Index: lib/Transforms/Scalar/MergeICmps.cpp
===================================================================
--- lib/Transforms/Scalar/MergeICmps.cpp
+++ lib/Transforms/Scalar/MergeICmps.cpp
@@ -24,6 +24,8 @@
 
 #include "llvm/ADT/APSInt.h"
 #include "llvm/Analysis/Loads.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/IntrinsicInst.h"
@@ -41,7 +43,7 @@
 
 #define DEBUG_TYPE "mergeicmps"
 
-#define MERGEICMPS_DOT_ON
+// #define MERGEICMPS_DOT_ON
 
 // A BCE atom.
 struct BCEAtom {
@@ -605,22 +607,33 @@
   bool runOnFunction(Function &F) override {
     if (skipFunction(F)) return false;
     const auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
-    auto PA = runImpl(F, &TLI);
+    const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
+    auto PA = runImpl(F, &TLI, &TTI);
     return !PA.areAllPreserved();
   }
 
  private:
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<TargetLibraryInfoWrapperPass>();
+    AU.addRequired<TargetTransformInfoWrapperPass>();
   }
 
-  PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI);
+  PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI,
+                            const TargetTransformInfo *TTI);
 };
 
 PreservedAnalyses MergeICmps::runImpl(Function &F,
-                                      const TargetLibraryInfo *TLI) {
+                                      const TargetLibraryInfo *TLI,
+                                      const TargetTransformInfo *TTI) {
   DEBUG(dbgs() << "MergeICmpsPass: " << F.getName() << "\n");
 
+  // We only try merging comparisons if the target wants to expand memcmp later.
+  // The rationale is to avoid turning small chains into memcmp calls.
+  unsigned MaxLoadSize;
+  if(!TTI->enableMemCmpExpansion(MaxLoadSize)) {
+    return PreservedAnalyses::all();
+  }
+
   bool MadeChange = false;
 
   for (auto BBIt = ++F.begin(); BBIt != F.end(); ++BBIt) {
@@ -640,6 +653,7 @@
 INITIALIZE_PASS_BEGIN(MergeICmps, "mergeicmps",
                       "Merge contiguous icmps into a memcmp", false, false)
 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
 INITIALIZE_PASS_END(MergeICmps, "mergeicmps",
                     "Merge contiguous icmps into a memcmp", false, false)
 
Index: test/Transforms/MergeICmps/pair-int32-int32.ll
===================================================================
--- test/Transforms/MergeICmps/pair-int32-int32.ll
+++ test/Transforms/MergeICmps/pair-int32-int32.ll
@@ -1,4 +1,5 @@
-; RUN: opt -mergeicmps -S -o - %s | FileCheck %s
+; RUN: opt -mergeicmps -mtriple=x86_64-unknown-unknown -S -o - %s | FileCheck %s --check-prefix=X86
+; RUN: opt -mergeicmps -S -o - %s | FileCheck %s --check-prefix=NOEXPANSION
 
 %"struct.std::pair" = type { i32, i32 }
 
@@ -26,20 +27,21 @@
   ret i1 %4
 ; CHECK-LABEL: @opeq1(
 ; The entry block with zero-offset GEPs is kept, loads are removed.
-; CHECK: entry
-; CHECK:     getelementptr {{.*}} i32 0
-; CHECK-NOT: load
-; CHECK:     getelementptr {{.*}} i32 0
-; CHECK-NOT: load
+; X86: entry
+; X86:     getelementptr {{.*}} i32 0
+; X86-NOT: load
+; X86:     getelementptr {{.*}} i32 0
+; X86-NOT: load
 ; The two 4 byte loads and compares are replaced with a single 8-byte memcmp.
-; CHECK:     @memcmp({{.*}}8)
-; CHECK:     icmp eq {{.*}} 0
+; X86:     @memcmp({{.*}}8)
+; X86:     icmp eq {{.*}} 0
 ; The branch is now a direct branch; the other block has been removed.
-; CHECK:     br label %opeq1.exit
-; CHECK-NOT: br
+; X86:     br label %opeq1.exit
+; X86-NOT: br
 ; The phi is updated.
-; CHECK:      phi i1 [ %{{[^,]*}}, %entry ]
-; CHECK-NEXT: ret
+; X86:      phi i1 [ %{{[^,]*}}, %entry ]
+; X86-NEXT: ret
+; NOEXPANSION-NOT: @memcmp({{.*}}8)
 }
 
 ; Same as above, but the two blocks are in inverse order.
@@ -68,19 +70,20 @@
 ; CHECK-LABEL: @opeq1_inverse(
 ; The second block with zero-offset GEPs is kept, loads are removed.
 ; CHECK: land.rhs.i
-; CHECK:     getelementptr {{.*}} i32 0
-; CHECK-NOT: load
-; CHECK:     getelementptr {{.*}} i32 0
-; CHECK-NOT: load
+; X86:     getelementptr {{.*}} i32 0
+; X86-NOT: load
+; X86:     getelementptr {{.*}} i32 0
+; X86-NOT: load
 ; The two 4 byte loads and compares are replaced with a single 8-byte memcmp.
-; CHECK:     @memcmp({{.*}}8)
-; CHECK:     icmp eq {{.*}} 0
+; X86:     @memcmp({{.*}}8)
+; X86:     icmp eq {{.*}} 0
 ; The branch is now a direct branch; the other block has been removed.
-; CHECK:     br label %opeq1.exit
-; CHECK-NOT: br
+; X86:     br label %opeq1.exit
+; X86-NOT: br
 ; The phi is updated.
-; CHECK:      phi i1 [ %{{[^,]*}}, %land.rhs.i ]
-; CHECK-NEXT: ret
+; X86:      phi i1 [ %{{[^,]*}}, %land.rhs.i ]
+; X86-NEXT: ret
+; NOEXPANSION-NOT:     @memcmp({{.*}}8)
 }
 
 
Index: test/Transforms/MergeICmps/tuple-four-int8.ll
===================================================================
--- test/Transforms/MergeICmps/tuple-four-int8.ll
+++ test/Transforms/MergeICmps/tuple-four-int8.ll
@@ -1,4 +1,4 @@
-; RUN: opt -mergeicmps -S -o - %s | FileCheck %s
+; RUN: opt -mergeicmps -mtriple=x86_64-unknown-unknown -S -o - %s | FileCheck %s
 
 ; This is a more involved test: clang generates this weird pattern for
 ; tuple<uint8_t, uint8_t, uint8_t, uint8_t>. Right now we skip the entry block
Index: test/Transforms/MergeICmps/volatile.ll
===================================================================
--- test/Transforms/MergeICmps/volatile.ll
+++ test/Transforms/MergeICmps/volatile.ll
@@ -1,4 +1,4 @@
-; RUN: opt -mergeicmps -S -o - %s | FileCheck %s
+; RUN: opt -mergeicmps -mtriple=x86_64-unknown-unknown -S -o - %s | FileCheck %s
 
 %"struct.std::pair" = type { i32, i32 }