[v3,09/11] riscv: thead: Add support for the XTheadMemPair ISA extension

diff --git a/gcc/config/riscv/peephole.md b/gcc/config/riscv/peephole.md
index 0ef0c04410b..ea696f3116c 100644
--- a/gcc/config/riscv/peephole.md
+++ b/gcc/config/riscv/peephole.md
@@ -38,3 +38,59 @@  (define_peephole2
 {
   operands[5] = GEN_INT (INTVAL (operands[2]) - INTVAL (operands[5]));
 })
+
+;; XTheadMemPair: merge two SI or DI loads
+(define_peephole2
+  [(set (match_operand:GPR 0 "register_operand" "")
+	(match_operand:GPR 1 "memory_operand" ""))
+   (set (match_operand:GPR 2 "register_operand" "")
+	(match_operand:GPR 3 "memory_operand" ""))]
+  "TARGET_XTHEADMEMPAIR
+  && thead_mempair_operands_p (operands, true, <GPR:MODE>mode)"
+  [(parallel [(set (match_dup 0) (match_dup 1))
+	          (set (match_dup 2) (match_dup 3))])]
+{
+  thead_mempair_order_operands (operands, true, <GPR:MODE>mode);
+})
+
+;; XTheadMemPair: merge two SI or DI stores
+(define_peephole2
+  [(set (match_operand:GPR 0 "memory_operand" "")
+	(match_operand:GPR 1 "register_operand" ""))
+   (set (match_operand:GPR 2 "memory_operand" "")
+	(match_operand:GPR 3 "register_operand" ""))]
+  "TARGET_XTHEADMEMPAIR
+  && thead_mempair_operands_p (operands, false, <GPR:MODE>mode)"
+  [(parallel [(set (match_dup 0) (match_dup 1))
+              (set (match_dup 2) (match_dup 3))])]
+{
+  thead_mempair_order_operands (operands, false, <GPR:MODE>mode);
+})
+
+;; XTheadMemPair: merge two SI loads with sign-extension
+(define_peephole2
+  [(set (match_operand:DI 0 "register_operand" "")
+	(sign_extend:DI (match_operand:SI 1 "memory_operand" "")))
+   (set (match_operand:DI 2 "register_operand" "")
+	(sign_extend:DI (match_operand:SI 3 "memory_operand" "")))]
+  "TARGET_XTHEADMEMPAIR && TARGET_64BIT
+  && thead_mempair_operands_p (operands, true, SImode)"
+  [(parallel [(set (match_dup 0) (sign_extend:DI (match_dup 1)))
+              (set (match_dup 2) (sign_extend:DI (match_dup 3)))])]
+{
+  thead_mempair_order_operands (operands, true, SImode);
+})
+
+;; XTheadMemPair: merge two SI loads with zero-extension
+(define_peephole2
+  [(set (match_operand:DI 0 "register_operand" "")
+	(zero_extend:DI (match_operand:SI 1 "memory_operand" "")))
+   (set (match_operand:DI 2 "register_operand" "")
+	(zero_extend:DI (match_operand:SI 3 "memory_operand" "")))]
+  "TARGET_XTHEADMEMPAIR && TARGET_64BIT
+  && thead_mempair_operands_p (operands, true, SImode)"
+  [(parallel [(set (match_dup 0) (zero_extend:DI (match_dup 1)))
+              (set (match_dup 2) (zero_extend:DI (match_dup 3)))])]
+{
+  thead_mempair_order_operands (operands, true, SImode);
+})
diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 5cf4fafd662..1b7ba02726d 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -54,7 +54,13 @@  extern bool riscv_split_64bit_move_p (rtx, rtx);
 extern void riscv_split_doubleword_move (rtx, rtx);
 extern const char *riscv_output_move (rtx, rtx);
 extern const char *riscv_output_return ();
+
+extern bool thead_mempair_operands_p (rtx[4], bool, machine_mode);
+extern void thead_mempair_order_operands (rtx[4], bool, machine_mode);
+
 #ifdef RTX_CODE
+extern const char *thead_mempair_output_move (rtx[4], bool, machine_mode, RTX_CODE);
+
 extern void riscv_expand_int_scc (rtx, enum rtx_code, rtx, rtx);
 extern void riscv_expand_float_scc (rtx, enum rtx_code, rtx, rtx);
 extern void riscv_expand_conditional_branch (rtx, enum rtx_code, rtx, rtx);
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 48f2cb399ae..33854393bd2 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -2943,6 +2943,332 @@  riscv_output_move (rtx dest, rtx src)
   gcc_unreachable ();
 }
 
+/* If MEM is in the form of "base+offset", extract the two parts
+   of address and set to BASE and OFFSET, otherwise return false
+   after clearing BASE and OFFSET.  */
