diff --git a/llvm/test/CodeGen/BPF/32-bit-subreg-alu.ll b/llvm/test/CodeGen/BPF/32-bit-subreg-alu.ll
new file mode 100644
index 000000000000..2a498494589a
--- /dev/null
+++ b/llvm/test/CodeGen/BPF/32-bit-subreg-alu.ll
@@ -0,0 +1,298 @@
+; RUN: llc -O2 -march=bpfel -mattr=+alu32 < %s | FileCheck %s
+;
+; int mov(int a)
+; {
+;   return a;
+; }
+;
+; int mov_ri(void)
+; {
+;   return 0xff;
+; }
+;
+; int add(int a, int b)
+; {
+;   return a + b;
+; }
+;
+; int add_i(int a)
+; {
+;   return a + 0x7fffffff;
+; }
+;
+; int sub(int a, int b)
+; {
+;   return a - b;
+; }
+;
+; int sub_i(int a)
+; {
+;   return a - 0xffffffff;
+; }
+;
+; int mul(int a, int b)
+; {
+;   return a * b;
+; }
+;
+; int mul_i(int a)
+; {
+;   return a * 0xf;
+; }
+;
+; unsigned div(unsigned a, unsigned b)
+; {
+;   return a / b;
+; }
+;
+; unsigned div_i(unsigned a)
+; {
+;   return a / 0xf;
+; }
+;
+; int or(int a, int b)
+; {
+;   return a | b;
+; }
+;
+; int or_i(int a)
+; {
+;   return a | 0xff;
+; }
+;
+; int xor(int a, int b)
+; {
+;   return a ^ b;
+; }
+;
+; int xor_i(int a)
+; {
+;   return a ^ 0xfff;
+; }
+;
+; int and(int a, int b)
+; {
+;   return a & b;
+; }
+;
+; int and_i(int a)
+; {
+;   return a & 0xffff;
+; }
+;
+; int sll(int a, int b)
+; {
+;   return a << b;
+; }
+;
+; int sll_i(int a)
+; {
+;   return a << 17;
+; }
+;
+; unsigned srl(unsigned a, unsigned b)
+; {
+;   return a >> b;
+; }
+;
+; unsigned srl_i(unsigned a, unsigned b)
+; {
+;   return a >> 31;
+; }
+;
+; int sra(int a, int b)
+; {
+;   return a >> b;
+; }
+;
+; int sra_i(int a, int b)
+; {
+;   return a >> 7;
+; }
+;
+; int neg(int a)
+; {
+;   return -a;
+; }
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @mov(i32 returned %a) local_unnamed_addr #0 {
+entry:
+  ret i32 %a
+; CHECK: w{{[0-9]+}} = w{{[0-9]+}}
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @mov_ri() local_unnamed_addr #0 {
+entry:
+  ret i32 255
+; CHECK: w{{[0-9]+}} = 255
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @add(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %add = add nsw i32 %b, %a
+; CHECK: w{{[0-9]+}} += w{{[0-9]+}}
+  ret i32 %add
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @add_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %add = add nsw i32 %a, 2147483647
+; CHECK: w{{[0-9]+}} += 2147483647
+  ret i32 %add
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sub(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %sub = sub nsw i32 %a, %b
+; CHECK: w{{[0-9]+}} -= w{{[0-9]+}}
+  ret i32 %sub
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sub_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %sub = add i32 %a, 1
+; CHECK: w{{[0-9]+}} += 1
+  ret i32 %sub
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @mul(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %mul = mul nsw i32 %b, %a
+; CHECK: w{{[0-9]+}} *= w{{[0-9]+}}
+  ret i32 %mul
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @mul_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %mul = mul nsw i32 %a, 15
+; CHECK: w{{[0-9]+}} *= 15
+  ret i32 %mul
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @div(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %div = udiv i32 %a, %b
+; CHECK: w{{[0-9]+}} /= w{{[0-9]+}}
+  ret i32 %div
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @div_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %div = udiv i32 %a, 15
+; CHECK: w{{[0-9]+}} /= 15
+  ret i32 %div
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @or(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %or = or i32 %b, %a
+; CHECK: w{{[0-9]+}} |= w{{[0-9]+}}
+  ret i32 %or
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @or_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %or = or i32 %a, 255
+; CHECK: w{{[0-9]+}} |= 255
+  ret i32 %or
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @xor(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %xor = xor i32 %b, %a
+; CHECK: w{{[0-9]+}} ^= w{{[0-9]+}}
+  ret i32 %xor
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @xor_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %xor = xor i32 %a, 4095
+; CHECK: w{{[0-9]+}} ^= 4095
+  ret i32 %xor
