[SystemZ] Recognize the higher cost of scalar i1 -> fp conversion

[jonpa]

Summary

Scalar i1 to fp conversions are done with a branch sequence, so it should have a higher cost.

[uweigand]

LGTM, thanks.

[jonpa]

r345818

lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

if (Opcode == Instruction::FPExt) {
       // -> fp128.  VF * lxdb/lxeb + extraction of elements.
       return VF + getScalarizationOverhead(Src, false, true);
     }
   }
   else { // Scalar
     assert (!Dst->isVectorTy());
 
     if (Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP)
-      return (SrcScalarBits >= 32 ? 1 : 2 /*i8/i16 extend*/);
+      return (SrcScalarBits >= 32
+                ? 1
+                : SrcScalarBits > 1 ? 2 /*i8/i16 extend*/ : 5 /*branch seq.*/);
 
     if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
         Src->isIntegerTy(1)) {
       if (ST->hasLoadStoreOnCond2())
         return 2; // li 0; loc 1
 
       // This should be extension of a compare i1 result, which is done with
       // ipm and a varying sequence of instructions.

test/Analysis/CostModel/SystemZ/cmp-tofp-scalar.ll

+; RUN: opt < %s -cost-model -analyze -mtriple=systemz-unknown -mcpu=z13 | FileCheck %s
+;
+; Costs for conversion of i1 to fp.
+
+define float @fun0(i64 %val1, i64 %val2) {
+  %cmp = icmp eq i64 %val1, %val2
+  %v = uitofp i1 %cmp to float
+  ret float %v
+
+; CHECK: fun0
+; CHECK: cost of 1 for instruction:   %cmp = icmp eq i64 %val1, %val2
+; CHECK: cost of 5 for instruction:   %v = uitofp i1 %cmp to float
+}
+
+define double @fun1(i64 %val1, i64 %val2) {
+  %cmp = icmp eq i64 %val1, %val2
+  %v = uitofp i1 %cmp to double
+  ret double %v
+
+; CHECK: fun1
+; CHECK: cost of 1 for instruction:   %cmp = icmp eq i64 %val1, %val2
+; CHECK: cost of 5 for instruction:   %v = uitofp i1 %cmp to double
+}