diff --git a/.gitmodules b/.gitmodules
new file mode 100644
index 0000000..2f8bb1f
--- /dev/null
+++ b/.gitmodules
@@ -0,0 +1,3 @@
+[submodule "asm/elf2hex"]
+	path = asm/elf2hex
+	url = https://github.com/sifive/elf2hex.git
diff --git a/asm/Makefile b/asm/Makefile
new file mode 100644
index 0000000..1681182
--- /dev/null
+++ b/asm/Makefile
@@ -0,0 +1,20 @@
+all: test.hex
+
+CC = riscv64-linux-gnu-gcc
+# CFLAGS = -march=rv32i -mabi=ilp32
+CFLAGS = -march=rv64i -mabi=lp64
+
+AS = riscv64-linux-gnu-as
+ASFLAGS = $(CFLAGS)
+
+LD = riscv64-linux-gnu-ld
+LDFLAGS = -T test.ld
+
+%.o: %.S
+	$(AS) $(ASFLAGS) $^ -o $@
+
+%.elf: %.o
+	$(LD) $(LDFLAGS) $^ -o $@
+
+test.hex: test.elf
+	riscv64-linux-gnu-objdump -s $^ | sed -n '/0000/,$$p' | cut -f3-6 -d ' ' | sed -e 's/ /\n/g' > $@
diff --git a/asm/test.S b/asm/test.S
new file mode 100644
index 0000000..d4f69c7
--- /dev/null
+++ b/asm/test.S
@@ -0,0 +1,36 @@
+.global _start
+.text
+_start:
+add     x1, x0, x0
+add     x2, x0, x0
+add     x3, x0, x0
+add     x4, x0, x0
+add     x5, x0, x0
+add     x6, x0, x0
+add     x7, x0, x0
+add     x8, x0, x0
+add     x9, x0, x0
+add     x10, x0, x0
+add     x11, x0, x0
+add     x12, x0, x0
+add     x13, x0, x0
+add     x14, x0, x0
+add     x15, x0, x0
+add     x16, x0, x0
+add     x17, x0, x0
+add     x18, x0, x0
+add     x19, x0, x0
+add     x20, x0, x0
+add     x21, x0, x0
+add     x22, x0, x0
+add     x23, x0, x0
+add     x24, x0, x0
+add     x25, x0, x0
+add     x26, x0, x0
+add     x27, x0, x0
+add     x28, x0, x0
+add     x29, x0, x0
+add     x30, x0, x0
+add     x31, x0, x0
+
+.data
diff --git a/asm/test.csv b/asm/test.csv
new file mode 100644
index 0000000..939004c
--- /dev/null
+++ b/asm/test.csv
@@ -0,0 +1,12 @@
+
+test.elf:     file format elf64-littleriscv
+
+Contents of section .text:
+ 0000 b3000000 33010000 b3010000 33020000  ....3.......3...
+ 0010 b3020000 33030000 b3030000 33040000  ....3.......3...
+ 0020 b3040000 33050000 b3050000 33060000  ....3.......3...
+ 0030 b3060000 33070000 b3070000 33080000  ....3.......3...
+ 0040 b3080000 33090000 b3090000 330a0000  ....3.......3...
+ 0050 b30a0000 330b0000 b30b0000 330c0000  ....3.......3...
+ 0060 b30c0000 330d0000 b30d0000 330e0000  ....3.......3...
+ 0070 b30e0000 330f0000 b30f0000           ....3.......    
diff --git a/asm/test.elf b/asm/test.elf
new file mode 100755
index 0000000..9a24b7f
Binary files /dev/null and b/asm/test.elf differ
diff --git a/asm/test.hex b/asm/test.hex
new file mode 100644
index 0000000..4301cf9
--- /dev/null
+++ b/asm/test.hex
@@ -0,0 +1,32 @@
+b3000000
+33010000
+b3010000
+33020000
+b3020000
+33030000
+b3030000
+33040000
+b3040000
+33050000
+b3050000
+33060000
+b3060000
+33070000
+b3070000
+33080000
+b3080000
+33090000
+b3090000
+330a0000
+b30a0000
+330b0000
+b30b0000
+330c0000
+b30c0000
+330d0000
+b30d0000
+330e0000
+b30e0000
+330f0000
+b30f0000
+
diff --git a/asm/test.ld b/asm/test.ld
new file mode 100644
index 0000000..4ede032
--- /dev/null
+++ b/asm/test.ld
@@ -0,0 +1,17 @@
+ENTRY(_start)
+
+MEMORY
+{
+    ROM (rx) : ORIGIN = 0x00000000, LENGTH = 1024
+}
+
+SECTIONS
+{
+    .text :
+    {
+        _text = .;
+        *(.text*)
+        *(.rodata*)
+        _etext = .;
+    } > ROM
+}
\ No newline at end of file
diff --git a/hdl/core.v b/hdl/core.v
index 1a66ea4..ef90044 100644
--- a/hdl/core.v
+++ b/hdl/core.v
@@ -30,6 +30,8 @@ reg [31:0] r_ex_s1, r_ex_s2, r_mem_s1, r_mem_s2;
 reg [31:0] r_mem_alu_out, r_wb_alu_out;
 reg r_mem_alu_zero;
 reg r_ex_jump;
+reg r_ex_store, r_mem_store;
+reg r_ex_load, r_mem_load;
 