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @and(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %and = and i32 %b, %a
+; CHECK: w{{[0-9]+}} &= w{{[0-9]+}}
+  ret i32 %and
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @and_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %and = and i32 %a, 65535
+; CHECK: w{{[0-9]+}} &= 65535
+  ret i32 %and
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sll(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %shl = shl i32 %a, %b
+; CHECK: w{{[0-9]+}} <<= w{{[0-9]+}}
+  ret i32 %shl
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sll_i(i32 %a) local_unnamed_addr #0 {
+entry:
+  %shl = shl i32 %a, 17
+; CHECK: w{{[0-9]+}} <<= 17
+  ret i32 %shl
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @srl(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %shr = lshr i32 %a, %b
+; CHECK: w{{[0-9]+}} >>= w{{[0-9]+}}
+  ret i32 %shr
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @srl_i(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %shr = lshr i32 %a, 31
+; CHECK: w{{[0-9]+}} >>= 31
+  ret i32 %shr
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sra(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %shr = ashr i32 %a, %b
+; CHECK: w{{[0-9]+}} s>>= w{{[0-9]+}}
+  ret i32 %shr
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @sra_i(i32 %a, i32 %b) local_unnamed_addr #0 {
+entry:
+  %shr = ashr i32 %a, 7
+; CHECK: w{{[0-9]+}} s>>= 7
+  ret i32 %shr
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @neg(i32 %a) local_unnamed_addr #0 {
+entry:
+  %sub = sub nsw i32 0, %a
+; CHECK: w{{[0-9]+}} = -w{{[0-9]+}}
+  ret i32 %sub
+}
diff --git a/llvm/test/CodeGen/BPF/32-bit-subreg-cond-select.ll b/llvm/test/CodeGen/BPF/32-bit-subreg-cond-select.ll
new file mode 100644
index 000000000000..a48141cf0404
--- /dev/null
+++ b/llvm/test/CodeGen/BPF/32-bit-subreg-cond-select.ll
@@ -0,0 +1,100 @@
+; RUN: llc -O2 -march=bpfel -mattr=+alu32 < %s | FileCheck %s
+;
+; unsigned int select_cc_32 (unsigned a, unsigned b, int c, int d)
+; {
+;   if (a > b)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; long long select_cc_32_64 (unsigned a, unsigned b, long long c, long long d)
+; {
+;   if (a > b)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; int select_cc_64_32 (long long a, long long b, int c, int d)
+; {
+;   if (a > b)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; int selecti_cc_32 (unsigned a, int c, int d)
+; {
+;   if (a > 10)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; long long selecti_cc_32_64 (unsigned a, long long c, long long d)
+; {
+;   if (a > 11)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; int selecti_cc_64_32 (long long a, int c, int d)
+; {
+;   if (a > 12)
+;     return c;
+;   else
+;     return d;
+; }
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @select_cc_32(i32 %a, i32 %b, i32 %c, i32 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp ugt i32 %a, %b
+  %c.d = select i1 %cmp, i32 %c, i32 %d
+  ret i32 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i64 @select_cc_32_64(i32 %a, i32 %b, i64 %c, i64 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp ugt i32 %a, %b
+  %c.d = select i1 %cmp, i64 %c, i64 %d
+  ret i64 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @select_cc_64_32(i64 %a, i64 %b, i32 %c, i32 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp sgt i64 %a, %b
+  %c.d = select i1 %cmp, i32 %c, i32 %d
+  ret i32 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @selecti_cc_32(i32 %a, i32 %c, i32 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp ugt i32 %a, 10
+  %c.d = select i1 %cmp, i32 %c, i32 %d
+  ret i32 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i64 @selecti_cc_32_64(i32 %a, i64 %c, i64 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp ugt i32 %a, 11
+  %c.d = select i1 %cmp, i64 %c, i64 %d
+  ret i64 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i32 @selecti_cc_64_32(i64 %a, i32 %c, i32 %d) local_unnamed_addr #0 {
+entry:
+  %cmp = icmp sgt i64 %a, 12
+  %c.d = select i1 %cmp, i32 %c, i32 %d
+  ret i32 %c.d
+}
+; There shouldn't be any type promotion, all of them are expected to be
+; eliminated by peephole optimization.