+
+static bool
+extract_base_offset_in_addr (rtx mem, rtx *base, rtx *offset)
+{
+  rtx addr;
+
+  gcc_assert (MEM_P (mem));
+
+  addr = XEXP (mem, 0);
+
+  if (REG_P (addr))
+    {
+      *base = addr;
+      *offset = const0_rtx;
+      return true;
+    }
+
+  if (GET_CODE (addr) == PLUS
+      && REG_P (XEXP (addr, 0)) && CONST_INT_P (XEXP (addr, 1)))
+    {
+      *base = XEXP (addr, 0);
+      *offset = XEXP (addr, 1);
+      return true;
+    }
+
+  *base = NULL_RTX;
+  *offset = NULL_RTX;
+
+  return false;
+}
+
+/* If X is a PLUS of a CONST_INT, return the two terms in *BASE_PTR
+   and *OFFSET_PTR.  Return X in *BASE_PTR and 0 in *OFFSET_PTR otherwise.  */
+
+static void
+riscv_split_plus (rtx x, rtx *base_ptr, HOST_WIDE_INT *offset_ptr)
+{
+  if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1)))
+    {
+      *base_ptr = XEXP (x, 0);
+      *offset_ptr = INTVAL (XEXP (x, 1));
+    }
+  else
+    {
+      *base_ptr = x;
+      *offset_ptr = 0;
+    }
+}
+
+/* Output a mempair instruction with the provided OPERANDS.
+   LOAD_P is true if a we have a pair of loads (stores otherwise).
+   MODE is the access mode (DI or SI).
+   CODE is the extension code (UNKNOWN, SIGN_EXTEND or ZERO_EXTEND).
+   This instruction does not handle invalid inputs gracefully,
+   but is full of assertions to ensure that only valid instructions
+   are emitted.  */
+
+const char *
+thead_mempair_output_move (rtx operands[4], bool load_p,
+			   machine_mode mode, RTX_CODE code)
+{
+  rtx reg1, reg2, mem1, mem2, base1, base2;
+  HOST_WIDE_INT offset1, offset2;
+  rtx output_operands[5];
+  const char* format;
+
+  gcc_assert (mode == SImode || mode == DImode);
+
+  /* Paired 64-bit access instructions have a fixed shift amount of 4.
+     Paired 32-bit access instructions have a fixed shift amount of 3.  */
+  unsigned shamt = (mode == DImode) ? 4 : 3;
+
+  if (load_p)
+    {
+	reg1 = copy_rtx (operands[0]);
+	reg2 = copy_rtx (operands[2]);
+	mem1 = copy_rtx (operands[1]);
+	mem2 = copy_rtx (operands[3]);
+
+	if (mode == SImode)
+	  if (code == ZERO_EXTEND)
+	    format = "th.lwud\t%0, %1, (%2), %3, %4";
+	  else //SIGN_EXTEND or UNKNOWN
+	    format = "th.lwd\t%0, %1, (%2), %3, %4";
+	else
+	  format = "th.ldd\t%0, %1, (%2), %3, %4";
+    }
+  else
+    {
+	reg1 = copy_rtx (operands[1]);
+	reg2 = copy_rtx (operands[3]);
+	mem1 = copy_rtx (operands[0]);
+	mem2 = copy_rtx (operands[2]);
+
+	if (mode == SImode)
+	  format = "th.swd\t%z0, %z1, (%2), %3, %4";
+	else
+	  format = "th.sdd\t%z0, %z1, (%2), %3, %4";
+    }
+
+  riscv_split_plus (XEXP (mem1, 0), &base1, &offset1);
+  riscv_split_plus (XEXP (mem2, 0), &base2, &offset2);
+  gcc_assert (rtx_equal_p (base1, base2));
+  auto size1 = MEM_SIZE (mem1);
+  auto size2 = MEM_SIZE (mem2);
+  gcc_assert (known_eq (size1, size2));
+  gcc_assert (known_eq (offset1 + size1, offset2));
+
+  HOST_WIDE_INT imm2 = offset1 >> shamt;
+
+  /* Make sure all mempair instruction constraints are met.  */
+  gcc_assert (imm2 >= 0 && imm2 < 4);
+  gcc_assert ((imm2 << shamt) == offset1);
+  gcc_assert (REG_P (reg1));
+  gcc_assert (REG_P (reg2));
+  gcc_assert (REG_P (base1));
+  if (load_p)
+    {
+      gcc_assert (REGNO (reg1) != REGNO (reg2));
+      gcc_assert (REGNO (reg1) != REGNO (base1));
+      gcc_assert (REGNO (reg2) != REGNO (base1));
+    }
+
+  /* Output the mempair instruction.  */
+  output_operands[0] = copy_rtx (reg1);
+  output_operands[1] = copy_rtx (reg2);
+  output_operands[2] = copy_rtx (base1);
+  output_operands[3] = gen_rtx_CONST_INT (mode, imm2);
+  output_operands[4] = gen_rtx_CONST_INT (mode, shamt);
+  output_asm_insn (format, output_operands);
+
+  return "";
+}
+
+/* Analyse if a pair of loads/stores MEM1 and MEM2 with given MODE
+   are consecutive so they can be merged into a mempair instruction.
