From 011a503f25cdbfff68bc28c8ac11f7df346674ef Mon Sep 17 00:00:00 2001
From: Jonas Paulsson <paulsson@linux.vnet.ibm.com>
Date: Wed, 28 Nov 2018 08:08:05 +0000
Subject: [PATCH] [SystemZ::TTI] Improved cost values for comparison against
 memory.

Single instructions exist for i8 and i16 comparisons of memory against a
small immediate.

This patch makes sure that if the load in these cases has a single user (the
ICmp), it gets a 0 cost (folded), and also that the ICmp gets a cost of 1.

Review: Ulrich Weigand
https://reviews.llvm.org/D54897

llvm-svn: 347733
---
 .../Target/SystemZ/SystemZTargetTransformInfo.cpp  | 17 +++++++++++++-
 llvm/test/Analysis/CostModel/SystemZ/cmp-mem.ll    | 27 ++++++++++++++++++++++
 2 files changed, 43 insertions(+), 1 deletion(-)
 create mode 100644 llvm/test/Analysis/CostModel/SystemZ/cmp-mem.ll

diff --git a/llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp b/llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
index 19e4448..fdb998e 100644
--- a/llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
+++ b/llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
@@ -835,8 +835,17 @@ int SystemZTTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
     switch (Opcode) {
     case Instruction::ICmp: {
       unsigned Cost = 1;
-      if (ValTy->isIntegerTy() && ValTy->getScalarSizeInBits() <= 16)
+      if (ValTy->isIntegerTy() && ValTy->getScalarSizeInBits() <= 16) {
+        if (I != nullptr) {
+          // Single instruction for comparison of memory with a small immediate.
+          if (const LoadInst* Ld = dyn_cast<LoadInst>(I->getOperand(0))) {
+            const Instruction *FoldedValue = nullptr;
+            if (isFoldableLoad(Ld, FoldedValue))
+              return Cost;
+          }
+        }
         Cost += 2; // extend both operands
+      }
       return Cost;
     }
     case Instruction::Select:
@@ -932,6 +941,12 @@ isFoldableLoad(const LoadInst *Ld, const Instruction *&FoldedValue) {
     if (SExtBits || ZExtBits)
       return false;
 
+    // Comparison between memory and immediate.
+    if (UserI->getOpcode() == Instruction::ICmp)
+      if (ConstantInt *CI = dyn_cast<ConstantInt>(UserI->getOperand(1)))
+        if (isUInt<16>(CI->getZExtValue()))
+          return true;
+
     unsigned LoadOrTruncBits = (TruncBits ? TruncBits : LoadedBits);
     return (LoadOrTruncBits == 32 || LoadOrTruncBits == 64);
     break;
diff --git a/llvm/test/Analysis/CostModel/SystemZ/cmp-mem.ll b/llvm/test/Analysis/CostModel/SystemZ/cmp-mem.ll
new file mode 100644
index 0000000..4f92d5b
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/SystemZ/cmp-mem.ll
@@ -0,0 +1,27 @@
+; RUN: opt < %s -cost-model -analyze -mtriple=systemz-unknown -mcpu=z13 | FileCheck %s
+;
+; Test costs for i8 and i16 comparisons against memory with a small immediate.
+
+define i32 @fun0(i8* %Src, i8* %Dst, i8 %Val) {
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun0':
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %Ld = load i8, i8* %Src
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %Cmp = icmp eq i8 %Ld, 123
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %Ret = zext i1 %Cmp to i32
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret i32 %Ret
+  %Ld = load i8, i8* %Src
+  %Cmp = icmp eq i8 %Ld, 123
+  %Ret = zext i1 %Cmp to i32
+  ret i32 %Ret
+}
+
+define i32 @fun1(i16* %Src, i16* %Dst, i16 %Val) {
+; CHECK: Printing analysis 'Cost Model Analysis' for function 'fun1':
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %Ld = load i16, i16* %Src
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %Cmp = icmp eq i16
+; CHECK: Cost Model: Found an estimated cost of 2 for instruction:   %Ret = zext i1 %Cmp to i32
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   ret i32 %Ret
+  %Ld = load i16, i16* %Src
+  %Cmp = icmp eq i16 %Ld, 1234
+  %Ret = zext i1 %Cmp to i32
+  ret i32 %Ret
+}
-- 
2.7.4