462010 - MOPS/FLACC failures when run on PPC

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Brent Baker]

Possibly an endian issue.

$> avmshell tests/acceptance/mops/mops.abc_

The following testcases fail:

Read test int w/ min sized = false FAILED! expected: true
Read test float w/ min sized = false FAILED! expected: true
Read test double w/ min sized = false FAILED! expected: true
Write test float w/ min sized = false FAILED! expected: true
Write test double w/ min sized = false FAILED! expected: true
Read/write int min = false FAILED! expected: true
Read/write float min = false FAILED! expected: true
Read/write double min = false FAILED! expected: true
Read/write int scratch = false FAILED! expected: true
Read/write float scratch = false FAILED! expected: true
Read/write double scratch = false FAILED! expected: true

Comment 1

[Brent Baker]

Pushed change to redux to skip the testcase (1031:82929bd5b049), needs to be
removed when this is fixed.

Comment 2

[Steven Johnson]

Is this running with -Dinterp, or without, or both?

Comment 3

[Steven Johnson]

Answer: it fails both ways.

(handy tip I just learned: you can force an Intel Mac to run a universal binary in PPC mode, via rosetta, by preceding command line with "arch -ppc").

Comment 4

[Steven Johnson]

Doh. Looks like the byte-swap functions for big-endian are just plain wrong for 32-bit and 64-bit.

Comment 5

[Steven Johnson]

Attachment: Patch

big-endian byte-swap routines were just plain wrong for 32-bit and 64-bit values in the MOPS opcodes. Also, a couple of the interpreter cases were incorrect as well (but only for big-endian). Consolidated the swap functions from MethodEnv.cpp and Interpreter.cpp into a single set, added strong typing to make some of the errors harder to make. Now passes the acceptance test for both Intel and PPC, in both interp and jit modes.

Comment 6

[Rick Reitmaier]

Comment on attachment 345329 [details] [diff] [review]
Patch

Probably not that critical but have you looked into using the shift/and/or version of swapping.  It's usually faster than byte extract.

Comment 7

[Steven Johnson]

Nope -- I mimicked what we were already doing in DataIO.cpp. Definitely open for improvements :-)

Comment 8

[Steven Johnson]

pushed to redux as changeset:   1038:c071833f0116

Comment 9

[Chris Peyer]

Resolved fixed engineering / work item that has been pushed.  Setting status to verified.

Comment 10

[Brent Baker]

removing QRB request, bug resolved/verified