[PATCH] D98002: [RISCV] Add scheduling resources for V

[Evandro Menezes via Phabricator via llvm-commits]

evandro added inline comments.


================
Comment at: llvm/lib/Target/RISCV/RISCVInstrInfoV.td:593
+  def _VS : VALUVV<funct6, OPMVV, opcodestr # ".vs">,
+            Sched<[WriteVIRedV, ReadVIRedV, ReadVIRedV, ReadVMask]>;
+}
----------------
frasercrmck wrote:
> I dunno if the first source of a reduction is likely to benefit from a different scheduling type than the second, since it's only reading the first element?
I don't know either.  It would probably be used with `ReadAdvance`.  


================
Comment at: llvm/lib/Target/RISCV/RISCVInstrInfoV.td:1178
 let Constraints = "@earlyclobber $vd" in {
-defm VFNCVT_XU_F_W : VALU_FV_VS2<"vfncvt.xu.f.w", 0b010010, 0b10000>;
-defm VFNCVT_X_F_W : VALU_FV_VS2<"vfncvt.x.f.w", 0b010010, 0b10001>;
-defm VFNCVT_RTZ_XU_F_W : VALU_FV_VS2<"vfncvt.rtz.xu.f.w", 0b010010, 0b10110>;
-defm VFNCVT_RTZ_X_F_W : VALU_FV_VS2<"vfncvt.rtz.x.f.w", 0b010010, 0b10111>;
-defm VFNCVT_F_XU_W : VALU_FV_VS2<"vfncvt.f.xu.w", 0b010010, 0b10010>;
-defm VFNCVT_F_X_W : VALU_FV_VS2<"vfncvt.f.x.w", 0b010010, 0b10011>;
-defm VFNCVT_F_F_W : VALU_FV_VS2<"vfncvt.f.f.w", 0b010010, 0b10100>;
-defm VFNCVT_ROD_F_F_W : VALU_FV_VS2<"vfncvt.rod.f.f.w", 0b010010, 0b10101>;
+defm VFNCVT_XU_F_W : VNCVTI_FV_VS2<"vfncvt.xu.f.w", 0b010010, 0b10000>;
+defm VFNCVT_X_F_W : VNCVTI_FV_VS2<"vfncvt.x.f.w", 0b010010, 0b10001>;
----------------
HsiangKai wrote:
> craig.topper wrote:
> > HsiangKai wrote:
> > > evandro wrote:
> > > > frasercrmck wrote:
> > > > > It might not really be important but it's not immediately obvious to me what some of these abbreviations signify. `VNCVTF` narrow-converts int to float but `VNCVTD` narrow-converts float to float? Why F and D?
> > > > Narrowing conversion from `double` to `float` or widening from `float` to `double`.
> > > ```
> > > vfncvt.f.x.w  vd, vs2, vm       # Convert double-width signed integer to float.
> > > vfncvt.f.f.w vd, vs2, vm        # Convert double-width float to single-width float.
> > > ```
> > > These two are both to convert to single-width float. Maybe use `DI`, `IF` and `DF` to distinguish them? 
> > vncvt.f.f.w is also used for float to half. So using something that stands for "double" and "float" could be misleading.
> I agree. It is misleading. How about use `F_X` and `F_F`  just like the instruction name.
> Besides the naming issue, the patch looks good to me.
The general scheme above is `V{,W,N}CVT{I,F}_{I,F}V_VS2`, where `{,W,N}` denotes the width of the result, whether the same, wider or narrower, and `{I,F}`, the type of the result, in the first instance, or of the argument, in the second instance, integer or FP.  


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D98002/new/

https://reviews.llvm.org/D98002