From ea9fed10ed4b945118e4fc13fc1f3362ffa52e3c Mon Sep 17 00:00:00 2001
From: Jonas Paulsson <paulsson@linux.vnet.ibm.com>
Date: Wed, 28 Nov 2018 08:58:27 +0000
Subject: [PATCH] [SystemZ::TTI]  Improve cost for compare of i64 with extended
 i32 load

CGF/CLGF compares an i64 register with a sign/zero extended loaded i32 value
in memory.

This patch makes such a load considered foldable and so gets a 0 cost.

Review: Ulrich Weigand
https://reviews.llvm.org/D54944

llvm-svn: 347735
---
 .../SystemZ/SystemZTargetTransformInfo.cpp     | 16 +++++++++-------
 .../SystemZ/memop-folding-int-arith.ll         | 18 +++++++++++++++++-
 2 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp b/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
index 362d2d3c79d..32883efbdab 100644
--- a/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
+++ b/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp
@@ -915,16 +915,19 @@ isFoldableLoad(const LoadInst *Ld, const Instruction *&FoldedValue) {
   switch (UserI->getOpcode()) {
   case Instruction::Add: // SE: 16->32, 16/32->64, z14:16->64. ZE: 32->64
   case Instruction::Sub:
+  case Instruction::ICmp:
     if (LoadedBits == 32 && ZExtBits == 64)
       return true;
     LLVM_FALLTHROUGH;
   case Instruction::Mul: // SE: 16->32, 32->64, z14:16->64
-    if (LoadedBits == 16 &&
-        (SExtBits == 32 ||
-         (SExtBits == 64 && ST->hasMiscellaneousExtensions2())))
-      return true;
-    if (LoadOrTruncBits == 16)
-      return true;
+    if (UserI->getOpcode() != Instruction::ICmp) {
+      if (LoadedBits == 16 &&
+          (SExtBits == 32 ||
+           (SExtBits == 64 && ST->hasMiscellaneousExtensions2())))
+        return true;
+      if (LoadOrTruncBits == 16)
+        return true;
+    }
     LLVM_FALLTHROUGH;
   case Instruction::SDiv:// SE: 32->64
     if (LoadedBits == 32 && SExtBits == 64)
@@ -934,7 +937,6 @@ isFoldableLoad(const LoadInst *Ld, const Instruction *&FoldedValue) {
   case Instruction::And:
   case Instruction::Or:
   case Instruction::Xor:
-  case Instruction::ICmp:
     // This also makes sense for float operations, but disabled for now due
     // to regressions.
     // case Instruction::FCmp:
diff --git a/test/Analysis/CostModel/SystemZ/memop-folding-int-arith.ll b/test/Analysis/CostModel/SystemZ/memop-folding-int-arith.ll
index afd923a9f3a..74d95ae2ea7 100644
--- a/test/Analysis/CostModel/SystemZ/memop-folding-int-arith.ll
+++ b/test/Analysis/CostModel/SystemZ/memop-folding-int-arith.ll
@@ -726,6 +726,16 @@ define void @icmp() {
   %tr_0 = trunc i64 %li64_2 to i32
   icmp eq i32 %tr_0, undef
 
+  ; Sign-extended load
+  %li32_2 = load i32, i32* undef
+  %sext = sext i32 %li32_2 to i64
+  icmp eq i64 %sext, undef
+
+  ; Zero-extended load
+  %li32_3 = load i32, i32* undef
+  %zext = zext i32 %li32_3 to i64
+  icmp eq i64 %zext, undef
+
   ; Loads with multiple uses are *not* folded
   %li64_3 = load i64, i64* undef
   %tr_1 = trunc i64 %li64_3 to i32
@@ -745,7 +755,13 @@ define void @icmp() {
 ; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %li64_2 = load i64, i64* undef
 ; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %tr_0 = trunc i64 %li64_2 to i32
 ; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %5 = icmp eq i32 %tr_0, undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %li32_2 = load i32, i32* undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %sext = sext i32 %li32_2 to i64
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %6 = icmp eq i64 %sext, undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %li32_3 = load i32, i32* undef
+; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %zext = zext i32 %li32_3 to i64
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %7 = icmp eq i64 %zext, undef
 ; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %li64_3 = load i64, i64* undef
 ; CHECK: Cost Model: Found an estimated cost of 0 for instruction:   %tr_1 = trunc i64 %li64_3 to i32
-; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %6 = icmp eq i64 %li64_3, undef
+; CHECK: Cost Model: Found an estimated cost of 1 for instruction:   %8 = icmp eq i64 %li64_3, undef
 }