X86 cost model: Vector shifts are expensive in most cases

The default logic does not correctly identify costs of casts because they are
marked as custom on x86.

For some cases, where the shift amount is a scalar we would be able to generate
better code. Unfortunately, when this is the case the value (the splat) will get
hoisted out of the loop, thereby making it invisible to ISel.

radar://13130673
radar://13537826

llvm-svn: 178703

5 files changed, 772 insertions, 2 deletions

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index 2336035bea0..a38b22bd789 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -182,6 +182,16 @@ unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty) const {
     { ISD::SRL,     MVT::v2i64,    1 },
     { ISD::SHL,     MVT::v4i64,    1 },
     { ISD::SRL,     MVT::v4i64,    1 },
+
+    { ISD::SHL,  MVT::v32i8,  42 }, // cmpeqb sequence.
+    { ISD::SHL,  MVT::v16i16,  16*10 }, // Scalarized.
+
+    { ISD::SRL,  MVT::v32i8,  32*10 }, // Scalarized.
+    { ISD::SRL,  MVT::v16i16,  8*10 }, // Scalarized.
+
+    { ISD::SRA,  MVT::v32i8,  32*10 }, // Scalarized.
+    { ISD::SRA,  MVT::v16i16,  16*10 }, // Scalarized.
+    { ISD::SRA,  MVT::v4i64,  4*10 }, // Scalarized.
   };
 
   // Look for AVX2 lowering tricks.
@@ -192,6 +202,38 @@ unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty) const {
       return LT.first * AVX2CostTable[Idx].Cost;
   }
 
+  static const CostTblEntry<MVT> SSE2CostTable[] = {
+    // We don't correctly identify costs of casts because they are marked as
+    // custom.
+    // For some cases, where the shift amount is a scalar we would be able
+    // to generate better code. Unfortunately, when this is the case the value
+    // (the splat) will get hoisted out of the loop, thereby making it invisible
+    // to ISel. The cost model must return worst case assumptions because it is
+    // used for vectorization and we don't want to make vectorized code worse
+    // than scalar code.
+    { ISD::SHL,  MVT::v16i8,  30 }, // cmpeqb sequence.
+    { ISD::SHL,  MVT::v8i16,  8*10 }, // Scalarized.
+    { ISD::SHL,  MVT::v4i32,  2*5 }, // We optimized this using mul.
+    { ISD::SHL,  MVT::v2i64,  2*10 }, // Scalarized.
+
+    { ISD::SRL,  MVT::v16i8,  16*10 }, // Scalarized.
+    { ISD::SRL,  MVT::v8i16,  8*10 }, // Scalarized.
+    { ISD::SRL,  MVT::v4i32,  4*10 }, // Scalarized.
+    { ISD::SRL,  MVT::v2i64,  2*10 }, // Scalarized.
+
+    { ISD::SRA,  MVT::v16i8,  16*10 }, // Scalarized.
+    { ISD::SRA,  MVT::v8i16,  8*10 }, // Scalarized.
+    { ISD::SRA,  MVT::v4i32,  4*10 }, // Scalarized.
+    { ISD::SRA,  MVT::v2i64,  2*10 }, // Scalarized.
+  };
+
+  if (ST->hasSSE2()) {
+    int Idx = CostTableLookup<MVT>(SSE2CostTable, array_lengthof(SSE2CostTable),
+                                   ISD, LT.second);
+    if (Idx != -1)
+      return LT.first * SSE2CostTable[Idx].Cost;
+  }
+
   static const CostTblEntry<MVT> AVX1CostTable[] = {
     // We don't have to scalarize unsupported ops. We can issue two half-sized
     // operations and we only need to extract the upper YMM half.
