On Fri, 3 Apr 2015, Maciej W. Rozycki wrote: > The ABS.fmt and NEG.fmt instructions have been specified as arithmetic > in the MIPS architecture, which in particular implies handling NaN data > in the usual way with qNaN bit patterns propagated unchanged and sNaN > bit patterns signalling the usual IEEE 754 Invalid Operation exception > and quieted by default. > > A series of changes applied over time to our implementation: > > c5033d78 [MIPS] ieee754[sd]p_neg workaround > cea2be44 MIPS: Fix abs.[sd] and neg.[sd] emulation for NaN operands > > has led to the current situation where the sign bit is updated according > to the operation requested even for NaN inputs. This is according to > these commits a workaround so that broken binaries produced by GCC > disregarding the properties of these instructions have a chance to work. > > For sNaN inputs this remains within IEEE Std 754 as the standard leaves > the choice of output qNaN bit patterns produced under the default > Invalid Operation exception handling for individual sNaN input bit > patterns to implementer's discretion, even though it still recommends as > much NaN input information to be preserved in NaN outputs. > > For qNaN inputs however it violates the standard as it requires a qNaN > input bit patterns to propagate unchanged to output. > > This is also unlike real MIPS FPU hardware behaves where sNaN and/or > qNaN processing has been fully implemented with no Unimplemented > Operation exception signalled. Such hardware propagates any input qNaN > bit pattern unchanged. It also quiets any input sNaN bit pattern in an > implementer-specific manner, for example the MIPS 74Kf processor returns > the default qNaN pattern with the sign bit always clear and the Broadcom > SB-1 and BMIPS5000 processors propagate the input sNaN bit pattern with > the sign bit unchanged and the quiet bit first cleared in the trailing > significand field and then the next lower bit set if clearing the quiet > bit left the field with no other bit set. > > Especially the latter observation indicates the limited usefulness of > the workaround as it will cover many hardware configurations, but not > all of them, only making it harder to discover such broken binaries that > need to be recompiled with GCC told to avoid the use of ABS.fmt and > NEG.fmt instructions where non-arithmetic semantics is required by the > algorithm used. > > Revert the damage done by the series of changes then, and take the > opportunity to simplify implementation by calling `ieee754dp_sub' and > `ieee754dp_add' as required and also the rounding mode set towards -Inf > temporarily so that the sign of 0 is correctly handled. > > Signed-off-by: Maciej W. Rozycki <macro@xxxxxxxxxxxxxx> > --- One point to make here is the use of `ieee754dp_sub' and `ieee754dp_add' makes emulated ABS.fmt and NEG.fmt respect FCSR.FS for denormals just as hardware does. I should have noted that in the commit message, perhaps it can be retrofitted? Maciej