 // IF
 reg s_if_halt;
@@ -64,6 +66,7 @@ reg [2:0] s_id_aluop;
 reg s_id_alu_seed;
 reg s_id_invalid;
 reg s_id_jump, s_id_branch;
+reg s_id_store, s_id_load;
 
 // RV32I / RV64I / RV32M
 localparam  OP_LUI      = 7'b0110111,
@@ -97,6 +100,8 @@ localparam  ALUOP_ADD   = 0,
 always @(*) begin
     s_id_halt = 0;
     s_id_invalid = 0;
+    s_id_store = 0;
+    s_id_load = 0;
 
     s_id_opcode = r_id_inst[6:0];
     s_id_rd     = r_id_inst[11:7];
@@ -333,19 +338,19 @@ always @(posedge clk) begin
         //     // stay halted forever
         // end else begin
             // IF
-            if (!s_mem_bp) begin
+            // if (!s_mem_bp) begin
                 r_if_pc     <= s_if_next_pc;
-            end
+            // end
 
             // ID
-            if (!s_mem_bp) begin
+            // if (!s_mem_bp) begin
                 r_id_stall  <= s_if_stall;
                 r_id_pc     <= r_if_pc;
                 r_id_inst   <= s_if_inst;
-            end
+            // end
 
             // EX
-            if (!s_mem_bp) begin
+            // if (!s_mem_bp) begin
                 // TODO: also stall EX if taking branch
                 r_ex_stall  <= r_id_stall;
                 r_ex_pc     <= r_id_pc;
@@ -356,11 +361,13 @@ always @(posedge clk) begin
                 r_ex_aluop  <= s_id_aluop;
                 r_ex_alu_seed <= s_id_alu_seed;
                 r_ex_jump   <= s_id_jump;
-            end
+                r_ex_store  <= s_id_store;
+                r_ex_load   <= s_id_load;
+            // end
             
 
             // MEM
-            if (!s_mem_bp) begin
+            // if (!s_mem_bp) begin
                 r_mem_stall <= r_ex_stall;
                 r_mem_pc    <= r_ex_pc;
                 r_mem_inst  <= r_ex_inst;
@@ -369,15 +376,17 @@ always @(posedge clk) begin
                 r_mem_s2    <= r_ex_s2;
                 r_mem_alu_out <= s_ex_alu_out;
                 r_mem_alu_zero <= s_ex_alu_zero;
-            end
+                r_mem_store <= r_ex_store;
+                r_mem_load  <= r_ex_load;
+            // end
 
             // WB
-            if (!s_mem_bp) begin
+            // if (!s_mem_bp) begin
                 r_wb_stall  <= r_mem_stall;
                 r_wb_pc     <= r_mem_pc;
                 r_wb_rd     <= r_mem_rd;
                 r_wb_alu_out <= r_mem_alu_out;
-            end
+            // end
 
             // Register File
             if (r_wb_rd != 0 && s_wb_write) begin