diff --git a/llvm/test/Analysis/CostModel/X86/arith.ll b/llvm/test/Analysis/CostModel/X86/arith.ll
index 5f9444157aa..85b442533f4 100644
--- a/llvm/test/Analysis/CostModel/X86/arith.ll
+++ b/llvm/test/Analysis/CostModel/X86/arith.ll
@@ -94,7 +94,7 @@ define void @shift() {
   ; AVX2: cost of 1 {{.*}} ashr
   %C0 = ashr <4 x i32> undef, undef
   ; AVX: cost of 6 {{.*}} ashr
-  ; AVX2: cost of 6 {{.*}} ashr
+  ; AVX2: cost of 20 {{.*}} ashr
   %C1 = ashr <2 x i64> undef, undef
 
   ret void
@@ -121,7 +121,7 @@ define void @avx2shift() {
   ; AVX2: cost of 1 {{.*}} ashr
   %C0 = ashr <8 x i32> undef, undef
   ; AVX: cost of 12 {{.*}} ashr
-  ; AVX2: cost of 12 {{.*}} ashr
+  ; AVX2: cost of 40 {{.*}} ashr
   %C1 = ashr <4 x i64> undef, undef
 
   ret void
diff --git a/llvm/test/Analysis/CostModel/X86/testshiftashr.ll b/llvm/test/Analysis/CostModel/X86/testshiftashr.ll
new file mode 100644
index 00000000000..d932b2a4c42
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/X86/testshiftashr.ll
@@ -0,0 +1,243 @@
+; RUN: llc -mtriple=x86_64-apple-darwin -mcpu=core2 < %s | FileCheck --check-prefix=SSE2-CODEGEN %s
+; RUN: opt -mtriple=x86_64-apple-darwin -mcpu=core2 -cost-model -analyze < %s | FileCheck --check-prefix=SSE2 %s
+
+%shifttype = type <2 x i16>
+define %shifttype @shift2i16(%shifttype %a, %shifttype %b) {
+entry:
+  ; SSE2: shift2i16
+  ; SSE2: cost of 20 {{.*}} ashr
+  ; SSE2-CODEGEN: shift2i16
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype %a , %b
+  ret %shifttype %0
+}
+
+%shifttype4i16 = type <4 x i16>
+define %shifttype4i16 @shift4i16(%shifttype4i16 %a, %shifttype4i16 %b) {
+entry:
+  ; SSE2: shift4i16
+  ; SSE2: cost of 40 {{.*}} ashr
+  ; SSE2-CODEGEN: shift4i16
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype4i16 %a , %b
+  ret %shifttype4i16 %0
+}
+
+%shifttype8i16 = type <8 x i16>
+define %shifttype8i16 @shift8i16(%shifttype8i16 %a, %shifttype8i16 %b) {
+entry:
+  ; SSE2: shift8i16
+  ; SSE2: cost of 80 {{.*}} ashr
+  ; SSE2-CODEGEN: shift8i16
+  ; SSE2-CODEGEN: sarw %cl
+
+  %0 = ashr %shifttype8i16 %a , %b
+  ret %shifttype8i16 %0
+}
+
+%shifttype16i16 = type <16 x i16>
+define %shifttype16i16 @shift16i16(%shifttype16i16 %a, %shifttype16i16 %b) {
+entry:
+  ; SSE2: shift16i16
+  ; SSE2: cost of 160 {{.*}} ashr
+  ; SSE2-CODEGEN: shift16i16
+  ; SSE2-CODEGEN: sarw %cl
+
+  %0 = ashr %shifttype16i16 %a , %b
+  ret %shifttype16i16 %0
+}
+
+%shifttype32i16 = type <32 x i16>
+define %shifttype32i16 @shift32i16(%shifttype32i16 %a, %shifttype32i16 %b) {
+entry:
+  ; SSE2: shift32i16
+  ; SSE2: cost of 320 {{.*}} ashr
+  ; SSE2-CODEGEN: shift32i16
+  ; SSE2-CODEGEN: sarw %cl
+
+  %0 = ashr %shifttype32i16 %a , %b
+  ret %shifttype32i16 %0
+}
+
+%shifttype2i32 = type <2 x i32>
+define %shifttype2i32 @shift2i32(%shifttype2i32 %a, %shifttype2i32 %b) {
+entry:
+  ; SSE2: shift2i32
+  ; SSE2: cost of 20 {{.*}} ashr
