[BasicTTI] Set scalarization cost of scalable vector casts to Invalid.

[sdesmalen]

Summary

When BasicTTIImpl::getCastInstrCost can't determine the cost of a
vector cast operation when the types need legalization, it falls
back to calculating scalarization costs. Instead of crashing on
cast<FixedVectorType>(DstVTy) when the type is a scalable vector,
return an Invalid cost.

[david-arm]

llvm/test/Analysis/CostModel/AArch64/sve-fptoi.ll

20

It seems a little inconsistent that we return a cost of one for converting <vscale x 1 x double>, whereas it's invalid for <vscale x 1 x float>. Not sure if we need to worry about this?

[sdesmalen]

llvm/test/Analysis/CostModel/AArch64/sve-fptoi.ll

20

Yes, it's inconsistent, but I don't think it's really something to worry about now. For all legal types or too-wide types (that require splitting) we should have sensible costs, whereas VFs with vscale x 1 types are currently avoided for SVE.

[david-arm]

LGTM! Happy to revisit the inconsistency another time.

llvm/include/llvm/CodeGen/BasicTTIImpl.h

if (DstVTy && SrcVTy) {
         // If both types need to be split then the split is free.
         InstructionCost SplitCost =
             (!SplitSrc || !SplitDst) ? TTI->getVectorSplitCost() : 0;
         return SplitCost +
                (2 * TTI->getCastInstrCost(Opcode, SplitDstTy, SplitSrcTy, CCH,
                                           CostKind, I));
       }
 
+      // Scalarization cost is Invalid, can't assume any num elements.
+      if (isa<ScalableVectorType>(DstVTy))
+        return InstructionCost::getInvalid();
+
       // In other cases where the source or destination are illegal, assume
       // the operation will get scalarized.
       unsigned Num = cast<FixedVectorType>(DstVTy)->getNumElements();
       InstructionCost Cost = thisT()->getCastInstrCost(
           Opcode, Dst->getScalarType(), Src->getScalarType(), CCH, CostKind, I);
 
       // Return the cost of multiple scalar invocation plus the cost of
       // inserting and extracting the values.

llvm/test/Analysis/CostModel/AArch64/sve-fptoi.ll

 ; RUN: opt -cost-model -analyze -mtriple aarch64-linux-gnu -mattr=+sve -o - -S < %s | FileCheck %s
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64-unknown-linux-gnu"
 
 define void @sve-fptoi() {
 ; CHECK-LABEL: 'sve-fptoi'
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_si8 = fptosi <vscale x 1 x half> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_ui8 = fptoui <vscale x 1 x half> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_si32 = fptosi <vscale x 1 x half> undef to <vscale x 1 x i32>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_ui32 = fptoui <vscale x 1 x half> undef to <vscale x 1 x i32>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_si64 = fptosi <vscale x 1 x half> undef to <vscale x 1 x i64>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f16_to_ui64 = fptoui <vscale x 1 x half> undef to <vscale x 1 x i64>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_si8 = fptosi <vscale x 1 x float> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_ui8 = fptoui <vscale x 1 x float> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_si16 = fptosi <vscale x 1 x float> undef to <vscale x 1 x i16>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_ui16 = fptoui <vscale x 1 x float> undef to <vscale x 1 x i16>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_si64 = fptosi <vscale x 1 x float> undef to <vscale x 1 x i64>
+; CHECK-NEXT: Cost Model: Invalid cost for instruction: %nv1f32_to_ui64 = fptoui <vscale x 1 x float> undef to <vscale x 1 x i64>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_si8 = fptosi <vscale x 1 x double> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_ui8 = fptoui <vscale x 1 x double> undef to <vscale x 1 x i8>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_si16 = fptosi <vscale x 1 x double> undef to <vscale x 1 x i16>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_ui16 = fptoui <vscale x 1 x double> undef to <vscale x 1 x i16>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_si32 = fptosi <vscale x 1 x double> undef to <vscale x 1 x i32>
+; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv1f64_to_ui32 = fptoui <vscale x 1 x double> undef to <vscale x 1 x i32>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_si8 = fptosi <vscale x 2 x half> undef to <vscale x 2 x i8>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_ui8 = fptoui <vscale x 2 x half> undef to <vscale x 2 x i8>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_si32 = fptosi <vscale x 2 x half> undef to <vscale x 2 x i32>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_ui32 = fptoui <vscale x 2 x half> undef to <vscale x 2 x i32>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_si64 = fptosi <vscale x 2 x half> undef to <vscale x 2 x i64>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f16_to_ui64 = fptoui <vscale x 2 x half> undef to <vscale x 2 x i64>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f32_to_si8 = fptosi <vscale x 2 x float> undef to <vscale x 2 x i8>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction:   %nv2f32_to_ui8 = fptoui <vscale x 2 x float> undef to <vscale x 2 x i8>

