Index: lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -108,6 +108,37 @@
   MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy;
 
   switch (ISD) {
+  case ISD::SHL:
+  case ISD::SRL:
+  case ISD::SRA: {
+    if (SLT == MVT::i64)
+      return get64BitInstrCost() * LT.first * NElts;
+
+    // i32
+    return getFullRateInstrCost() * LT.first * NElts;
+  }
+  case ISD::ADD:
+  case ISD::SUB:
+  case ISD::AND:
+  case ISD::OR:
+  case ISD::XOR: {
+    if (SLT == MVT::i64){
+      // and, or and xor are typically split into 2 VALU instructions.
+      return 2 * getFullRateInstrCost() * LT.first * NElts;
+    }
+
+    return LT.first * NElts * getFullRateInstrCost();
+  }
+  case ISD::MUL: {
+    const int QuarterRateCost = getQuarterRateInstrCost();
+    if (SLT == MVT::i64) {
+      const int FullRateCost = getFullRateInstrCost();
+      return (4 * QuarterRateCost + (2 * 2) * FullRateCost) * LT.first * NElts;
+    }
+
+    // i32
+    return QuarterRateCost * NElts * LT.first;
+  }
   case ISD::FADD:
   case ISD::FSUB:
   case ISD::FMUL:
Index: test/Analysis/CostModel/AMDGPU/add-sub.ll
===================================================================
--- /dev/null
+++ test/Analysis/CostModel/AMDGPU/add-sub.ll
@@ -0,0 +1,138 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=+half-rate-64-ops < %s | FileCheck %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck %s
+
+; CHECK: 'add_i32'
+; CHECK: estimated cost of 1 for {{.*}} add i32
+define void @add_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %add = add i32 %vec, %b
+  store i32 %add, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v2i32'
+; CHECK: estimated cost of 2 for {{.*}} add <2 x i32>
+define void @add_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %vaddr, <2 x i32> %b) #0 {
+  %vec = load <2 x i32>, <2 x i32> addrspace(1)* %vaddr
+  %add = add <2 x i32> %vec, %b
+  store <2 x i32> %add, <2 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v3i32'
+; CHECK: estimated cost of 3 for {{.*}} add <3 x i32>
+define void @add_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> addrspace(1)* %vaddr, <3 x i32> %b) #0 {
+  %vec = load <3 x i32>, <3 x i32> addrspace(1)* %vaddr
+  %add = add <3 x i32> %vec, %b
+  store <3 x i32> %add, <3 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v4i32'
+; CHECK: estimated cost of 4 for {{.*}} add <4 x i32>
+define void @add_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %vaddr, <4 x i32> %b) #0 {
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %vaddr
+  %add = add <4 x i32> %vec, %b
+  store <4 x i32> %add, <4 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_i64'
+; CHECK: estimated cost of 2 for {{.*}} add i64
+define void @add_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %add = add i64 %vec, %b
+  store i64 %add, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v2i64'
+; CHECK: estimated cost of 4 for {{.*}} add <2 x i64>
+define void @add_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> addrspace(1)* %vaddr, <2 x i64> %b) #0 {
+  %vec = load <2 x i64>, <2 x i64> addrspace(1)* %vaddr
+  %add = add <2 x i64> %vec, %b
+  store <2 x i64> %add, <2 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v3i64'
+; CHECK: estimated cost of 6 for {{.*}} add <3 x i64>
+define void @add_v3i64(<3 x i64> addrspace(1)* %out, <3 x i64> addrspace(1)* %vaddr, <3 x i64> %b) #0 {
+  %vec = load <3 x i64>, <3 x i64> addrspace(1)* %vaddr
+  %add = add <3 x i64> %vec, %b
+  store <3 x i64> %add, <3 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v4i64'
+; CHECK: estimated cost of 8 for {{.*}} add <4 x i64>
+define void @add_v4i64(<4 x i64> addrspace(1)* %out, <4 x i64> addrspace(1)* %vaddr, <4 x i64> %b) #0 {
+  %vec = load <4 x i64>, <4 x i64> addrspace(1)* %vaddr
+  %add = add <4 x i64> %vec, %b