+   RESERVED will be set to true, if a reversal of the accesses is
+   required (false otherwise). Returns true if the accesses can be
+   merged (even if reversing is necessary) and false if not.  */
+
+static bool
+thead_mempair_check_consecutive_mems (machine_mode mode, rtx *mem1, rtx *mem2,
+				      bool *reversed)
+{
+  rtx base1, base2, offset1, offset2;
+  extract_base_offset_in_addr (*mem1, &base1, &offset1);
+  extract_base_offset_in_addr (*mem2, &base2, &offset2);
+
+  /* Make sure both mems are in base+offset form.  */
+  if (!base1 || !base2)
+    return false;
+
+  /* If both mems use the same base register, just check the offsets.  */
+  if (rtx_equal_p (base1, base2))
+    {
+      auto size = GET_MODE_SIZE (mode);
+
+      if (known_eq (UINTVAL (offset1) + size, UINTVAL (offset2)))
+	{
+	  *reversed = false;
+	  return true;
+	}
+
+      if (known_eq (UINTVAL (offset2) + size, UINTVAL (offset1)))
+	{
+	  *reversed = true;
+	  return true;
+	}
+
+      return false;
+    }
+
+  return false;
+}
+
+/* Check if the given MEM can be used to define the address of a mempair
+   instruction.  */
+
+static bool
+thead_mempair_operand_p (rtx mem, machine_mode mode)
+{
+  if (!MEM_SIZE_KNOWN_P (mem))
+    return false;
+
+  /* Only DI or SI mempair instructions exist.  */
+  auto mem_sz = MEM_SIZE (mem);
+  auto mode_sz = GET_MODE_SIZE (mode);
+  if (!known_eq (mem_sz, mode_sz))
+    return false;
+
+  /* Paired 64-bit access instructions have a fixed shift amount of 4.
+     Paired 32-bit access instructions have a fixed shift amount of 3.  */
+  machine_mode mem_mode = GET_MODE (mem);
+  unsigned shamt = (mem_mode == DImode) ? 4 : 3;
+
+  rtx base;
+  HOST_WIDE_INT offset;
+  riscv_split_plus (XEXP (mem, 0), &base, &offset);
+  HOST_WIDE_INT imm2 = offset >> shamt;
+
+  if (imm2 < 0 || imm2 >= 4)
+    return false;
+
+  if ((imm2 << shamt) != offset)
+    return false;
+
+  return true;
+}
+
+static bool
+thead_mempair_load_overlap_p (rtx reg1, rtx reg2, rtx mem)
+{
+  if (REGNO (reg1) == REGNO (reg2))
+    return true;
+
+  if (reg_overlap_mentioned_p (reg1, mem))
+    return true;
+
+  rtx base;
+  HOST_WIDE_INT offset;
+  riscv_split_plus (XEXP (mem, 0), &base, &offset);
+
+  if (!REG_P (base))
+    return true;
+
+  if (REG_P (base))
+    {
+      if (REGNO (base) == REGNO (reg1)
+	  || REGNO (base) == REGNO (reg2))
+	return true;
+    }
+
+  return false;
+}
+
+/* Given OPERANDS of consecutive load/store, check if we can merge
+   them into load-pair or store-pair instructions.
+   LOAD is true if they are load instructions.
