[LLVMdev] Floating point instructions patch
sabre at nondot.org
Thu Apr 28 15:21:58 PDT 2005
On Thu, 28 Apr 2005, Morten Ofstad wrote:
> I have been gone for a while, finishing work on my Master's thesis...
Hi Morten, congrats! :)
> Now that I'm back I updated LLVM to the most recent version and found
> that my FP_ABS SelectionDAGNode type and code generation was now
> conflicting with the new FABS node type. I brought the rest of my local
> modifications in line with the FABS implementation, so here is my patch
> that includes sqrt, sin and cos also. The only thing missing is
> expansion to lib calls for these node types for targets that don't
> support them... I'm sure someone who understands the LLVM internals a
> bit better than me can add that in no time ;) Here is the patch, hope to
> see it applied soon!
Ok, overall it looks good. I've applied several pieces of it here:
Add and legalize new nodes:
Add fabs/fabsf support to x86 isel simple:
Add fsqrt support to x86 pattern isel:
New X86 instrs:
The patches I didn't apply are these:
1. Match (Y < 0) ? -Y : Y -> FABS in the SelectionDAGISel.cpp file. We
already catch this at the DAG level. If we aren't, please let me know.
2. Codegen fsin/fcos to fsin/fcos for X86. We cannot do this, except
under the control of something like -enable-unsafe-fp-math. These
instructions are architected to have a limited range.
3. Codegen fsqrt/fsqrtf C functions to the FSQRT dag node. These
functions can set errno, so this is not a safe transformation. The
proper way to do this is to introduce an llvm.sqrt.f32/llvm.sqrt.f64
pair of intrinsics to represent these operations without errno. The
optimizer can then turn fsqrt calls into these intrinsics when errno is
provably never used or if a compiler option is specified to ignore
errno. Also, your work could just generate llvm.sqrt.* calls
4. Codegen of fsin/fcos libm calls to FSIN/FCOS dag nodes. We really do
need to be able to legalize these for targets that don't support them
before we can do this. The easiest way to do this is to check the
TargetLowering information for the operation to see if they are
supported. If not, leave them as function calls in the
SelectionDAGISel.cpp file. The alternative way is to have legalize
lower them to libcalls. This is also needed for fsqrt, though sin/cos
don't set errno.
I hope that committing the patches I did will significantly help your
merging. To get the rest in, here are what needs to be done:
For #1, verify that it is still needed.
For #2, you can choose to add a flag to llvm/Target/TargetOptions.h named
-enable-unsafe-fp-math. This would default to off, but you can obviously
set it to on for your environment. This could also be used to do a lot of
other simplifications and optimizations in the code generator.
For #3, just add the llvm intrinsics.
For #4, implement it or find someone to do it :)
Please lemme know if you have any questions, thanks for the patches!
> I also noticed that the DoesntAccessMemoryTable in BasicAliasAnalysis.cpp
> includes "sinh", "cosh" and friends -- as far as I know, these can actually
> set errno so they should not be in the table...
You're right, I should not believe man pages. Fixed, thanks!
More information about the llvm-dev