[SystemZ::TTI] Let i8/i16 uint/sint to fp conversions cost 1 if operand is a load.

[jonpa]

Summary

Since the load already does the extension, there is no extra cost (previously returned 2).

This affects ~5 loop vectorizer decisions, seemingly for the better. See

report_int2fp_extload71 KBDownload

[uweigand]

LGTM, thanks.

[jonpa]

r346009

lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

if (Opcode == Instruction::FPExt) {
       }
       // -> fp128.  VF * lxdb/lxeb + extraction of elements.
       return VF + getScalarizationOverhead(Src, false, Extract);
     }
   }
   else { // Scalar
     assert (!Dst->isVectorTy());
 
-    if (Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP)
-      return (SrcScalarBits >= 32
-                ? 1
-                : SrcScalarBits > 1 ? 2 /*i8/i16 extend*/ : 5 /*branch seq.*/);
+    if (Opcode == Instruction::SIToFP || Opcode == Instruction::UIToFP) {
+      if (SrcScalarBits >= 32 ||
+          (I != nullptr && isa<LoadInst>(I->getOperand(0))))
+        return 1;
+      return 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/fp-cast.ll

 ; CHECK: Cost Model: Found an estimated cost of 48 for instruction:   %v51 = uitofp <16 x i32> undef to <16 x float>
 ; CHECK: Cost Model: Found an estimated cost of 64 for instruction:   %v52 = uitofp <16 x i16> undef to <16 x double>
 ; CHECK: Cost Model: Found an estimated cost of 64 for instruction:   %v53 = uitofp <16 x i16> undef to <16 x float>
 ; CHECK: Cost Model: Found an estimated cost of 64 for instruction:   %v54 = uitofp <16 x i8> undef to <16 x double>
 ; CHECK: Cost Model: Found an estimated cost of 64 for instruction:   %v55 = uitofp <16 x i8> undef to <16 x float>
 
   ret void;
 }
+
+define void @sitofp_extload(i16 *%src16, i8 *%src8) {
+  %ld16 = load i16, i16 *%src16
+  %v6 = sitofp i16 %ld16 to fp128
+  %v7 = sitofp i16 %ld16 to double
+  %v8 = sitofp i16 %ld16 to float
+
+  %ld8 = load i8, i8 *%src8
+  %v9 = sitofp i8 %ld8 to fp128
+  %v10 = sitofp i8 %ld8 to double
+  %v11 = sitofp i8 %ld8 to float
+
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %ld16 = load i16, i16* %src16
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v6 = sitofp i16 %ld16 to fp128
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v7 = sitofp i16 %ld16 to double
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v8 = sitofp i16 %ld16 to float
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %ld8 = load i8, i8* %src8
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v9 = sitofp i8 %ld8 to fp128
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v10 = sitofp i8 %ld8 to double
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v11 = sitofp i8 %ld8 to float
+
+  ret void;
+}
+
+define void @uitofp_extload(i16 *%src16, i8 *%src8) {
+  %ld16 = load i16, i16 *%src16
+  %v6 = uitofp i16 %ld16 to fp128
+  %v7 = uitofp i16 %ld16 to double
+  %v8 = uitofp i16 %ld16 to float
+
+  %ld8 = load i8, i8 *%src8
+  %v9 = uitofp i8 %ld8 to fp128
+  %v10 = uitofp i8 %ld8 to double
+  %v11 = uitofp i8 %ld8 to float
+
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %ld16 = load i16, i16* %src16
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v6 = uitofp i16 %ld16 to fp128
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v7 = uitofp i16 %ld16 to double
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v8 = uitofp i16 %ld16 to float
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %ld8 = load i8, i8* %src8
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v9 = uitofp i8 %ld8 to fp128
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v10 = uitofp i8 %ld8 to double
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %v11 = uitofp i8 %ld8 to float
+
+  ret void;
+}