+   MODE is the mode of memory operation.  */
+
+bool
+thead_mempair_operands_p (rtx operands[4], bool load_p,
+			  machine_mode mode)
+{
+  rtx mem_1, mem_2, reg_1, reg_2;
+
+  if (load_p)
+    {
+      reg_1 = operands[0];
+      mem_1 = operands[1];
+      reg_2 = operands[2];
+      mem_2 = operands[3];
+      if (!REG_P (reg_1) || !REG_P (reg_2))
+	return false;
+      if (thead_mempair_load_overlap_p (reg_1, reg_2, mem_1))
+	return false;
+      if (thead_mempair_load_overlap_p (reg_1, reg_2, mem_2))
+	return false;
+    }
+  else
+    {
+      mem_1 = operands[0];
+      reg_1 = operands[1];
+      mem_2 = operands[2];
+      reg_2 = operands[3];
+    }
+
+  /* Check if the registers are GP registers.  */
+  if (!REG_P (reg_1) || !GP_REG_P (REGNO (reg_1))
+      || !REG_P (reg_2) || !GP_REG_P (REGNO (reg_2)))
+    return false;
+
+  /* The mems cannot be volatile.  */
+  if (!MEM_P (mem_1) || !MEM_P (mem_2))
+    return false;
+  if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2))
+    return false;
+
+  /* If we have slow unaligned access, we only accept aligned memory.  */
+  if (riscv_slow_unaligned_access_p
+      && known_lt (MEM_ALIGN (mem_1), GET_MODE_SIZE (mode) * BITS_PER_UNIT))
+    return false;
+
+  /* Check if the addresses are in the form of [base+offset].  */
+  bool reversed = false;
+  if (!thead_mempair_check_consecutive_mems (mode, &mem_1, &mem_2, &reversed))
+    return false;
+
+  /* The first memory accesses must be a mempair operand.  */
+  if ((!reversed && !thead_mempair_operand_p (mem_1, mode))
+      || (reversed && !thead_mempair_operand_p (mem_2, mode)))
+    return false;
+
+  /* The operands must be of the same size.  */
+  gcc_assert (known_eq (GET_MODE_SIZE (GET_MODE (mem_1)),
+			GET_MODE_SIZE (GET_MODE (mem_2))));
+
+  return true;
+}
+
+/* Given OPERANDS of consecutive load/store that can be merged,
+   swap them if they are not in ascending order.
+   Return true if swap was performed.  */
+void
+thead_mempair_order_operands (rtx operands[4], bool load_p, machine_mode mode)
+{
+  int mem_op = load_p ? 1 : 0;
+  bool reversed = false;
+  if (!thead_mempair_check_consecutive_mems (mode,
+					     operands + mem_op,
+					     operands + mem_op + 2,
+					     &reversed))
+    gcc_unreachable ();
+
+  if (reversed)
+    {
+      /* Irrespective of whether this is a load or a store,
+	 we do the same swap.  */
+      std::swap (operands[0], operands[2]);
+      std::swap (operands[1], operands[3]);
+    }
+}
+
 const char *
 riscv_output_return ()
 {
@@ -4963,6 +5289,61 @@  riscv_set_return_address (rtx address, rtx scratch)
   riscv_emit_move (gen_frame_mem (GET_MODE (address), slot_address), address);
 }
 
+/* Save register REG to MEM.  Make the instruction frame-related.  */
+
+static void
+riscv_save_reg (rtx reg, rtx mem)
+{
+  riscv_emit_move (mem, reg);
+  riscv_set_frame_expr (riscv_frame_set (mem, reg));
+}
+
+/* Save register 2x REG to MEM.  Make the instruction frame-related.  */
+
+static void
+th_mempair_save_reg (rtx operands[4])
+{
+  rtx set1 = gen_rtx_SET (operands[0], operands[1]);
+  rtx set2 = gen_rtx_SET (operands[2], operands[3]);
+  rtx insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set1, set2)));
+  RTX_FRAME_RELATED_P (insn) = 1;
+  add_reg_note (insn, REG_CFA_OFFSET, copy_rtx (set1));
+  add_reg_note (insn, REG_CFA_OFFSET, copy_rtx (set2));
+}
+
+/* Restore register REG from MEM.  */