+; CHECK-NOT: r{{[0-9]+}} <<= 32
diff --git a/llvm/test/CodeGen/BPF/32-bit-subreg-load-store.ll b/llvm/test/CodeGen/BPF/32-bit-subreg-load-store.ll
new file mode 100644
index 000000000000..9906cee35dde
--- /dev/null
+++ b/llvm/test/CodeGen/BPF/32-bit-subreg-load-store.ll
@@ -0,0 +1,107 @@
+; RUN: llc -O2 -march=bpfel -mattr=+alu32 < %s | FileCheck %s
+;
+; unsigned char loadu8(unsigned char *p)
+; {
+;   return *p;
+; }
+;
+; unsigned short loadu16(unsigned short *p)
+; {
+;   return *p;
+; }
+;
+; unsigned loadu32(unsigned *p)
+; {
+;   return *p;
+; }
+;
+; unsigned long long loadu64(unsigned long long *p)
+; {
+;   return *p;
+; }
+;
+; void storeu8(unsigned char *p, unsigned long long v)
+; {
+;   *p = (unsigned char)v;
+; }
+;
+; void storeu16(unsigned short *p, unsigned long long v)
+; {
+;   *p = (unsigned short)v;
+; }
+;
+; void storeu32(unsigned *p, unsigned long long v)
+; {
+;   *p = (unsigned)v;
+; }
+;
+; void storeu64(unsigned long long *p, unsigned long long v)
+; {
+;   *p = v;
+; }
+; Function Attrs: norecurse nounwind readonly
+define dso_local zeroext i8 @loadu8(i8* nocapture readonly %p) local_unnamed_addr #0 {
+entry:
+  %0 = load i8, i8* %p, align 1
+; CHECK: w{{[0-9]+}} = *(u8 *)(r{{[0-9]+}} + 0)
+  ret i8 %0
+}
+
+; Function Attrs: norecurse nounwind readonly
+define dso_local zeroext i16 @loadu16(i16* nocapture readonly %p) local_unnamed_addr #0 {
+entry:
+  %0 = load i16, i16* %p, align 2
+; CHECK: w{{[0-9]+}} = *(u16 *)(r{{[0-9]+}} + 0)
+  ret i16 %0
+}
+
+; Function Attrs: norecurse nounwind readonly
+define dso_local i32 @loadu32(i32* nocapture readonly %p) local_unnamed_addr #0 {
+entry:
+  %0 = load i32, i32* %p, align 4
+; CHECK: w{{[0-9]+}} = *(u32 *)(r{{[0-9]+}} + 0)
+  ret i32 %0
+}
+
+; Function Attrs: norecurse nounwind readonly
+define dso_local i64 @loadu64(i64* nocapture readonly %p) local_unnamed_addr #0 {
+entry:
+  %0 = load i64, i64* %p, align 8
+; CHECK: r{{[0-9]+}} = *(u64 *)(r{{[0-9]+}} + 0)
+  ret i64 %0
+}
+
+; Function Attrs: norecurse nounwind
+define dso_local void @storeu8(i8* nocapture %p, i64 %v) local_unnamed_addr #1 {
+entry:
+  %conv = trunc i64 %v to i8
+  store i8 %conv, i8* %p, align 1
+; CHECK: *(u8 *)(r{{[0-9]+}} + 0) = w{{[0-9]+}}
+  ret void
+}
+
+; Function Attrs: norecurse nounwind
+define dso_local void @storeu16(i16* nocapture %p, i64 %v) local_unnamed_addr #1 {
+entry:
+  %conv = trunc i64 %v to i16
+  store i16 %conv, i16* %p, align 2
+; CHECK: *(u16 *)(r{{[0-9]+}} + 0) = w{{[0-9]+}}
+  ret void
+}
+
+; Function Attrs: norecurse nounwind
+define dso_local void @storeu32(i32* nocapture %p, i64 %v) local_unnamed_addr #1 {
+entry:
+  %conv = trunc i64 %v to i32
+  store i32 %conv, i32* %p, align 4
+; CHECK: *(u32 *)(r{{[0-9]+}} + 0) = w{{[0-9]+}}
+  ret void
+}
+
+; Function Attrs: norecurse nounwind
+define dso_local void @storeu64(i64* nocapture %p, i64 %v) local_unnamed_addr #1 {
+entry:
+  store i64 %v, i64* %p, align 8
+; CHECK: *(u64 *)(r{{[0-9]+}} + 0) = r{{[0-9]+}}
+  ret void
+}
diff --git a/llvm/test/CodeGen/BPF/32-bit-subreg-peephole.ll b/llvm/test/CodeGen/BPF/32-bit-subreg-peephole.ll
new file mode 100644
index 000000000000..8ec74ca3bf22
--- /dev/null
+++ b/llvm/test/CodeGen/BPF/32-bit-subreg-peephole.ll
@@ -0,0 +1,36 @@
+; RUN: llc -O2 -march=bpfel -mattr=+alu32 < %s | FileCheck %s
+;
+; long long select_u(unsigned a, unsigned b, long long c, long long d)
+; {
+;   if (a > b)
+;     return c;
+;   else
+;     return d;
+; }
+;
+; long long select_s(signed a, signed b, long long c, long long d)
+; {
+;   if (a > b)
+;     return c;
+;   else
+;     return d;
+;}
+; Function Attrs: norecurse nounwind readnone
+define dso_local i64 @select_u(i32 %a, i32 %b, i64 %c, i64 %d) local_unnamed_addr #0 {
+; CHECK-LABEL: select_u:
+entry:
+  %cmp = icmp ugt i32 %a, %b
+  %c.d = select i1 %cmp, i64 %c, i64 %d
+; CHECK: if r{{[0-9]+}} {{<|>}} r{{[0-9]+}} goto
+  ret i64 %c.d
+}
+
+; Function Attrs: norecurse nounwind readnone
+define dso_local i64 @select_s(i32 %a, i32 %b, i64 %c, i64 %d) local_unnamed_addr #0 {
+; CHECK-LABEL: select_s:
+entry:
+  %cmp = icmp sgt i32 %a, %b
+  %c.d = select i1 %cmp, i64 %c, i64 %d
+; CHECK: if r{{[0-9]+}} s{{<|>}} r{{[0-9]+}} goto
+  ret i64 %c.d
+}