583731 - misc/testJitOrDie testcase is causing an assert in nanojit::LInsPrinter::formatRef()

Status: VERIFIED FIXED

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect)

Description

[Brent Baker]

Attachment: gdb stacktrace

gdb $AVM misc/testJitordieSwitch.abc

Reason: KERN_INVALID_ADDRESS at address: 0x0000000000000018
0x000000010014b7fb in nanojit::LInsPrinter::formatRef (this=0x0, buf=0x7fff5fbfd3f0, ref=0x101267e40, showImmValue=true) at /Users/brbaker/hg/tamarin-redux/nanojit/LIR.cpp:1680
1680	        const char* name = lirNameMap->lookupName(ref);

Comment 1

[Brent Baker]

This failure was injected in recent nanojit merge to d8dccd548db136e88ea1a8be0627088b6a78cc78

Testcase passed in tr rev# 4943 and then failed in the next build after the merge tr# 4956

Comment 2

[Brent Baker]

Marking in-testsuite? so that the testcase is re-enabled in the deep phase once this issue is resolved. Testcase was removed from deep testing in tamarin-redux changeset 4977:e3a2c970a774

Comment 3

[Rick Reitmaier]

Attachment: v1

This only occurs in debug builds correct?

The problem is that ValidateWriter is trying to complain (i.e. print) about an AccSet (non)issue.  

accSet is set to VARS but CseFilter decided to split the store into 2 instructions since disp > 16bits resulting in a different value of 'base' than what ValidateWriter is checking against.

Patch teaches ValidateWriter.checkAccSet how to peek under the hood for a LIR_addp.

Comment 4

[Edwin Smith]

Comment on attachment 468534 [details] [diff] [review]
v1

Logically fine, but would be better to use && instead of a conditional:

bool isVars = (base == vars) || 
              (base->opcode() == LIR_addp && base->oprnd1() == vars);

R+ with that fixed.

Comment 5

[Brent Baker]

(In reply to comment #3)
> Created attachment 468534 [details] [diff] [review]
> v1
> 
> This only occurs in debug builds correct?
 Yes only seen on debug builds, nondebug builds are still executing cleanly.

Comment 6

[Rick Reitmaier]

http://hg.mozilla.org/tamarin-redux/rev/467cdedf877d

Comment 7

[Brent Baker]

Re-enable testcase misc/testJitordieSwitch testcase in changeset 5112:929bfbe908a9