+  store <4 x i64> %add, <4 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v16i64'
+; CHECK: estimated cost of 32 for {{.*}} add <16 x i64>
+define void @add_v16i64(<16 x i64> addrspace(1)* %out, <16 x i64> addrspace(1)* %vaddr, <16 x i64> %b) #0 {
+  %vec = load <16 x i64>, <16 x i64> addrspace(1)* %vaddr
+  %add = add <16 x i64> %vec, %b
+  store <16 x i64> %add, <16 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_i16'
+; CHECK: estimated cost of 1 for {{.*}} add i16
+define void @add_i16(i16 addrspace(1)* %out, i16 addrspace(1)* %vaddr, i16 %b) #0 {
+  %vec = load i16, i16 addrspace(1)* %vaddr
+  %add = add i16 %vec, %b
+  store i16 %add, i16 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'add_v2i16'
+; CHECK: estimated cost of 2 for {{.*}} add <2 x i16>
+define void @add_v2i16(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %vaddr, <2 x i16> %b) #0 {
+  %vec = load <2 x i16>, <2 x i16> addrspace(1)* %vaddr
+  %add = add <2 x i16> %vec, %b
+  store <2 x i16> %add, <2 x i16> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'sub_i32'
+; CHECK: estimated cost of 1 for {{.*}} sub i32
+define void @sub_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %sub = sub i32 %vec, %b
+  store i32 %sub, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'sub_i64'
+; CHECK: estimated cost of 2 for {{.*}} sub i64
+define void @sub_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %sub = sub i64 %vec, %b
+  store i64 %sub, i64 addrspace(1)* %out
+  ret void
+}
+; CHECK: 'sub_i16'
+; CHECK: estimated cost of 1 for {{.*}} sub i16
+define void @sub_i16(i16 addrspace(1)* %out, i16 addrspace(1)* %vaddr, i16 %b) #0 {
+  %vec = load i16, i16 addrspace(1)* %vaddr
+  %sub = sub i16 %vec, %b
+  store i16 %sub, i16 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'sub_v2i16'
+; CHECK: estimated cost of 2 for {{.*}} sub <2 x i16>
+define void @sub_v2i16(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %vaddr, <2 x i16> %b) #0 {
+  %vec = load <2 x i16>, <2 x i16> addrspace(1)* %vaddr
+  %sub = sub <2 x i16> %vec, %b
+  store <2 x i16> %sub, <2 x i16> addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
Index: test/Analysis/CostModel/AMDGPU/add.ll
===================================================================
--- test/Analysis/CostModel/AMDGPU/add.ll
+++ /dev/null
@@ -1,56 +0,0 @@
-; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
-
-; CHECK: 'add_i32'
-; CHECK: estimated cost of 1 for {{.*}} add i32
-define void @add_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
-  %vec = load i32, i32 addrspace(1)* %vaddr
-  %add = add i32 %vec, %b
-  store i32 %add, i32 addrspace(1)* %out
-  ret void
-}
-
-; CHECK: 'add_v2i32'
-; CHECK: estimated cost of 2 for {{.*}} add <2 x i32>
-define void @add_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %vaddr, <2 x i32> %b) #0 {
-  %vec = load <2 x i32>, <2 x i32> addrspace(1)* %vaddr
-  %add = add <2 x i32> %vec, %b
-  store <2 x i32> %add, <2 x i32> addrspace(1)* %out
-  ret void
-}
-
-; CHECK: 'add_v3i32'
-; CHECK: estimated cost of 3 for {{.*}} add <3 x i32>
-define void @add_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> addrspace(1)* %vaddr, <3 x i32> %b) #0 {
-  %vec = load <3 x i32>, <3 x i32> addrspace(1)* %vaddr
-  %add = add <3 x i32> %vec, %b
-  store <3 x i32> %add, <3 x i32> addrspace(1)* %out
-  ret void
-}
-
-; CHECK: 'add_i64'
-; CHECK: estimated cost of 1 for {{.*}} add i64
-define void @add_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
-  %vec = load i64, i64 addrspace(1)* %vaddr
-  %add = add i64 %vec, %b
-  store i64 %add, i64 addrspace(1)* %out
-  ret void
-}
-
-; CHECK: 'add_v2i64'
-; CHECK: estimated cost of 2 for {{.*}} add <2 x i64>
-define void @add_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> addrspace(1)* %vaddr, <2 x i64> %b) #0 {