+  ; SSE2-CODEGEN: shift2i32
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype2i32 %a , %b
+  ret %shifttype2i32 %0
+}
+
+%shifttype4i32 = type <4 x i32>
+define %shifttype4i32 @shift4i32(%shifttype4i32 %a, %shifttype4i32 %b) {
+entry:
+  ; SSE2: shift4i32
+  ; SSE2: cost of 40 {{.*}} ashr
+  ; SSE2-CODEGEN: shift4i32
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype4i32 %a , %b
+  ret %shifttype4i32 %0
+}
+
+%shifttype8i32 = type <8 x i32>
+define %shifttype8i32 @shift8i32(%shifttype8i32 %a, %shifttype8i32 %b) {
+entry:
+  ; SSE2: shift8i32
+  ; SSE2: cost of 80 {{.*}} ashr
+  ; SSE2-CODEGEN: shift8i32
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype8i32 %a , %b
+  ret %shifttype8i32 %0
+}
+
+%shifttype16i32 = type <16 x i32>
+define %shifttype16i32 @shift16i32(%shifttype16i32 %a, %shifttype16i32 %b) {
+entry:
+  ; SSE2: shift16i32
+  ; SSE2: cost of 160 {{.*}} ashr
+  ; SSE2-CODEGEN: shift16i32
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype16i32 %a , %b
+  ret %shifttype16i32 %0
+}
+
+%shifttype32i32 = type <32 x i32>
+define %shifttype32i32 @shift32i32(%shifttype32i32 %a, %shifttype32i32 %b) {
+entry:
+  ; SSE2: shift32i32
+  ; SSE2: cost of 256 {{.*}} ashr
+  ; SSE2-CODEGEN: shift32i32
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype32i32 %a , %b
+  ret %shifttype32i32 %0
+}
+
+%shifttype2i64 = type <2 x i64>
+define %shifttype2i64 @shift2i64(%shifttype2i64 %a, %shifttype2i64 %b) {
+entry:
+  ; SSE2: shift2i64
+  ; SSE2: cost of 20 {{.*}} ashr
+  ; SSE2-CODEGEN: shift2i64
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype2i64 %a , %b
+  ret %shifttype2i64 %0
+}
+
+%shifttype4i64 = type <4 x i64>
+define %shifttype4i64 @shift4i64(%shifttype4i64 %a, %shifttype4i64 %b) {
+entry:
+  ; SSE2: shift4i64
+  ; SSE2: cost of 40 {{.*}} ashr
+  ; SSE2-CODEGEN: shift4i64
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype4i64 %a , %b
+  ret %shifttype4i64 %0
+}
+
+%shifttype8i64 = type <8 x i64>
+define %shifttype8i64 @shift8i64(%shifttype8i64 %a, %shifttype8i64 %b) {
+entry:
+  ; SSE2: shift8i64
+  ; SSE2: cost of 80 {{.*}} ashr
+  ; SSE2-CODEGEN: shift8i64
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype8i64 %a , %b
+  ret %shifttype8i64 %0
+}
+
+%shifttype16i64 = type <16 x i64>
+define %shifttype16i64 @shift16i64(%shifttype16i64 %a, %shifttype16i64 %b) {
+entry:
+  ; SSE2: shift16i64
+  ; SSE2: cost of 160 {{.*}} ashr
+  ; SSE2-CODEGEN: shift16i64
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype16i64 %a , %b
+  ret %shifttype16i64 %0
+}
+
+%shifttype32i64 = type <32 x i64>
+define %shifttype32i64 @shift32i64(%shifttype32i64 %a, %shifttype32i64 %b) {
+entry:
+  ; SSE2: shift32i64
+  ; SSE2: cost of 256 {{.*}} ashr
+  ; SSE2-CODEGEN: shift32i64
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype32i64 %a , %b
+  ret %shifttype32i64 %0
+}
+
+%shifttype2i8 = type <2 x i8>