+
+static void
+riscv_restore_reg (rtx reg, rtx mem)
+{
+  rtx insn = riscv_emit_move (reg, mem);
+  rtx dwarf = NULL_RTX;
+  dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf);
+
+  if (epilogue_cfa_sp_offset && REGNO (reg) == HARD_FRAME_POINTER_REGNUM)
+    {
+      rtx cfa_adjust_rtx = gen_rtx_PLUS (Pmode, stack_pointer_rtx,
+					 GEN_INT (epilogue_cfa_sp_offset));
+      dwarf = alloc_reg_note (REG_CFA_DEF_CFA, cfa_adjust_rtx, dwarf);
+    }
+
+  REG_NOTES (insn) = dwarf;
+  RTX_FRAME_RELATED_P (insn) = 1;
+}
+
+/* Restore register 2x REG from MEM.  Make the instruction frame-related.  */
+
+static void
+th_mempair_restore_reg (rtx operands[4])
+{
+  rtx set1 = gen_rtx_SET (operands[0], operands[1]);
+  rtx set2 = gen_rtx_SET (operands[2], operands[3]);
+  rtx insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set1, set2)));
+  RTX_FRAME_RELATED_P (insn) = 1;
+  add_reg_note (insn, REG_CFA_RESTORE, operands[0]);
+  add_reg_note (insn, REG_CFA_RESTORE, operands[2]);
+}
+
 /* A function to save or store a register.  The first argument is the
    register and the second is the stack slot.  */
 typedef void (*riscv_save_restore_fn) (rtx, rtx);
@@ -5057,6 +5438,46 @@  riscv_for_each_saved_reg (poly_int64 sp_offset, riscv_save_restore_fn fn,
 	  && riscv_is_eh_return_data_register (regno))
 	continue;
 
+      if (TARGET_XTHEADMEMPAIR)
+	{
+	  /* Get the next reg/offset pair.  */
+	  HOST_WIDE_INT offset2 = offset;
+	  unsigned int regno2 = riscv_next_saved_reg (regno, limit, &offset2);
+
+	  /* Prepare the operands for a mempair instruction.  */
+	  bool load_p = (fn == riscv_restore_reg);
+	  int reg_op = load_p ? 0 : 1;
+	  int mem_op = load_p ? 1 : 0;
+	  rtx operands[4];
+	  operands[reg_op] = gen_rtx_REG (word_mode, regno);
+	  rtx mem1 = plus_constant (word_mode, stack_pointer_rtx, offset);
+	  mem1 = gen_frame_mem (word_mode, mem1);
+	  operands[mem_op] = mem1;
+	  operands[2 + reg_op] = gen_rtx_REG (word_mode, regno2);
+	  rtx mem2 = plus_constant (word_mode, stack_pointer_rtx, offset2);
+	  mem2 = gen_frame_mem (word_mode, mem2);
+	  operands[2 + mem_op] = mem2;
+
+	  /* Validate everything before emitting a mempair instruction.  */
+	  if (regno2 != INVALID_REGNUM
+	      && !cfun->machine->reg_is_wrapped_separately[regno2]
+	      && !(epilogue && !maybe_eh_return
+		   && riscv_is_eh_return_data_register (regno2))
+	      && thead_mempair_operands_p (operands, load_p, word_mode))
+	    {
+	      thead_mempair_order_operands (operands, load_p, word_mode);
+
+	      if (load_p)
+		th_mempair_restore_reg (operands);
+	      else
+		th_mempair_save_reg (operands);
+
+	      offset = offset2;
+	      regno = regno2;
+	      continue;
+	    }
+	}
+
       riscv_save_restore_reg (word_mode, regno, offset, fn);
     }
 
@@ -5075,35 +5496,6 @@  riscv_for_each_saved_reg (poly_int64 sp_offset, riscv_save_restore_fn fn,
       }
 }
 
-/* Save register REG to MEM.  Make the instruction frame-related.  */
-
-static void
-riscv_save_reg (rtx reg, rtx mem)
-{
-  riscv_emit_move (mem, reg);
-  riscv_set_frame_expr (riscv_frame_set (mem, reg));
-}
-
-/* Restore register REG from MEM.  */
-
-static void
-riscv_restore_reg (rtx reg, rtx mem)
-{
-  rtx insn = riscv_emit_move (reg, mem);
-  rtx dwarf = NULL_RTX;
-  dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf);
-
-  if (epilogue_cfa_sp_offset && REGNO (reg) == HARD_FRAME_POINTER_REGNUM)
-    {
