[gcc(refs/users/meissner/heads/work161-dmf)] PowerPC: Add support for 1, 024 bit DMR registers.

[Michael Meissner <meissner@gcc.gnu.org>]

https://gcc.gnu.org/g:4276b6f8707c890cd45b120b9405b713aec5c477

commit 4276b6f8707c890cd45b120b9405b713aec5c477
Author: Michael Meissner <meissner@linux.ibm.com>
Date:   Tue Mar 5 02:46:24 2024 -0500

    PowerPC: Add support for 1,024 bit DMR registers.
    
    This patch is a prelimianry patch to add the full 1,024 bit dense math register
    (DMRs) for -mcpu=future.  The MMA 512-bit accumulators map onto the top of the
    DMR register.
    
    This patch only adds the new 1,024 bit register support.  It does not add
    support for any instructions that need 1,024 bit registers instead of 512 bit
    registers.
    
    I used the new mode 'TDOmode' to be the opaque mode used for 1,024 bit
    registers.  The 'wD' constraint added in previous patches is used for these
    registers.  I added support to do load and store of DMRs via the VSX registers,
    since there are no load/store dense math instructions.  I added the new keyword
    '__dmr' to create 1,024 bit types that can be loaded into DMRs.  At present, I
    don't have aliases for __dmr512 and __dmr1024 that we've discussed internally.
    
    The patches have been tested on both little and big endian systems.  Can I check
    it into the master branch?
    
    2024-03-05   Michael Meissner  <meissner@linux.ibm.com>
    
    gcc/
    
            * config/rs6000/mma.md (UNSPEC_DM_INSERT512_UPPER): New unspec.
            (UNSPEC_DM_INSERT512_LOWER): Likewise.
            (UNSPEC_DM_EXTRACT512): Likewise.
            (UNSPEC_DMR_RELOAD_FROM_MEMORY): Likewise.
            (UNSPEC_DMR_RELOAD_TO_MEMORY): Likewise.
            (movtdo): New define_expand and define_insn_and_split to implement 1,024
            bit DMR registers.
            (movtdo_insert512_upper): New insn.
            (movtdo_insert512_lower): Likewise.
            (movtdo_extract512): Likewise.
            (reload_dmr_from_memory): Likewise.
            (reload_dmr_to_memory): Likewise.
            * config/rs6000/rs6000-builtin.cc (rs6000_type_string): Add DMR
            support.
            (rs6000_init_builtins): Add support for __dmr keyword.
            * config/rs6000/rs6000-call.cc (rs6000_return_in_memory): Add support
            for TDOmode.
            (rs6000_function_arg): Likewise.
            * config/rs6000/rs6000-modes.def (TDOmode): New mode.
            * config/rs6000/rs6000.cc (rs6000_hard_regno_nregs_internal): Add
            support for TDOmode.
            (rs6000_hard_regno_mode_ok_uncached): Likewise.
            (rs6000_hard_regno_mode_ok): Likewise.
            (rs6000_modes_tieable_p): Likewise.
            (rs6000_debug_reg_global): Likewise.
            (rs6000_setup_reg_addr_masks): Likewise.
            (rs6000_init_hard_regno_mode_ok): Add support for TDOmode.  Setup reload
            hooks for DMR mode.
            (reg_offset_addressing_ok_p): Add support for TDOmode.
            (rs6000_emit_move): Likewise.
            (rs6000_secondary_reload_simple_move): Likewise.
            (rs6000_preferred_reload_class): Likewise.
            (rs6000_secondary_reload_class): Likewise.
            (rs6000_mangle_type): Add mangling for __dmr type.
            (rs6000_dmr_register_move_cost): Add support for TDOmode.
            (rs6000_split_multireg_move): Likewise.
            (rs6000_invalid_conversion): Likewise.
            * config/rs6000/rs6000.h (VECTOR_ALIGNMENT_P): Add TDOmode.
            (enum rs6000_builtin_type_index): Add DMR type nodes.
            (dmr_type_node): Likewise.
            (ptr_dmr_type_node): Likewise.
    
    gcc/testsuite/
    
            * gcc.target/powerpc/dm-1024bit.c: New test.

Diff:
---
 gcc/config/rs6000/mma.md                      | 155 ++++++++++++++++++++++++++
 gcc/config/rs6000/rs6000-builtin.cc           |  17 +++
 gcc/config/rs6000/rs6000-call.cc              |  10 +-
 gcc/config/rs6000/rs6000-modes.def            |   4 +
 gcc/config/rs6000/rs6000.cc                   | 102 ++++++++++++-----
 gcc/config/rs6000/rs6000.h                    |   6 +-
 gcc/testsuite/gcc.target/powerpc/dm-1024bit.c |  63 +++++++++++
 7 files changed, 323 insertions(+), 34 deletions(-)

diff --git a/gcc/config/rs6000/mma.md b/gcc/config/rs6000/mma.md
index 1a93c60418f..f13a7a2f877 100644
--- a/gcc/config/rs6000/mma.md
+++ b/gcc/config/rs6000/mma.md
@@ -91,6 +91,11 @@
    UNSPEC_MMA_XVI8GER4SPP
    UNSPEC_MMA_XXMFACC
    UNSPEC_MMA_XXMTACC
+   UNSPEC_DM_INSERT512_UPPER
+   UNSPEC_DM_INSERT512_LOWER
+   UNSPEC_DM_EXTRACT512
+   UNSPEC_DMR_RELOAD_FROM_MEMORY
+   UNSPEC_DMR_RELOAD_TO_MEMORY
   ])
 
