Index: lib/Target/WebAssembly/WebAssemblyInstrMemory.td
===================================================================
--- lib/Target/WebAssembly/WebAssemblyInstrMemory.td
+++ lib/Target/WebAssembly/WebAssemblyInstrMemory.td
@@ -25,9 +25,9 @@
 let Defs = [ARGUMENTS] in {
 
 // Basic load.
-def LOAD_I32 : I<(outs I32:$dst), (ins I32:$addr),
-                 [(set I32:$dst, (load I32:$addr))],
-                 "i32.load\t$dst, $addr">;
+def LOAD_I32 : I<(outs I32:$dst), (ins I32:$off, I32:$addr),
+                 [(set I32:$dst, (i32 (load (add I32:$addr, (i32 imm:$off)))))],
+                 "i32.load\t$dst, $off($addr)">;
 def LOAD_I64 : I<(outs I64:$dst), (ins I32:$addr),
                  [(set I64:$dst, (load I32:$addr))],
                  "i64.load\t$dst, $addr">;
@@ -72,6 +72,9 @@
 
 } // Defs = [ARGUMENTS]
 
+// Select loads with no constant offset.
+def : Pat<(i32 (load I32:$addr)), (LOAD_I32 0, $addr)>;
+
 // "Don't care" extending load become zero-extending load.
 def : Pat<(i32 (extloadi8 I32:$addr)),  (LOAD8_U_I32 $addr)>;
 def : Pat<(i32 (extloadi16 I32:$addr)), (LOAD16_U_I32 $addr)>;
@@ -87,8 +90,8 @@
 // instruction definition patterns that don't reference all of the output
 // operands.
 // Note: WebAssembly inverts SelectionDAG's usual operand order.
-def STORE_I32  : I<(outs I32:$dst), (ins I32:$addr, I32:$val), [],
-                   "i32.store\t$dst, $addr, $val">;
+def STORE_I32  : I<(outs I32:$dst), (ins I32:$off, I32:$addr, I32:$val), [],
+                   "i32.store\t$dst, $off($addr), $val">;
 def STORE_I64  : I<(outs I64:$dst), (ins I32:$addr, I64:$val), [],
                    "i64.store\t$dst, $addr, $val">;
 def STORE_F32  : I<(outs F32:$dst), (ins I32:$addr, F32:$val), [],
@@ -98,7 +101,11 @@
 
 } // Defs = [ARGUMENTS]
 
-def : Pat<(store I32:$val, I32:$addr), (STORE_I32 I32:$addr, I32:$val)>;
+def : Pat<(store I32:$val, I32:$addr), (STORE_I32 0, I32:$addr, I32:$val)>;
+def : Pat<(store I32:$val, (add I32:$addr, (i32 imm:$off))),
+          (STORE_I32 imm:$off, I32:$addr, I32:$val)>;
+
+
 def : Pat<(store I64:$val, I32:$addr), (STORE_I64 I32:$addr, I64:$val)>;
 def : Pat<(store F32:$val, I32:$addr), (STORE_F32 I32:$addr, F32:$val)>;
 def : Pat<(store F64:$val, I32:$addr), (STORE_F64 I32:$addr, F64:$val)>;
Index: lib/Target/WebAssembly/WebAssemblyStoreResults.cpp
===================================================================
--- lib/Target/WebAssembly/WebAssemblyStoreResults.cpp
+++ lib/Target/WebAssembly/WebAssemblyStoreResults.cpp
@@ -85,7 +85,7 @@
       case WebAssembly::STORE_I32:
       case WebAssembly::STORE_I64:
         unsigned ToReg = MI.getOperand(0).getReg();
-        unsigned FromReg = MI.getOperand(2).getReg();
+        unsigned FromReg = MI.getOperand(3).getReg();
         for (auto I = MRI.use_begin(FromReg), E = MRI.use_end(); I != E;) {
           MachineOperand &O = *I++;
           MachineInstr *Where = O.getParent();
Index: test/CodeGen/WebAssembly/cfg-stackify.ll
===================================================================
--- test/CodeGen/WebAssembly/cfg-stackify.ll
+++ test/CodeGen/WebAssembly/cfg-stackify.ll
@@ -184,7 +184,7 @@
 ; CHECK-LABEL: minimal_loop:
 ; CHECK-NOT: br
 ; CHECK: BB7_1:
-; CHECK: i32.store $discard=, $0, $pop{{[0-9]+}}{{$}}
+; CHECK: i32.store $discard=, 0($0), $pop{{[0-9]+}}{{$}}
 ; CHECK: br BB7_1{{$}}
 ; CHECK: BB7_2:
 define i32 @minimal_loop(i32* %p) {
Index: test/CodeGen/WebAssembly/global.ll
===================================================================
--- test/CodeGen/WebAssembly/global.ll
+++ test/CodeGen/WebAssembly/global.ll
@@ -11,7 +11,7 @@
 