-      rtx cfa_adjust_rtx = gen_rtx_PLUS (Pmode, stack_pointer_rtx,
-					 GEN_INT (epilogue_cfa_sp_offset));
-      dwarf = alloc_reg_note (REG_CFA_DEF_CFA, cfa_adjust_rtx, dwarf);
-    }
-
-  REG_NOTES (insn) = dwarf;
-  RTX_FRAME_RELATED_P (insn) = 1;
-}
-
 /* For stack frames that can't be allocated with a single ADDI instruction,
    compute the best value to initially allocate.  It must at a minimum
    allocate enough space to spill the callee-saved registers.  If TARGET_RVC,
diff --git a/gcc/config/riscv/thead.md b/gcc/config/riscv/thead.md
index 6ec3613071f..cbfe51cbd85 100644
--- a/gcc/config/riscv/thead.md
+++ b/gcc/config/riscv/thead.md
@@ -297,3 +297,55 @@  (define_insn "*th_sextw_msubhisi4"
   [(set_attr "type" "imul")
    (set_attr "mode" "SI")]
 )
+
+;; XTheadMemPair
+
+;; MEMPAIR load 64/32 bit
+(define_insn "*th_mempair_load_<GPR:mode>2"
+  [(set (match_operand:GPR 0 "register_operand" "=r")
+	(match_operand:GPR 1 "memory_operand" "m"))
+   (set (match_operand:GPR 2 "register_operand" "=r")
+	(match_operand:GPR 3 "memory_operand" "m"))]
+  "TARGET_XTHEADMEMPAIR && reload_completed
+   && thead_mempair_operands_p (operands, true, <GPR:MODE>mode)"
+  { return thead_mempair_output_move (operands, true, <GPR:MODE>mode, UNKNOWN); }
+  [(set_attr "move_type" "load")
+   (set_attr "mode" "<GPR:MODE>")])
+
+;; MEMPAIR store 64/32 bit
+(define_insn "*th_mempair_store_<GPR:mode>2"
+  [(set (match_operand:GPR 0 "memory_operand" "=m")
+	(match_operand:GPR 1 "register_operand" "r"))
+   (set (match_operand:GPR 2 "memory_operand" "=m")
+	(match_operand:GPR 3 "register_operand" "r"))]
+  "TARGET_XTHEADMEMPAIR && reload_completed
+   && thead_mempair_operands_p (operands, false, <GPR:MODE>mode)"
+  { return thead_mempair_output_move (operands, false, <GPR:MODE>mode, UNKNOWN); }
+  [(set_attr "move_type" "store")
+   (set_attr "mode" "<GPR:MODE>")])
+
+;; MEMPAIR load DI extended signed SI
+(define_insn "*th_mempair_load_extendsidi2"
+  [(set (match_operand 0 "register_operand" "=r")
+	(sign_extend:DI (match_operand 1 "memory_operand" "m")))
+   (set (match_operand 2 "register_operand" "=r")
+	(sign_extend:DI (match_operand 3 "memory_operand" "m")))]
+  "TARGET_XTHEADMEMPAIR && TARGET_64BIT && reload_completed
+   && thead_mempair_operands_p (operands, true, SImode)"
+  { return thead_mempair_output_move (operands, true, SImode, SIGN_EXTEND); }
+  [(set_attr "move_type" "load")
+   (set_attr "mode" "DI")
+   (set_attr "length" "8")])
+
+;; MEMPAIR load DI extended unsigned SI
+(define_insn "*th_mempair_load_zero_extendsidi2"
+  [(set (match_operand 0 "register_operand" "=r")
+	(zero_extend:DI (match_operand 1 "memory_operand" "m")))
+   (set (match_operand 2 "register_operand" "=r")
+	(zero_extend:DI (match_operand 3 "memory_operand" "m")))]
+  "TARGET_XTHEADMEMPAIR && TARGET_64BIT && reload_completed
+   && thead_mempair_operands_p (operands, true, SImode)"
+  { return thead_mempair_output_move (operands, true, SImode, ZERO_EXTEND); }
+  [(set_attr "move_type" "load")
+   (set_attr "mode" "DI")
+   (set_attr "length" "8")])
diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmempair-1.c b/gcc/testsuite/gcc.target/riscv/xtheadmempair-1.c
new file mode 100644
index 00000000000..c324555fd35
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/xtheadmempair-1.c
@@ -0,0 +1,98 @@ 
+/* { dg-do compile } */
+/* { dg-skip-if "" { *-*-* } { "-O0" "-O1" "-g" "-Oz" "-Os" "-flto" } } */