+define %shifttype2i8 @shift2i8(%shifttype2i8 %a, %shifttype2i8 %b) {
+entry:
+  ; SSE2: shift2i8
+  ; SSE2: cost of 20 {{.*}} ashr
+  ; SSE2-CODEGEN: shift2i8
+  ; SSE2-CODEGEN: sarq %cl
+
+  %0 = ashr %shifttype2i8 %a , %b
+  ret %shifttype2i8 %0
+}
+
+%shifttype4i8 = type <4 x i8>
+define %shifttype4i8 @shift4i8(%shifttype4i8 %a, %shifttype4i8 %b) {
+entry:
+  ; SSE2: shift4i8
+  ; SSE2: cost of 40 {{.*}} ashr
+  ; SSE2-CODEGEN: shift4i8
+  ; SSE2-CODEGEN: sarl %cl
+
+  %0 = ashr %shifttype4i8 %a , %b
+  ret %shifttype4i8 %0
+}
+
+%shifttype8i8 = type <8 x i8>
+define %shifttype8i8 @shift8i8(%shifttype8i8 %a, %shifttype8i8 %b) {
+entry:
+  ; SSE2: shift8i8
+  ; SSE2: cost of 80 {{.*}} ashr
+  ; SSE2-CODEGEN: shift8i8
+  ; SSE2-CODEGEN: sarw %cl
+
+  %0 = ashr %shifttype8i8 %a , %b
+  ret %shifttype8i8 %0
+}
+
+%shifttype16i8 = type <16 x i8>
+define %shifttype16i8 @shift16i8(%shifttype16i8 %a, %shifttype16i8 %b) {
+entry:
+  ; SSE2: shift16i8
+  ; SSE2: cost of 160 {{.*}} ashr
+  ; SSE2-CODEGEN: shift16i8
+  ; SSE2-CODEGEN: sarb %cl
+
+  %0 = ashr %shifttype16i8 %a , %b
+  ret %shifttype16i8 %0
+}
+
+%shifttype32i8 = type <32 x i8>
+define %shifttype32i8 @shift32i8(%shifttype32i8 %a, %shifttype32i8 %b) {
+entry:
+  ; SSE2: shift32i8
+  ; SSE2: cost of 320 {{.*}} ashr
+  ; SSE2-CODEGEN: shift32i8
+  ; SSE2-CODEGEN: sarb %cl
+
+  %0 = ashr %shifttype32i8 %a , %b
+  ret %shifttype32i8 %0
+}
+
diff --git a/llvm/test/Analysis/CostModel/X86/testshiftlshr.ll b/llvm/test/Analysis/CostModel/X86/testshiftlshr.ll
new file mode 100644
index 00000000000..7d665fc9c61
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/X86/testshiftlshr.ll
@@ -0,0 +1,243 @@
+; RUN: llc -mtriple=x86_64-apple-darwin -mcpu=core2 < %s | FileCheck --check-prefix=SSE2-CODEGEN %s
+; RUN: opt -mtriple=x86_64-apple-darwin -mcpu=core2 -cost-model -analyze < %s | FileCheck --check-prefix=SSE2 %s
+
+%shifttype = type <2 x i16>
+define %shifttype @shift2i16(%shifttype %a, %shifttype %b) {
+entry:
+  ; SSE2: shift2i16
+  ; SSE2: cost of 20 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i16
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype %a , %b
+  ret %shifttype %0
+}
+
+%shifttype4i16 = type <4 x i16>
+define %shifttype4i16 @shift4i16(%shifttype4i16 %a, %shifttype4i16 %b) {
+entry:
+  ; SSE2: shift4i16
+  ; SSE2: cost of 40 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i16
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype4i16 %a , %b
+  ret %shifttype4i16 %0
+}
+
+%shifttype8i16 = type <8 x i16>
+define %shifttype8i16 @shift8i16(%shifttype8i16 %a, %shifttype8i16 %b) {
+entry:
+  ; SSE2: shift8i16
+  ; SSE2: cost of 80 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i16
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype8i16 %a , %b
+  ret %shifttype8i16 %0
+}
+
+%shifttype16i16 = type <16 x i16>
+define %shifttype16i16 @shift16i16(%shifttype16i16 %a, %shifttype16i16 %b) {
+entry:
+  ; SSE2: shift16i16
+  ; SSE2: cost of 160 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i16
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype16i16 %a , %b
+  ret %shifttype16i16 %0
+}
+
+%shifttype32i16 = type <32 x i16>
+define %shifttype32i16 @shift32i16(%shifttype32i16 %a, %shifttype32i16 %b) {
+entry:
+  ; SSE2: shift32i16
+  ; SSE2: cost of 320 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i16
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype32i16 %a , %b
+  ret %shifttype32i16 %0
+}
+
+%shifttype2i32 = type <2 x i32>
+define %shifttype2i32 @shift2i32(%shifttype2i32 %a, %shifttype2i32 %b) {
+entry:
+  ; SSE2: shift2i32
+  ; SSE2: cost of 20 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i32
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype2i32 %a , %b
+  ret %shifttype2i32 %0
+}
+
+%shifttype4i32 = type <4 x i32>
+define %shifttype4i32 @shift4i32(%shifttype4i32 %a, %shifttype4i32 %b) {
+entry:
+  ; SSE2: shift4i32
+  ; SSE2: cost of 40 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i32
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype4i32 %a , %b
+  ret %shifttype4i32 %0
+}
+
+%shifttype8i32 = type <8 x i32>
+define %shifttype8i32 @shift8i32(%shifttype8i32 %a, %shifttype8i32 %b) {
+entry:
+  ; SSE2: shift8i32
+  ; SSE2: cost of 80 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i32
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype8i32 %a , %b
+  ret %shifttype8i32 %0
+}
+
+%shifttype16i32 = type <16 x i32>
+define %shifttype16i32 @shift16i32(%shifttype16i32 %a, %shifttype16i32 %b) {
+entry:
+  ; SSE2: shift16i32
+  ; SSE2: cost of 160 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i32
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype16i32 %a , %b
+  ret %shifttype16i32 %0
+}
+
+%shifttype32i32 = type <32 x i32>
+define %shifttype32i32 @shift32i32(%shifttype32i32 %a, %shifttype32i32 %b) {
+entry:
+  ; SSE2: shift32i32
+  ; SSE2: cost of 256 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i32
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype32i32 %a , %b
+  ret %shifttype32i32 %0
+}
+
+%shifttype2i64 = type <2 x i64>
+define %shifttype2i64 @shift2i64(%shifttype2i64 %a, %shifttype2i64 %b) {
+entry:
+  ; SSE2: shift2i64
+  ; SSE2: cost of 20 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i64
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype2i64 %a , %b
+  ret %shifttype2i64 %0
+}
+
+%shifttype4i64 = type <4 x i64>
+define %shifttype4i64 @shift4i64(%shifttype4i64 %a, %shifttype4i64 %b) {
+entry:
+  ; SSE2: shift4i64
+  ; SSE2: cost of 40 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i64
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype4i64 %a , %b
+  ret %shifttype4i64 %0
+}
+
+%shifttype8i64 = type <8 x i64>
+define %shifttype8i64 @shift8i64(%shifttype8i64 %a, %shifttype8i64 %b) {
+entry:
+  ; SSE2: shift8i64
