fpu/softfloat: round_to_int_and_pack refactor broke TriCore ftoi insns

After the refactor from ab52f973a504f8de0c5df64631ba4caea70a7d9e the bahaviour of int32_to_float32() was altered.

helper_ftoi() in target/tricore/fpu_helper.c relied on int32_to_float32 to raise the invalid flag if the input was NaN to properly return 0. Likewise if the input is infinity.

The obvious fix for softfloat would be to raise this flag in round_to_int_and_pack(). However,
I'm not sure if this breaks other targets and I have no easy way to test it.

Yeah it looks like it was missed, the round_to_uint code does it.

Do you have a test case I can verify?

On 04/10/2018 10:07 PM, Alex Bennée wrote:
> Yeah it looks like it was missed, the round_to_uint code does it.
> 
> Do you have a test case I can verify?
> 

For the NaN input 0xffffffff the expected result for the flags is that
flag_invalid is raised.

I can provide you with some TriCore asm, but it is a bit of pain to get
the gnu assembler to build, since the public version is a decade old.

Cheers,
Bastian



Bastian Koppelmann <email address hidden> writes:

> On 04/10/2018 10:07 PM, Alex Bennée wrote:
>> Yeah it looks like it was missed, the round_to_uint code does it.
>>
>> Do you have a test case I can verify?
>>
>
> For the NaN input 0xffffffff the expected result for the flags is that
> flag_invalid is raised.
>
> I can provide you with some TriCore asm, but it is a bit of pain to get
> the gnu assembler to build, since the public version is a decade old.

I'll trust you if you send me a static binary for this particular
verification ;-)

It would be nice to TriCore tests building in tests/tcg/ but I guess we
need an up to date cross compile environment somewhere.

>
> Cheers,
> Bastian


--
Alex Bennée


On 04/11/2018 01:01 PM, Alex Bennée wrote:
> 
> Bastian Koppelmann <email address hidden> writes:
> 
>> On 04/10/2018 10:07 PM, Alex Bennée wrote:
>>> Yeah it looks like it was missed, the round_to_uint code does it.
>>>
>>> Do you have a test case I can verify?
>>>
>>
>> For the NaN input 0xffffffff the expected result for the flags is that
>> flag_invalid is raised.
>>
>> I can provide you with some TriCore asm, but it is a bit of pain to get
>> the gnu assembler to build, since the public version is a decade old.
> 
> I'll trust you if you send me a static binary for this particular
> verification ;-)
> 
> It would be nice to TriCore tests building in tests/tcg/ but I guess we
> need an up to date cross compile environment somewhere.

That is the problem. There is a gcc/binutils port done by some german
company. And it's not easy to get the source. The one they provide is
pretty old and needs some patching to get in building on modern
machines. Right now I'm trying to set up a test environment.

Cheers,
Bastian


On 04/11/2018 01:01 PM, Alex Bennée wrote:
> Bastian Koppelmann <email address hidden> writes:
> 
>> On 04/10/2018 10:07 PM, Alex Bennée wrote:
>>> Yeah it looks like it was missed, the round_to_uint code does it.
>>>
>>> Do you have a test case I can verify?
>>>
>>
>> For the NaN input 0xffffffff the expected result for the flags is that
>> flag_invalid is raised.
>>
>> I can provide you with some TriCore asm, but it is a bit of pain to get
>> the gnu assembler to build, since the public version is a decade old.
> 
> I'll trust you if you send me a static binary for this particular
> verification ;-)

I set up a github repo with working binutils and the corresponding testcase:

https://github.com/bkoppelmann/tricore-fpu

The one caveat is, that we cannot produce any binaries with the TriCore
ISA > 1.3.

In this testcase the last ftoi instruction is supposed to raise the
invalid flag and return 0 since the input was NaN. We did that by only
checking for NaN if any flag was raised after ftoi, then do the NaN
check, and if positive, return 0.

Cheers,
Bastian



Bastian Koppelmann <email address hidden> writes:

> On 04/11/2018 01:01 PM, Alex Bennée wrote:
>> Bastian Koppelmann <email address hidden> writes:
>>
>>> On 04/10/2018 10:07 PM, Alex Bennée wrote:
>>>> Yeah it looks like it was missed, the round_to_uint code does it.
>>>>
>>>> Do you have a test case I can verify?
>>>>
>>>
>>> For the NaN input 0xffffffff the expected result for the flags is that
>>> flag_invalid is raised.
>>>
>>> I can provide you with some TriCore asm, but it is a bit of pain to get
>>> the gnu assembler to build, since the public version is a decade old.
>>
>> I'll trust you if you send me a static binary for this particular
>> verification ;-)
>
> I set up a github repo with working binutils and the corresponding
> testcase:
>
> https://github.com/bkoppelmann/tricore-fpu
>
> The one caveat is, that we cannot produce any binaries with the TriCore
> ISA > 1.3.
>
> In this testcase the last ftoi instruction is supposed to raise the
> invalid flag and return 0 since the input was NaN. We did that by only
> checking for NaN if any flag was raised after ftoi, then do the NaN
> check, and if positive, return 0.

Well it builds and I get an fpu-test.elf but I'm a bit stuck on how to
run it. What are the runes for launching the test?

>
> Cheers,
> Bastian


--
Alex Bennée


On 04/12/2018 03:41 PM, Alex Bennée wrote:
> Bastian Koppelmann <email address hidden> writes:
> 
>> On 04/11/2018 01:01 PM, Alex Bennée wrote:
>>> Bastian Koppelmann <email address hidden> writes:
>>>
>>>> On 04/10/2018 10:07 PM, Alex Bennée wrote:
>>>>> Yeah it looks like it was missed, the round_to_uint code does it.
>>>>>
>>>>> Do you have a test case I can verify?
>>>>>
>>>>
>>>> For the NaN input 0xffffffff the expected result for the flags is that
>>>> flag_invalid is raised.
>>>>
>>>> I can provide you with some TriCore asm, but it is a bit of pain to get
>>>> the gnu assembler to build, since the public version is a decade old.
>>>
>>> I'll trust you if you send me a static binary for this particular
>>> verification ;-)
>>
>> I set up a github repo with working binutils and the corresponding
>> testcase:
>>
>> https://github.com/bkoppelmann/tricore-fpu
>>
>> The one caveat is, that we cannot produce any binaries with the TriCore
>> ISA > 1.3.
>>
>> In this testcase the last ftoi instruction is supposed to raise the
>> invalid flag and return 0 since the input was NaN. We did that by only
>> checking for NaN if any flag was raised after ftoi, then do the NaN
>> check, and if positive, return 0.
> 
> Well it builds and I get an fpu-test.elf but I'm a bit stuck on how to
> run it. What are the runes for launching the test?

qemu-system-tricore -M tricore_testboard -kernel fpu-test.elf -nographic

However, this is not really a test. I usually run these against
Infineons golden (sadly proprietary) simulator and compare register
dumps. I have some more regression tests but I haven't had the time make
them into proper tests, such that we don't need this golden model.

Cheers,
Bastian