[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH v1 5/7] fpu/softfloat: avoid undefined behaviour when normali
From: |
Richard Henderson |
Subject: |
Re: [PATCH v1 5/7] fpu/softfloat: avoid undefined behaviour when normalising empty sigs |
Date: |
Fri, 27 Mar 2020 15:27:45 -0700 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.4.1 |
On 3/27/20 3:09 AM, Peter Maydell wrote:
> On Fri, 27 Mar 2020 at 09:49, Alex Bennée <address@hidden> wrote:
>>
>> The undefined behaviour checker pointed out that a shift of 64 would
>> lead to undefined behaviour.
>>
>> Signed-off-by: Alex Bennée <address@hidden>
>> ---
>> fpu/softfloat.c | 11 ++++++++---
>> 1 file changed, 8 insertions(+), 3 deletions(-)
>>
>> diff --git a/fpu/softfloat.c b/fpu/softfloat.c
>> index 301ce3b537b..444d35920dd 100644
>> --- a/fpu/softfloat.c
>> +++ b/fpu/softfloat.c
>> @@ -3834,9 +3834,14 @@ void normalizeFloatx80Subnormal(uint64_t aSig,
>> int32_t *zExpPtr,
>> {
>> int8_t shiftCount;
>>
>> - shiftCount = clz64(aSig);
>> - *zSigPtr = aSig<<shiftCount;
>> - *zExpPtr = 1 - shiftCount;
>> + if (aSig) {
>> + shiftCount = clz64(aSig);
>> + *zSigPtr = aSig << shiftCount;
>> + *zExpPtr = 1 - shiftCount;
>> + } else {
>> + *zSigPtr = 0;
>> + *zExpPtr = 1 - 64;
>> + }
>> }
>
> Ah yes, I saw this one in Coverity: CID 1421991.
>
> RTH marked the Coverity issue as a false positive with the rationale
> "We assume an out-of-range shift count is merely IMPLEMENTATION DEFINED
> and not UNDEFINED (in the Arm ARM sense), and so cannot turn a 0 value
> into a non-zero value."
> but I think I disagree with that. We can assume that for the TCG IR
> where we get to define shift semantics because we're doing the codegen,
> but we can't assume it in C code, because it's not included in the set
> of extended guarantees provided by -fwrapv as far as I know.
Perhaps. Of course we also know from our broad knowledge of architectures,
that a compiler would really have to go out of its way for this to happen.
I really hate C in this way, sometimes.
I wonder if I have the energy to petition the committee to drop, for C202? all
of the "undefined" nonsense that only applies to sign-magnitute and
ones-compliment computers, which haven't been seen since the 70's...
> That said, is it valid for this function to be called with a zero
> aSig value ? I think all these normalizeFloat*Subnormal() functions
> assume non-zero sig input, and the only callsite where it's not clearly
> obvious that this is obvious that the sig input is non-zero is the call to
> normalizeFloatx80Subnormal() from addFloatx80Sigs(). So perhaps we
> just need to check and fix that callsite ??
You're right -- addFloatx80Sigs is the only use out of 26 that doesn't have a
preceding check for 0.
r~
- Re: [PATCH v1 3/7] tests/tcg: remove extraneous pasting macros, (continued)
- [PATCH v1 2/7] linux-user: protect fcntl64 with an #ifdef, Alex Bennée, 2020/03/27
- [PATCH v1 4/7] linux-user: more debug for init_guest_space, Alex Bennée, 2020/03/27
- [PATCH v1 6/7] target/xtensa: add FIXME for translation memory leak, Alex Bennée, 2020/03/27
- [PATCH v1 5/7] fpu/softfloat: avoid undefined behaviour when normalising empty sigs, Alex Bennée, 2020/03/27
- Re: [PATCH v1 5/7] fpu/softfloat: avoid undefined behaviour when normalising empty sigs, Aleksandar Markovic, 2020/03/27
- [PATCH v1 7/7] gdbstub: fix compiler complaining, Alex Bennée, 2020/03/27
- Re: [PATCH for 5.0 v1 0/7] A selection of sanitiser fixes, no-reply, 2020/03/27