 ; CHECK-NEXT: Cost Model: Found an estimated cost of 8 for instruction:   %nv8f32_to_si64 = fptosi <vscale x 8 x float> undef to <vscale x 8 x i64>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 8 for instruction:   %nv8f32_to_ui64 = fptoui <vscale x 8 x float> undef to <vscale x 8 x i64>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 7 for instruction:   %nv8f64_to_si8 = fptosi <vscale x 8 x double> undef to <vscale x 8 x i8>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 7 for instruction:   %nv8f64_to_ui8 = fptoui <vscale x 8 x double> undef to <vscale x 8 x i8>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 7 for instruction:   %nv8f64_to_si16 = fptosi <vscale x 8 x double> undef to <vscale x 8 x i16>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 7 for instruction:   %nv8f64_to_ui16 = fptoui <vscale x 8 x double> undef to <vscale x 8 x i16>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 6 for instruction:   %nv8f64_to_si32 = fptosi <vscale x 8 x double> undef to <vscale x 8 x i32>
 ; CHECK-NEXT: Cost Model: Found an estimated cost of 6 for instruction:   %nv8f64_to_ui32 = fptoui <vscale x 8 x double> undef to <vscale x 8 x i32>
+  %nv1f16_to_si8  = fptosi <vscale x 1 x half> undef to <vscale x 1 x i8>
+  %nv1f16_to_ui8  = fptoui <vscale x 1 x half> undef to <vscale x 1 x i8>
+  %nv1f16_to_si32 = fptosi <vscale x 1 x half> undef to <vscale x 1 x i32>
+  %nv1f16_to_ui32 = fptoui <vscale x 1 x half> undef to <vscale x 1 x i32>
+  %nv1f16_to_si64 = fptosi <vscale x 1 x half> undef to <vscale x 1 x i64>
+  %nv1f16_to_ui64 = fptoui <vscale x 1 x half> undef to <vscale x 1 x i64>
+
+  %nv1f32_to_si8  = fptosi <vscale x 1 x float> undef to <vscale x 1 x i8>
+  %nv1f32_to_ui8  = fptoui <vscale x 1 x float> undef to <vscale x 1 x i8>
+  %nv1f32_to_si16 = fptosi <vscale x 1 x float> undef to <vscale x 1 x i16>
+  %nv1f32_to_ui16 = fptoui <vscale x 1 x float> undef to <vscale x 1 x i16>
+  %nv1f32_to_si64 = fptosi <vscale x 1 x float> undef to <vscale x 1 x i64>
+  %nv1f32_to_ui64 = fptoui <vscale x 1 x float> undef to <vscale x 1 x i64>
+
+  %nv1f64_to_si8  = fptosi <vscale x 1 x double> undef to <vscale x 1 x i8>
+  %nv1f64_to_ui8  = fptoui <vscale x 1 x double> undef to <vscale x 1 x i8>
+  %nv1f64_to_si16 = fptosi <vscale x 1 x double> undef to <vscale x 1 x i16>
+  %nv1f64_to_ui16 = fptoui <vscale x 1 x double> undef to <vscale x 1 x i16>
+  %nv1f64_to_si32 = fptosi <vscale x 1 x double> undef to <vscale x 1 x i32>
+  %nv1f64_to_ui32 = fptoui <vscale x 1 x double> undef to <vscale x 1 x i32>
 
   %nv2f16_to_si8  = fptosi <vscale x 2 x half> undef to <vscale x 2 x i8>
   %nv2f16_to_ui8  = fptoui <vscale x 2 x half> undef to <vscale x 2 x i8>
   %nv2f16_to_si32 = fptosi <vscale x 2 x half> undef to <vscale x 2 x i32>
   %nv2f16_to_ui32 = fptoui <vscale x 2 x half> undef to <vscale x 2 x i32>
   %nv2f16_to_si64 = fptosi <vscale x 2 x half> undef to <vscale x 2 x i64>
   %nv2f16_to_ui64 = fptoui <vscale x 2 x half> undef to <vscale x 2 x i64>