We don't need to figure out the weird extended rounding modes or
handle offsets to keep the lookup table in 64-bits.
Details
Diff Detail
Event Timeline
Code LGTM but I'm not well-versed in this kind of floating point magic so I will leave it up to someone that knows more to approve :)
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | ||
---|---|---|
3639–3641 | nit: inline UseReducedTable? |
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | ||
---|---|---|
3639–3641 | Then you lose the name-as-comment? |
Did you run adequate testing on this (OCL conformance maybe?) ? If so then I can approve it I think
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | ||
---|---|---|
3657–3670 | Does that take a lot of effort to add? Why not add it now? |
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | ||
---|---|---|
3646 | This is just AMDGPU::FltRoundToHWConversionTable & 0xFFFF isn't it? I'm not sure it's worth having all the boilerplate in SIModeRegisterDefaults.cpp to define another table. |
There's no wired up user of this
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | ||
---|---|---|
3657–3670 | There are enough cases to think about in one patch as it is |
nit: inline UseReducedTable?