search for: d60261

...ops, where chaining happens through the reduction value's PHI node (i.e. the scalar reduction value from one iteration will be the input to the next iteration). This intrinsic would also naturally match reduction instructions of ARM SVE and NEC SX-aurora. For Option A (https://reviews.llvm.org/D60261), there is an argument that code creating or operating on these intrinsics can treat ordered and unordered reductions the same (in that they have the same arguments). Fast-math flags determine whether or not the intrinsic needs to be evaluated in strict order. Codegen for non-strict reductions shou...

...lso propose renaming the 'accumulator' operand to 'start value' because for fmul this is the start value of the reduction, rather than a value to which the fmul reduction is accumulated into. [Option A] Always using the start value operand in the reduction (https://reviews.llvm.org/D60261) declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %start_value, <4 x float> %vec) This means that if the start value is 'undef', the result will be undef and all code creating such a reduction will need to ensure it has a sensible start value (e.g. 0.0 for fa...

...scope. But yes we can easily rename the intrinsics again when the VP proposal lands. > > Btw, if you are at EuroLLVM. There is a BoF at 2pm today on LLVM-VP. > >>>> >>>> [Option A] Always using the start value operand in the reduction (https://reviews.llvm.org/D60261) >>>> >>>> declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float %start_value, <4 x float> %vec) >>>> >>>> This means that if the start value is 'undef', the result will be undef and all code creating such a reduct...

...operand to 'start value' because >> for fmul this is the start value of the reduction, rather than a >> value to which the fmul reduction is accumulated into. >> >> [Option A] Always using the start value operand in the reduction >> (https://reviews.llvm.org/D60261) >> >>   declare float >> @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float >> %start_value, <4 x float> %vec) >> >> This means that if the start value is 'undef', the result will be >> undef and all code creating such a reduction wi...