[SystemZ::TTI] Return zero cost for icmp in case of Load-And-Test

[jonpa]

Summary

A loaded value with multiple users compared with 0 will become a load and test single instruction. The load is not folded in this case (multiple users), but the compare instruction is eliminated.

This patch returns 0 cost for the icmp in these cases.

This changes just 33 instruction query results. One file changed on spec - two loops are now are kept scalar (not vectorized), see:

report_LT44 KBDownload

I tried also handling the load i32 ->sext i64 case, but this doubled the LOCs of the patch, while changing absolutely nothing (not a single LV query / file), so this seemed not useful enough to keep.

[uweigand]

LGTM, thanks!

[jonpa]

r348141

lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

else { // Called with a select instruction.
           getVectorBitmaskConversionCost(CmpOpTy, ValTy);
 
       return getNumVectorRegs(ValTy) /*vsel*/ + PackCost;
     }
   }
   else { // Scalar
     switch (Opcode) {
     case Instruction::ICmp: {
+      // A loaded value compared with 0 with multiple users becomes Load and
+      // Test. The load is then not foldable, so return 0 cost for the ICmp.
+      unsigned ScalarBits = ValTy->getScalarSizeInBits();
+      if (I != nullptr && ScalarBits >= 32)
+        if (LoadInst *Ld = dyn_cast<LoadInst>(I->getOperand(0)))
+          if (const ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
+            if (!Ld->hasOneUse() && Ld->getParent() == I->getParent() &&
+                C->getZExtValue() == 0)
+              return 0;
+
       unsigned Cost = 1;
       if (ValTy->isIntegerTy() && ValTy->getScalarSizeInBits() <= 16)
         Cost += (I != nullptr ? getOperandsExtensionCost(I) : 2);
       return Cost;
     }
     case Instruction::Select:
       if (ValTy->isFloatingPointTy())
         return 4; // No load on condition for FP - costs a conditional jump.

test/Analysis/CostModel/SystemZ/load-and-test.ll

+; RUN: opt < %s -cost-model -analyze -mtriple=systemz-unknown -mcpu=z13 | FileCheck %s
+;
+; Test that load and test results in 0 cost for the compare.
+
+define i64 @fun0(i64* %Src, i64 %Arg) {
+  %Ld1 = load i64, i64* %Src
+  %Cmp = icmp eq i64 %Ld1, 0
+  %S   = select i1 %Cmp, i64 %Arg, i64 %Ld1
+  ret i64 %S
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun0':
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %Ld1 = load i64, i64* %Src
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %Cmp = icmp eq i64 %Ld1, 0
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %S = select
+}
+
+define i32 @fun1(i32* %Src, i32 %Arg) {
+  %Ld1 = load i32, i32* %Src
+  %Cmp = icmp eq i32 %Ld1, 0
+  %S   = select i1 %Cmp, i32 %Arg, i32 %Ld1
+  ret i32 %S
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun1':
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %Ld1 = load i32, i32* %Src
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %Cmp = icmp eq i32 %Ld1, 0
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %S = select
+}