+  ; SSE2: cost of 80 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i64
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype8i64 %a , %b
+  ret %shifttype8i64 %0
+}
+
+%shifttype16i64 = type <16 x i64>
+define %shifttype16i64 @shift16i64(%shifttype16i64 %a, %shifttype16i64 %b) {
+entry:
+  ; SSE2: shift16i64
+  ; SSE2: cost of 160 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i64
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype16i64 %a , %b
+  ret %shifttype16i64 %0
+}
+
+%shifttype32i64 = type <32 x i64>
+define %shifttype32i64 @shift32i64(%shifttype32i64 %a, %shifttype32i64 %b) {
+entry:
+  ; SSE2: shift32i64
+  ; SSE2: cost of 256 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i64
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype32i64 %a , %b
+  ret %shifttype32i64 %0
+}
+
+%shifttype2i8 = type <2 x i8>
+define %shifttype2i8 @shift2i8(%shifttype2i8 %a, %shifttype2i8 %b) {
+entry:
+  ; SSE2: shift2i8
+  ; SSE2: cost of 20 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i8
+  ; SSE2-CODEGEN: shrq %cl
+
+  %0 = lshr %shifttype2i8 %a , %b
+  ret %shifttype2i8 %0
+}
+
+%shifttype4i8 = type <4 x i8>
+define %shifttype4i8 @shift4i8(%shifttype4i8 %a, %shifttype4i8 %b) {
+entry:
+  ; SSE2: shift4i8
+  ; SSE2: cost of 40 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i8
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype4i8 %a , %b
+  ret %shifttype4i8 %0
+}
+
+%shifttype8i8 = type <8 x i8>
+define %shifttype8i8 @shift8i8(%shifttype8i8 %a, %shifttype8i8 %b) {
+entry:
+  ; SSE2: shift8i8
+  ; SSE2: cost of 80 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i8
+  ; SSE2-CODEGEN: shrl %cl
+
+  %0 = lshr %shifttype8i8 %a , %b
+  ret %shifttype8i8 %0
+}
+
+%shifttype16i8 = type <16 x i8>
+define %shifttype16i8 @shift16i8(%shifttype16i8 %a, %shifttype16i8 %b) {
+entry:
+  ; SSE2: shift16i8
+  ; SSE2: cost of 160 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i8
+  ; SSE2-CODEGEN: shrb %cl
+
+  %0 = lshr %shifttype16i8 %a , %b
+  ret %shifttype16i8 %0
+}
+
+%shifttype32i8 = type <32 x i8>
+define %shifttype32i8 @shift32i8(%shifttype32i8 %a, %shifttype32i8 %b) {
+entry:
+  ; SSE2: shift32i8
+  ; SSE2: cost of 320 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i8
+  ; SSE2-CODEGEN: shrb %cl
+
+  %0 = lshr %shifttype32i8 %a , %b
+  ret %shifttype32i8 %0
+}
+
diff --git a/llvm/test/Analysis/CostModel/X86/testshiftshl.ll b/llvm/test/Analysis/CostModel/X86/testshiftshl.ll
new file mode 100644
index 00000000000..897d9832e50
--- /dev/null
+++ b/llvm/test/Analysis/CostModel/X86/testshiftshl.ll
@@ -0,0 +1,242 @@
+; RUN: llc -mtriple=x86_64-apple-darwin -mcpu=core2 < %s | FileCheck --check-prefix=SSE2-CODEGEN %s
+; RUN: opt -mtriple=x86_64-apple-darwin -mcpu=core2 -cost-model -analyze < %s | FileCheck --check-prefix=SSE2 %s
+
+%shifttype = type <2 x i16>
+define %shifttype @shift2i16(%shifttype %a, %shifttype %b) {
+entry:
+  ; SSE2: shift2i16
+  ; SSE2: cost of 20 {{.*}} shl
+  ; SSE2-CODEGEN: shift2i16
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype %a , %b
+  ret %shifttype %0
+}
+
+%shifttype4i16 = type <4 x i16>
+define %shifttype4i16 @shift4i16(%shifttype4i16 %a, %shifttype4i16 %b) {
+entry:
+  ; SSE2: shift4i16
+  ; SSE2: cost of 10 {{.*}} shl
+  ; SSE2-CODEGEN: shift4i16
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype4i16 %a , %b
+  ret %shifttype4i16 %0
+}
+
+%shifttype8i16 = type <8 x i16>
+define %shifttype8i16 @shift8i16(%shifttype8i16 %a, %shifttype8i16 %b) {
+entry:
+  ; SSE2: shift8i16
+  ; SSE2: cost of 80 {{.*}} shl
+  ; SSE2-CODEGEN: shift8i16
+  ; SSE2-CODEGEN: shll %cl
+
+  %0 = shl %shifttype8i16 %a , %b
+  ret %shifttype8i16 %0
+}
+
+%shifttype16i16 = type <16 x i16>
+define %shifttype16i16 @shift16i16(%shifttype16i16 %a, %shifttype16i16 %b) {
+entry:
+  ; SSE2: shift16i16
+  ; SSE2: cost of 160 {{.*}} shl
+  ; SSE2-CODEGEN: shift16i16
+  ; SSE2-CODEGEN: shll %cl
+
+  %0 = shl %shifttype16i16 %a , %b
+  ret %shifttype16i16 %0
+}
+
+%shifttype32i16 = type <32 x i16>
+define %shifttype32i16 @shift32i16(%shifttype32i16 %a, %shifttype32i16 %b) {
+entry:
+  ; SSE2: shift32i16
+  ; SSE2: cost of 320 {{.*}} shl
+  ; SSE2-CODEGEN: shift32i16
+  ; SSE2-CODEGEN: shll %cl
+
+  %0 = shl %shifttype32i16 %a , %b
+  ret %shifttype32i16 %0
+}
+
+%shifttype2i32 = type <2 x i32>
+define %shifttype2i32 @shift2i32(%shifttype2i32 %a, %shifttype2i32 %b) {
+entry:
+  ; SSE2: shift2i32
+  ; SSE2: cost of 20 {{.*}} shl
+  ; SSE2-CODEGEN: shift2i32
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype2i32 %a , %b
+  ret %shifttype2i32 %0
+}
+
+%shifttype4i32 = type <4 x i32>
+define %shifttype4i32 @shift4i32(%shifttype4i32 %a, %shifttype4i32 %b) {
+entry:
+  ; SSE2: shift4i32
+  ; SSE2: cost of 10 {{.*}} shl
+  ; SSE2-CODEGEN: shift4i32
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype4i32 %a , %b
+  ret %shifttype4i32 %0
+}
+
+%shifttype8i32 = type <8 x i32>
+define %shifttype8i32 @shift8i32(%shifttype8i32 %a, %shifttype8i32 %b) {
+entry:
+  ; SSE2: shift8i32
+  ; SSE2: cost of 20 {{.*}} shl
+  ; SSE2-CODEGEN: shift8i32
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype8i32 %a , %b
+  ret %shifttype8i32 %0
+}
+
+%shifttype16i32 = type <16 x i32>
+define %shifttype16i32 @shift16i32(%shifttype16i32 %a, %shifttype16i32 %b) {
+entry:
+  ; SSE2: shift16i32
+  ; SSE2: cost of 40 {{.*}} shl
+  ; SSE2-CODEGEN: shift16i32
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype16i32 %a , %b
+  ret %shifttype16i32 %0
+}
+
+%shifttype32i32 = type <32 x i32>
+define %shifttype32i32 @shift32i32(%shifttype32i32 %a, %shifttype32i32 %b) {
+entry:
+  ; SSE2: shift32i32
+  ; SSE2: cost of 256 {{.*}} shl
+  ; SSE2-CODEGEN: shift32i32
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype32i32 %a , %b
+  ret %shifttype32i32 %0
+}
+
+%shifttype2i64 = type <2 x i64>
