I noted a number of code generation differences between various gcc versions for aarch64, but there is something different going on. When I copied VM binaries produced on the Chromebook Plus to Raspberry Pi 3 (ArchLinuxArm), the same 2 unit tests fail in Cuis! So I am now guessing there must be a shared library difference (perhaps libm ?) Having the same 2 tests fail on x86_64 would preclude a bad CPU chip. ??? -KenD
Hi Ken,
On Wed, Feb 13, 2019 at 1:58 PM ken.dickey@whidbey.com wrote:
I noted a number of code generation differences between various gcc versions for aarch64, but there is something different going on. When I copied VM binaries produced on the Chromebook Plus to Raspberry Pi 3 (ArchLinuxArm), the same 2 unit tests fail in Cuis! So I am now guessing there must be a shared library difference (perhaps libm ?) Having the same 2 tests fail on x86_64 would preclude a bad CPU chip.
What are the failing tests? If they are comparing float values could it simply be a precision error that requires inexact float comparison? (e.g. (a - b) abs < epsilon)
??? -KenD
_,,,^..^,,,_ best, Eliot
On 2019-02-13 18:57, Eliot Miranda wrote:
What are the failing tests? If they are comparing float values could it simply be a precision error that requires inexact float comparison? (e.g. (a - b) abs < epsilon)
Ah! Yes. That nailed it! Thinking. Gotta try that next time! Radical approach to the solution, however. ;^)
Thanks for all, -KenD
Something doesn't look right there. The test in question appears to be adding all integers from 1 to 9 as floating point numbers, and seems to be expecting the result is 45 as a floating point number. That calculation should need no epsilons, it's lossless in floating point land (under the usual assumptions). What's going on, exactly?
On 2/14/19 09:17, ken.dickey@whidbey.com wrote:
On 2019-02-13 18:57, Eliot Miranda wrote:
What are the failing tests? If they are comparing float values could it simply be a precision error that requires inexact float comparison? (e.g. (a - b) abs < epsilon)
Ah! Yes. That nailed it! Thinking. Gotta try that next time! Radical approach to the solution, however. ;^)
Thanks for all, -KenD
On 2019-02-14 16:39, Andres Valloud wrote:
Something doesn't look right there. The test in question appears to be adding all integers from 1 to 9 as floating point numbers, and seems to be expecting the result is 45 as a floating point number. That calculation should need no epsilons, it's lossless in floating point land (under the usual assumptions). What's going on, exactly?
I am clueless. Different compiler versions. Different silocon. All gcc+Linux. All "git cloned" from same https://github.com/KenDickey/opensmalltalk-vm (cloned from OPenSmalltalk w minor changes).
Vanilla (epsilon-less) test: - Failed on Raspberry Pi 3 (Broadcom aarch64) and older x86_64. - Passed on aarch64 Samsung Chromebook Plus (Rockchip) and LePotato (AMLogic). I have stepped through the call to ffiTestDoubles9 with gdb, which returns float(45.) but get lost in the various alloc,copy,shift,mask mechanics as things unwind in the interpreter vm.
It is beyond my poor shrinking brain.. -KenD
On 2/15/2019 1:55 PM, ken.dickey@whidbey.com wrote:
On 2019-02-14 16:39, Andres Valloud wrote:
Something doesn't look right there. The test in question appears to be adding all integers from 1 to 9 as floating point numbers, and seems to be expecting the result is 45 as a floating point number. That calculation should need no epsilons, it's lossless in floating point land (under the usual assumptions). What's going on, exactly?
I am clueless. Different compiler versions. Different silocon. All gcc+Linux. All "git cloned" from same https://github.com/KenDickey/opensmalltalk-vm (cloned from OPenSmalltalk w minor changes).
Vanilla (epsilon-less) test:
- Failed on Raspberry Pi 3 (Broadcom aarch64) and older x86_64.
- Passed on aarch64 Samsung Chromebook Plus (Rockchip) and LePotato
(AMLogic). I have stepped through the call to ffiTestDoubles9 with gdb, which returns float(45.) but get lost in the various alloc,copy,shift,mask mechanics as things unwind in the interpreter vm.
It is beyond my poor shrinking brain.. -KenD
When it fails, what are the 8 bytes you get in gdb? What are the 8 bytes you get in Cuis (in the Cuis debugger, at the moment of the failure).
On 2019-02-17 05:44, Juan Vuletich wrote:
When it fails, what are the 8 bytes you get in gdb? What are the 8 bytes you get in Cuis (in the Cuis debugger, at the moment of the failure).
Hi Juan, happen to be on old amd x86_64 Linux box right now.
If I set a gdb breakpoint at ffiTestDoubles9 and step through it, I end up in src/plugins/SqueakFFIPrims/IA32FFIPlugin.c
result2 = floatObjectOf(floatRet1);
At this point floatRet1 = 0x4046800000000000 after step result2 = 0x846800000000004
which looks right to me.
Continuing, the result is Smalltalk is SmallInteger 0.0 !
Result seems to get lost somewhere.
I have my RasPi unplugged at the moment.. Have to get back w result there.
HTH, -KenD
vm-dev@lists.squeakfoundation.org
-
Andres Valloud -
Eliot Miranda -
Juan Vuletich -
ken.dickey＠whidbey.com