-  %vec = load <2 x i64>, <2 x i64> addrspace(1)* %vaddr
-  %add = add <2 x i64> %vec, %b
-  store <2 x i64> %add, <2 x i64> addrspace(1)* %out
-  ret void
-}
-
-; CHECK: 'add_v3i64'
-; CHECK: estimated cost of 3 for {{.*}} add <3 x i64>
-define void @add_v3i64(<3 x i64> addrspace(1)* %out, <3 x i64> addrspace(1)* %vaddr, <3 x i64> %b) #0 {
-  %vec = load <3 x i64>, <3 x i64> addrspace(1)* %vaddr
-  %add = add <3 x i64> %vec, %b
-  store <3 x i64> %add, <3 x i64> addrspace(1)* %out
-  ret void
-}
-
Index: test/Analysis/CostModel/AMDGPU/bit-ops.ll
===================================================================
--- /dev/null
+++ test/Analysis/CostModel/AMDGPU/bit-ops.ll
@@ -0,0 +1,59 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+
+; CHECK: 'or_i32'
+; CHECK: estimated cost of 1 for {{.*}} or i32
+define void @or_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = or i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'or_i64'
+; CHECK: estimated cost of 2 for {{.*}} or i64
+define void @or_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = or i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'xor_i32'
+; CHECK: estimated cost of 1 for {{.*}} xor i32
+define void @xor_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = xor i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'xor_i64'
+; CHECK: estimated cost of 2 for {{.*}} xor i64
+define void @xor_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = xor i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+
+; CHECK: 'and_i32'
+; CHECK: estimated cost of 1 for {{.*}} and i32
+define void @and_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = and i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'and_i64'
+; CHECK: estimated cost of 2 for {{.*}} and i64
+define void @and_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = and i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+
+attributes #0 = { nounwind }
Index: test/Analysis/CostModel/AMDGPU/mul.ll
===================================================================
--- /dev/null
+++ test/Analysis/CostModel/AMDGPU/mul.ll
@@ -0,0 +1,85 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck %s
+
+; CHECK: 'mul_i32'
+; CHECK: estimated cost of 3 for {{.*}} mul i32
+define void @mul_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %mul = mul i32 %vec, %b
+  store i32 %mul, i32 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v2i32'
+; CHECK: estimated cost of 6 for {{.*}} mul <2 x i32>
+define void @mul_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %vaddr, <2 x i32> %b) #0 {
+  %vec = load <2 x i32>, <2 x i32> addrspace(1)* %vaddr
+  %mul = mul <2 x i32> %vec, %b
+  store <2 x i32> %mul, <2 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v3i32'
+; CHECK: estimated cost of 9 for {{.*}} mul <3 x i32>
+define void @mul_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> addrspace(1)* %vaddr, <3 x i32> %b) #0 {
+  %vec = load <3 x i32>, <3 x i32> addrspace(1)* %vaddr
+  %mul = mul <3 x i32> %vec, %b
+  store <3 x i32> %mul, <3 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v4i32'
+; CHECK: estimated cost of 12 for {{.*}} mul <4 x i32>
+define void @mul_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %vaddr, <4 x i32> %b) #0 {
+  %vec = load <4 x i32>, <4 x i32> addrspace(1)* %vaddr
+  %mul = mul <4 x i32> %vec, %b
+  store <4 x i32> %mul, <4 x i32> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_i64'
+; CHECK: estimated cost of 16 for {{.*}} mul i64
+define void @mul_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %mul = mul i64 %vec, %b
+  store i64 %mul, i64 addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v2i64'
+; CHECK: estimated cost of 32 for {{.*}} mul <2 x i64>
+define void @mul_v2i64(<2 x i64> addrspace(1)* %out, <2 x i64> addrspace(1)* %vaddr, <2 x i64> %b) #0 {
