltjen/hello-xmlvm-emscripten.git

Note that there are two tags - "desired" (which uses setjmp) and "workaround" (which replaces all setjmps with 0, effectively disabling Java exception-handling).

There is a Makefile in the dist/ folder that is already set up to use emcc, with -O1 enabled. This will produce a file "dist/build/hello.js"; running with a JavaScript interpreter (I have been using node) should print "Hello!" to the console, but falls into an infinite loop when using -O1 on the "desired" branch. (You may need to manually clean -remove the build directory - when switching between tags, if make reports no changes)

(Apologies for the size of the test case, but I haven't been able to reproduce this in a smaller form. The cross-compiled sources above include close to 500 uses of setjmp, so it may be that the issue only occurs at a certain complexity.)

Some notes:

• Cross-compiling with clang does not exhibit this problem, even when using -O1 on the "desired" branch
• As an experiment, I tried changing the XMLVM_SETJMP macro to check if the system time was 0 (that is, to something that can't be determined at compile time, but will always be false in practice). This did not reproduce the issue under -O1 (I wanted to rule out the possibility that explicitly setting this macro to 0 was fixing the issue because half the if branch could be pruned at compile time)

[kripken]

I can't find a straightforward way to automate the debugging of this, I'm afraid. The best course of action looks like bisecting on the uses of XMLVM_SETJMP, that is, change more and more into 0 until we find the single change that makes a difference in the output. There are two appearances in .c files, changing those makes no difference. So it looks like bisection would need to be done on XMLVM_TRY_BEGIN which uses the former macro. It appears in 43 files, 492 appearances. Which sounds like a lot, but bisection and search and replace could get us to an answer without too much pain I hope. Unless there is a better way to change how these macros are used?

[waywardmonkeys]

Looking at the code, this looks like it might well be the classic problem of not having the right things marked as volatile (and therefore a bug in XMLVM)?

See http://pubs.opengroup.org/onlinepubs/7908799/xsh/setjmp.html for details.

[VWoeltjen]

Thanks waywardmonkeys! Marking variables as volatile does indeed make the issue go away. This can be fixed on the XMLVM side.

[kripken]

Great, closing this.