+define %shifttype2i64 @shift2i64(%shifttype2i64 %a, %shifttype2i64 %b) {
+entry:
+  ; SSE2: shift2i64
+  ; SSE2: cost of 20 {{.*}} shl
+  ; SSE2-CODEGEN: shift2i64
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype2i64 %a , %b
+  ret %shifttype2i64 %0
+}
+
+%shifttype4i64 = type <4 x i64>
+define %shifttype4i64 @shift4i64(%shifttype4i64 %a, %shifttype4i64 %b) {
+entry:
+  ; SSE2: shift4i64
+  ; SSE2: cost of 40 {{.*}} shl
+  ; SSE2-CODEGEN: shift4i64
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype4i64 %a , %b
+  ret %shifttype4i64 %0
+}
+
+%shifttype8i64 = type <8 x i64>
+define %shifttype8i64 @shift8i64(%shifttype8i64 %a, %shifttype8i64 %b) {
+entry:
+  ; SSE2: shift8i64
+  ; SSE2: cost of 80 {{.*}} shl
+  ; SSE2-CODEGEN: shift8i64
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype8i64 %a , %b
+  ret %shifttype8i64 %0
+}
+
+%shifttype16i64 = type <16 x i64>
+define %shifttype16i64 @shift16i64(%shifttype16i64 %a, %shifttype16i64 %b) {
+entry:
+  ; SSE2: shift16i64
+  ; SSE2: cost of 160 {{.*}} shl
+  ; SSE2-CODEGEN: shift16i64
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype16i64 %a , %b
+  ret %shifttype16i64 %0
+}
+
+%shifttype32i64 = type <32 x i64>
+define %shifttype32i64 @shift32i64(%shifttype32i64 %a, %shifttype32i64 %b) {
+entry:
+  ; SSE2: shift32i64
+  ; SSE2: cost of 256 {{.*}} shl
+  ; SSE2-CODEGEN: shift32i64
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype32i64 %a , %b
+  ret %shifttype32i64 %0
+}
+
+%shifttype2i8 = type <2 x i8>
+define %shifttype2i8 @shift2i8(%shifttype2i8 %a, %shifttype2i8 %b) {
+entry:
+  ; SSE2: shift2i8
+  ; SSE2: cost of 20 {{.*}} shl
+  ; SSE2-CODEGEN: shift2i8
+  ; SSE2-CODEGEN: shlq %cl
+
+  %0 = shl %shifttype2i8 %a , %b
+  ret %shifttype2i8 %0
+}
+
+%shifttype4i8 = type <4 x i8>
+define %shifttype4i8 @shift4i8(%shifttype4i8 %a, %shifttype4i8 %b) {
+entry:
+  ; SSE2: shift4i8
+  ; SSE2: cost of 10 {{.*}} shl
+  ; SSE2-CODEGEN: shift4i8
+  ; SSE2-CODEGEN: pmuludq
+
+  %0 = shl %shifttype4i8 %a , %b
+  ret %shifttype4i8 %0
+}
+
+%shifttype8i8 = type <8 x i8>
+define %shifttype8i8 @shift8i8(%shifttype8i8 %a, %shifttype8i8 %b) {
+entry:
+  ; SSE2: shift8i8
+  ; SSE2: cost of 80 {{.*}} shl
+  ; SSE2-CODEGEN: shift8i8
+  ; SSE2-CODEGEN: shll
+
+  %0 = shl %shifttype8i8 %a , %b
+  ret %shifttype8i8 %0
+}
+
+%shifttype16i8 = type <16 x i8>
+define %shifttype16i8 @shift16i8(%shifttype16i8 %a, %shifttype16i8 %b) {
+entry:
+  ; SSE2: shift16i8
+  ; SSE2: cost of 30 {{.*}} shl
+  ; SSE2-CODEGEN: shift16i8
+  ; SSE2-CODEGEN: cmpeqb
+
+  %0 = shl %shifttype16i8 %a , %b
+  ret %shifttype16i8 %0
+}
+
+%shifttype32i8 = type <32 x i8>
+define %shifttype32i8 @shift32i8(%shifttype32i8 %a, %shifttype32i8 %b) {
+entry:
+  ; SSE2: shift32i8
+  ; SSE2: cost of 60 {{.*}} shl
+  ; SSE2-CODEGEN: shift32i8
+  ; SSE2-CODEGEN: cmpeqb
+
+  %0 = shl %shifttype32i8 %a , %b
+  ret %shifttype32i8 %0
+}