+  %vec = load <2 x i64>, <2 x i64> addrspace(1)* %vaddr
+  %mul = mul <2 x i64> %vec, %b
+  store <2 x i64> %mul, <2 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v3i64'
+; CHECK: estimated cost of 48 for {{.*}} mul <3 x i64>
+define void @mul_v3i64(<3 x i64> addrspace(1)* %out, <3 x i64> addrspace(1)* %vaddr, <3 x i64> %b) #0 {
+  %vec = load <3 x i64>, <3 x i64> addrspace(1)* %vaddr
+  %mul = mul <3 x i64> %vec, %b
+  store <3 x i64> %mul, <3 x i64> addrspace(1)* %out
+  ret void
+}
+
+; CHECK: 'mul_v4i64'
+; CHECK: estimated cost of 64 for {{.*}} mul <4 x i64>
+define void @mul_v4i64(<4 x i64> addrspace(1)* %out, <4 x i64> addrspace(1)* %vaddr, <4 x i64> %b) #0 {
+  %vec = load <4 x i64>, <4 x i64> addrspace(1)* %vaddr
+  %mul = mul <4 x i64> %vec, %b
+  store <4 x i64> %mul, <4 x i64> addrspace(1)* %out
+  ret void
+}
+
+
+; CHECK: 'mul_v8i64'
+; CHECK: estimated cost of 128 for {{.*}} mul <8 x i64>
+define void @mul_v8i64(<8 x i64> addrspace(1)* %out, <8 x i64> addrspace(1)* %vaddr, <8 x i64> %b) #0 {
+  %vec = load <8 x i64>, <8 x i64> addrspace(1)* %vaddr
+  %mul = mul <8 x i64> %vec, %b
+  store <8 x i64> %mul, <8 x i64> addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
Index: test/Analysis/CostModel/AMDGPU/shifts.ll
===================================================================
--- /dev/null
+++ test/Analysis/CostModel/AMDGPU/shifts.ll
@@ -0,0 +1,61 @@
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=+half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=FAST64 %s
+; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefix=ALL -check-prefix=SLOW64 %s
+
+; ALL: 'shl_i32'
+; ALL: estimated cost of 1 for {{.*}} shl i32
+define void @shl_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = shl i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'shl_i64'
+; FAST64: estimated cost of 2 for {{.*}} shl i64
+; SLOW64: estimated cost of 3 for {{.*}} shl i64
+define void @shl_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = shl i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'lshr_i32'
+; ALL: estimated cost of 1 for {{.*}} lshr i32
+define void @lshr_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = lshr i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'lshr_i64'
+; FAST64: estimated cost of 2 for {{.*}} lshr i64
+; SLOW64: estimated cost of 3 for {{.*}} lshr i64
+define void @lshr_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = lshr i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'ashr_i32'
+; ALL: estimated cost of 1 for {{.*}} ashr i32
+define void @ashr_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %vaddr, i32 %b) #0 {
+  %vec = load i32, i32 addrspace(1)* %vaddr
+  %or = ashr i32 %vec, %b
+  store i32 %or, i32 addrspace(1)* %out
+  ret void
+}
+
+; ALL: 'ashr_i64'
+; FAST64: estimated cost of 2 for {{.*}} ashr i64
+; SLOW64: estimated cost of 3 for {{.*}} ashr i64
+define void @ashr_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %vaddr, i64 %b) #0 {
+  %vec = load i64, i64 addrspace(1)* %vaddr
+  %or = ashr i64 %vec, %b
+  store i64 %or, i64 addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
Index: test/Transforms/IndVarSimplify/AMDGPU/lit.local.cfg
===================================================================
--- /dev/null
+++ test/Transforms/IndVarSimplify/AMDGPU/lit.local.cfg
@@ -0,0 +1,2 @@
+if not 'AMDGPU' in config.root.targets:
+    config.unsupported = True
Index: test/Transforms/IndVarSimplify/AMDGPU/no-widen-to-i64.ll
===================================================================
--- /dev/null
+++ test/Transforms/IndVarSimplify/AMDGPU/no-widen-to-i64.ll
@@ -0,0 +1,98 @@
+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -indvars %s | FileCheck %s
+
+; Bug 21148
+
+; Induction variables should not be widened for 64-bit integers,
+; despite being a legal type.