+/* { dg-options "-march=rv64gc_xtheadmempair -mtune=thead-c906" { target { rv64 } } } */
+/* { dg-options "-march=rv32gc_xtheadmempair -mtune=thead-c906" { target { rv32 } } } */
+
+#include <inttypes.h>
+
+#if __riscv_xlen == 32
+typedef uint32_t xlen_t;
+#else
+typedef uint64_t xlen_t;
+#endif
+
+void foof (xlen_t*, xlen_t, xlen_t);
+void foor (xlen_t*, xlen_t, xlen_t);
+void foowu (uint32_t*, uint64_t, uint64_t);
+void foows (int32_t*, int64_t, int64_t);
+
+#define LxD_TEST(f, T, i1, i2)		\
+void					\
+f ## i1 ## i2(T *arr)			\
+{					\
+  foo ## f(arr, arr[i1], arr[i2]);	\
+}
+
+// works
+LxD_TEST(f, xlen_t, 0, 1)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(f, xlen_t, 1, 2)
+// works
+LxD_TEST(f, xlen_t, 2, 3)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(f, xlen_t, 3, 4)
+// works
+LxD_TEST(f, xlen_t, 4, 5)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(f, xlen_t, 5, 6)
+// works
+LxD_TEST(f, xlen_t, 6, 7)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(f, xlen_t, 7, 8)
+// does not work (out of range)
+LxD_TEST(f, xlen_t, 8, 9)
+
+// works with reordering
+LxD_TEST(r, xlen_t, 1, 0)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(r, xlen_t, 2, 1)
+// works with reordering
+LxD_TEST(r, xlen_t, 3, 2)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(r, xlen_t, 4, 3)
+// works with reordering
+LxD_TEST(r, xlen_t, 5, 4)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(r, xlen_t, 6, 5)
+// works with reordering
+LxD_TEST(r, xlen_t, 7, 6)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(r, xlen_t, 8, 7)
+// does not work (out of range)
+LxD_TEST(r, xlen_t, 9, 8)
+
+#if __riscv_xlen != 32
+// works
+LxD_TEST(wu, uint32_t, 0, 1)
+LxD_TEST(ws, int32_t, 0, 1)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(wu, uint32_t, 1, 2)
+LxD_TEST(ws, int32_t, 1, 2)
+// works
+LxD_TEST(wu, uint32_t, 2, 3)
+LxD_TEST(ws, int32_t, 2, 3)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(wu, uint32_t, 3, 4)
+LxD_TEST(ws, int32_t, 3, 4)
+// works
+LxD_TEST(wu, uint32_t, 4, 5)
+LxD_TEST(ws, int32_t, 4, 5)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(wu, uint32_t, 5, 6)
+LxD_TEST(ws, int32_t, 5, 6)
+// works
+LxD_TEST(wu, uint32_t, 6, 7)
+LxD_TEST(ws, int32_t, 6, 7)
+// does not work (can't merge with unaligned offset)
+LxD_TEST(wu, uint32_t, 7, 8)
+LxD_TEST(ws, int32_t, 7, 8)
+// does not work (out of range)
+LxD_TEST(wu, uint32_t, 8, 9)
+LxD_TEST(ws, int32_t, 8, 9)
+#endif
+
+/* { dg-final { scan-assembler-times "th.ldd\t" 8 { target { rv64 } } } } */
+/* { dg-final { scan-assembler-times "th.luwd\t" 4 { target { rv64 } } } } */
+/* { dg-final { scan-assembler-times "th.lwd\t" 4 { target { rv64 } } } } */
+
+/* { dg-final { scan-assembler-times "th.lwd\t" 8 { target { rv32 } } } } */
diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmempair-2.c b/gcc/testsuite/gcc.target/riscv/xtheadmempair-2.c
new file mode 100644
index 00000000000..ed639709196
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/xtheadmempair-2.c
@@ -0,0 +1,84 @@ 
+/* { dg-do compile } */
+/* { dg-skip-if "" { *-*-* } { "-O0" "-O1" "-g" "-Oz" "-Os" "-flto" } } */
+/* { dg-options "-march=rv64gc_xtheadmempair -mtune=thead-c906" { target { rv64 } } } */
+/* { dg-options "-march=rv32gc_xtheadmempair -mtune=thead-c906" { target { rv32 } } } */
+
+#include <inttypes.h>
+
+#if __riscv_xlen == 32
+typedef uint32_t xlen_t;
+#else
+typedef uint64_t xlen_t;