 ; CHECK: foo:
 ; CHECK: i32.const $push0=, answer{{$}}
-; CHECK-NEXT: i32.load $push1=, $pop0{{$}}
+; CHECK-NEXT: i32.load $push1=, 0($pop0){{$}}
 ; CHECK-NEXT: return $pop1{{$}}
 define i32 @foo() {
   %a = load i32, i32* @answer
Index: test/CodeGen/WebAssembly/load.ll
===================================================================
--- test/CodeGen/WebAssembly/load.ll
+++ test/CodeGen/WebAssembly/load.ll
@@ -8,7 +8,7 @@
 ; CHECK-LABEL: ldi32:
 ; CHECK-NEXT: .param i32{{$}}
 ; CHECK-NEXT: .result i32{{$}}
-; CHECK-NEXT: i32.load $push[[NUM:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: i32.load $push[[NUM:[0-9]+]]=, 0($0){{$}}
 ; CHECK-NEXT: return $pop[[NUM]]{{$}}
 define i32 @ldi32(i32 *%p) {
   %v = load i32, i32* %p
@@ -44,3 +44,21 @@
   %v = load double, double* %p
   ret double %v
 }
+
+; Check that a constant offset is folded into a load
+; CHECK-LABEL: ldoff:
+; CHECK: i32.load $push0=, 12($0)
+define i32 @ldoff(i32 *%p) {
+  %addr = getelementptr i32, i32* %p, i32 3
+  %v = load i32, i32* %addr
+  ret i32 %v
+}
+
+; Check that a variable offset is not folded
+; CHECK-LABEL: ldoff_var
+; CHECK: i32.load $push{{[0-9]+}}=, 0($pop{{[0-9]}})
+define i32 @ldoff_var(i32 *%p, i32 %o) {
+  %addr = getelementptr i32, i32* %p, i32 %o
+  %v = load i32, i32* %addr
+  ret i32 %v
+}
\ No newline at end of file
Index: test/CodeGen/WebAssembly/store-results.ll
===================================================================
--- test/CodeGen/WebAssembly/store-results.ll
+++ test/CodeGen/WebAssembly/store-results.ll
@@ -9,7 +9,7 @@
 ; CHECK-LABEL: single_block:
 ; CHECK-NOT: .local
 ; CHECK: i32.const $push{{[0-9]+}}=, 0
-; CHECK: i32.store $push[[STORE:[0-9]+]]=, $0, $pop{{[0-9]+}}
+; CHECK: i32.store $push[[STORE:[0-9]+]]=, 0($0), $pop{{[0-9]+}}
 ; CHECK: return $pop[[STORE]]{{$}}
 define i32 @single_block(i32* %p) {
 entry:
Index: test/CodeGen/WebAssembly/store.ll
===================================================================
--- test/CodeGen/WebAssembly/store.ll
+++ test/CodeGen/WebAssembly/store.ll
@@ -7,7 +7,7 @@
 
 ; CHECK-LABEL: sti32:
 ; CHECK-NEXT: .param i32, i32{{$}}
-; CHECK-NEXT: i32.store $discard=, $0, $1{{$}}
+; CHECK-NEXT: i32.store $discard=, 0($0), $1{{$}}
 ; CHECK-NEXT: return{{$}}
 define void @sti32(i32 *%p, i32 %v) {
   store i32 %v, i32* %p
@@ -40,3 +40,12 @@
   store double %v, double* %p
   ret void
 }
+
+; Check that a constant offset is folded into a store
+; CHECK-LABEL: stoff:
+; CHECK: i32.store $discard=, 12($0), $1
+define void @stoff(i32 *%p, i32 %v) {
+  %addr = getelementptr i32, i32* %p, i32 3
+  store i32 %v, i32* %addr
+  ret void
+}