+;
+; The cost of basic arithmetic instructions on a 64-bit integer are
+; twice as expensive as that on a 32-bit integer, or split into 2
+; 32-bit components.
+
+target datalayout = "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64"
+
+; CHECK-LABEL: @indvar_32_bit(
+; CHECK-NOT: sext i32
+; CHECK: phi i32
+define void @indvar_32_bit(i32 %n, i32* nocapture %output) {
+entry:
+  %cmp5 = icmp sgt i32 %n, 0
+  br i1 %cmp5, label %for.body.preheader, label %for.end
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.06 = phi i32 [ 0, %for.body.preheader ], [ %add, %for.body ]
+  %mul = mul nsw i32 %i.06, %i.06
+  %tmp0 = sext i32 %i.06 to i64
+  %arrayidx = getelementptr inbounds i32, i32* %output, i64 %tmp0
+  store i32 %mul, i32* %arrayidx, align 4
+  %add = add nsw i32 %i.06, 3
+  %cmp = icmp slt i32 %add, %n
+  br i1 %cmp, label %for.body, label %for.end.loopexit
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
+; CHECK-LABEL: @no_promote_i32(
+; CHECK-NOT: sext i32
+; CHECK: br
+; CHECK-NOT: shl i64
+; CHECK-NOT: ashr i64
+; CHECK-NOT: mul nsw i64
+; CHECK-NOT: add nsw i64
+define void @no_promote_i32(i32 addrspace(1)* %out, i32 %a, i32 %b) {
+entry:
+  br label %for.body
+
+for.body:
+  %inc = phi i32 [ 0, %entry ], [ %inc.i, %for.body ]
+  %tmp0 = add i32 %a, %inc
+  %shl = shl i32 %inc, 8
+  %shr = ashr exact i32 %shl, 8
+  %mul = mul nsw i32 %shr, %a
+  %add = add nsw i32 %mul, %b
+  %tmp1 = sext i32 %add to i64
+  %arrayidx1 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 %tmp1
+  store i32 %tmp0, i32 addrspace(1)* %arrayidx1, align 4
+  %inc.i = add nsw i32 %inc, 1
+  %cmp = icmp slt i32 %inc.i, 16
+  br i1 %cmp, label %for.body, label %for.end
+
+for.end:
+  ret void
+}
+
+; FIXME: This should really be promoted to i64, since it will need to
+; be legalized anyway.
+
+; CHECK-LABEL: @indvar_48_bit(
+define void @indvar_48_bit(i48 %n, i48* nocapture %output) {
+entry:
+  %cmp5 = icmp sgt i48 %n, 0
+  br i1 %cmp5, label %for.body.preheader, label %for.end
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.06 = phi i48 [ 0, %for.body.preheader ], [ %add, %for.body ]
+  %mul = mul nsw i48 %i.06, %i.06
+  %tmp0 = sext i48 %i.06 to i64
+  %arrayidx = getelementptr inbounds i48, i48* %output, i64 %tmp0
+  store i48 %mul, i48* %arrayidx, align 4
+  %add = add nsw i48 %i.06, 3
+  %cmp = icmp slt i48 %add, %n
+  br i1 %cmp, label %for.body, label %for.end.loopexit
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}