+#endif
+
+#define SxD_TEST(f, T, i1, i2)		\
+void					\
+f ## i1 ## i2(T *arr, T x, T y)		\
+{					\
+  arr[i1] = x;				\
+  arr[i2] = y;				\
+}
+
+// works
+SxD_TEST(f, xlen_t, 0, 1)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(f, xlen_t, 1, 2)
+// works
+SxD_TEST(f, xlen_t, 2, 3)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(f, xlen_t, 3, 4)
+// works
+SxD_TEST(f, xlen_t, 4, 5)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(f, xlen_t, 5, 6)
+// works
+SxD_TEST(f, xlen_t, 6, 7)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(f, xlen_t, 7, 8)
+// does not work (out of range)
+SxD_TEST(f, xlen_t, 8, 9)
+
+// works with reordering
+SxD_TEST(r, xlen_t, 1, 0)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(r, xlen_t, 2, 1)
+// works with reordering
+SxD_TEST(r, xlen_t, 3, 2)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(r, xlen_t, 4, 3)
+// works with reordering
+SxD_TEST(r, xlen_t, 5, 4)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(r, xlen_t, 6, 5)
+// works with reordering
+SxD_TEST(r, xlen_t, 7, 6)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(r, xlen_t, 8, 7)
+// does not work (out of range)
+SxD_TEST(r, xlen_t, 9, 8)
+
+#if __riscv_xlen != 32
+// works
+SxD_TEST(w, uint32_t, 0, 1)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(w, uint32_t, 1, 2)
+// works
+SxD_TEST(w, uint32_t, 2, 3)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(w, uint32_t, 3, 4)
+// works
+SxD_TEST(w, uint32_t, 4, 5)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(w, uint32_t, 5, 6)
+// works
+SxD_TEST(w, uint32_t, 6, 7)
+// does not work (can't merge with unaligned offset)
+SxD_TEST(w, uint32_t, 7, 8)
+// does not work (out of range)
+SxD_TEST(w, uint32_t, 8, 9)
+#endif
+
+/* { dg-final { scan-assembler-times "th.sdd\t" 8 { target { rv64 } } } } */
+/* { dg-final { scan-assembler-times "th.swd\t" 4 { target { rv64 } } } } */
+
+/* { dg-final { scan-assembler-times "th.swd\t" 8 { target { rv32 } } } } */
diff --git a/gcc/testsuite/gcc.target/riscv/xtheadmempair-3.c b/gcc/testsuite/gcc.target/riscv/xtheadmempair-3.c
new file mode 100644
index 00000000000..5dec702819a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/xtheadmempair-3.c
@@ -0,0 +1,29 @@ 
+/* { dg-do compile } */
+/* { dg-skip-if "" { *-*-* } { "-O0" "-O1" "-g" "-Oz" "-Os" "-flto" } } */
+/* { dg-options "-march=rv64gc_xtheadmempair -mtune=thead-c906" { target { rv64 } } } */
+/* { dg-options "-march=rv32gc_xtheadmempair -mtune=thead-c906" { target { rv32 } } } */
+
+#include <inttypes.h>
+
+#if __riscv_xlen == 32
+typedef uint32_t xlen_t;
+#else
+typedef uint64_t xlen_t;
+#endif
+
+void foo (xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t);
+void bar (xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t, xlen_t);
+
+void baz (xlen_t a, xlen_t b, xlen_t c, xlen_t d, xlen_t e, xlen_t f, xlen_t g, xlen_t h)
+{
+  foo (a, b, c, d, e, f, g, h);
+  /* RV64: We don't use 0(sp), therefore we can only get 3 mempairs.  */
+  /* RV32: We don't use 0(sp)-8(sp), therefore we can only get 2 mempairs.  */
+  bar (a, b, c, d, e, f, g, h);
+}
+
+/* { dg-final { scan-assembler-times "th.ldd\t" 3 { target { rv64 } } } } */
+/* { dg-final { scan-assembler-times "th.sdd\t" 3 { target { rv64 } } } } */
+
+/* { dg-final { scan-assembler-times "th.lwd\t" 2 { target { rv32 } } } } */
+/* { dg-final { scan-assembler-times "th.swd\t" 2 { target { rv32 } } } } */