 (define_c_enum "unspecv"
@@ -779,3 +784,153 @@
   [(set_attr "type" "mma")
    (set_attr "prefixed" "yes")
    (set_attr "isa" "dm,not_dm,not_dm")])
+
+\f
+;; TDOmode (i.e. __dmr).
+(define_expand "movtdo"
+  [(set (match_operand:TDO 0 "nonimmediate_operand")
+	(match_operand:TDO 1 "input_operand"))]
+  "TARGET_DENSE_MATH"
+{
+  rs6000_emit_move (operands[0], operands[1], TDOmode);
+  DONE;
+})
+
+(define_insn_and_split "*movtdo"
+  [(set (match_operand:TDO 0 "nonimmediate_operand" "=wa,m,wa,wD,wD,wa")
+	(match_operand:TDO 1 "input_operand" "m,wa,wa,wa,wD,wD"))]
+  "TARGET_DENSE_MATH
+   && (gpc_reg_operand (operands[0], TDOmode)
+       || gpc_reg_operand (operands[1], TDOmode))"
+  "@
+   #
+   #
+   #
+   #
+   dmmr %0,%1
+   #"
+  "&& reload_completed
+   && (!dmr_operand (operands[0], TDOmode) || !dmr_operand (operands[1], TDOmode))"
+  [(const_int 0)]
+{
+  rtx op0 = operands[0];
+  rtx op1 = operands[1];
+
+  if (REG_P (op0) && REG_P (op1))
+    {
+      int regno0 = REGNO (op0);
+      int regno1 = REGNO (op1);
+
+      if (DMR_REGNO_P (regno0) && VSX_REGNO_P (regno1))
+	{
+	  rtx op1_upper = gen_rtx_REG (XOmode, regno1);
+	  rtx op1_lower = gen_rtx_REG (XOmode, regno1 + 4);
+	  emit_insn (gen_movtdo_insert512_upper (op0, op1_upper));
+	  emit_insn (gen_movtdo_insert512_lower (op0, op0, op1_lower));
+	  DONE;
+	}
+
+      else if (VSX_REGNO_P (regno0) && DMR_REGNO_P (regno1))
+	{
+	  rtx op0_upper = gen_rtx_REG (XOmode, regno0);
+	  rtx op0_lower = gen_rtx_REG (XOmode, regno0 + 4);
+	  emit_insn (gen_movtdo_extract512 (op0_upper, op1, const0_rtx));
+	  emit_insn (gen_movtdo_extract512 (op0_lower, op1, const1_rtx));
+	  DONE;
+	}
+
+     else
+	gcc_assert (VSX_REGNO_P (regno0) && VSX_REGNO_P (regno1));
+    }
+
+  rs6000_split_multireg_move (operands[0], operands[1]);
+  DONE;
+}
+  [(set_attr "type" "vecload,vecstore,vecmove,vecmove,vecmove,vecmove")
+   (set_attr "length" "*,*,32,8,*,8")
+   (set_attr "max_prefixed_insns" "4,4,*,*,*,*")])
+
+;; Move from VSX registers to DMR registers via two insert 512 bit
+;; instructions.
+(define_insn "movtdo_insert512_upper"
+  [(set (match_operand:TDO 0 "dmr_operand" "=wD")
+	(unspec:TDO [(match_operand:XO 1 "vsx_register_operand" "wa")]
+		    UNSPEC_DM_INSERT512_UPPER))]
+  "TARGET_DENSE_MATH"
+  "dmxxinstdmr512 %0,%1,%Y1,0"
+  [(set_attr "type" "mma")])
+
+(define_insn "movtdo_insert512_lower"
+  [(set (match_operand:TDO 0 "dmr_operand" "=wD")
+	(unspec:TDO [(match_operand:TDO 1 "dmr_operand" "0")
+		     (match_operand:XO 2 "vsx_register_operand" "wa")]
+		    UNSPEC_DM_INSERT512_LOWER))]
+  "TARGET_DENSE_MATH"
+  "dmxxinstdmr512 %0,%2,%Y2,1"
+  [(set_attr "type" "mma")])
+
+;; Move from DMR registers to VSX registers via two extract 512 bit
+;; instructions.
+(define_insn "movtdo_extract512"
+  [(set (match_operand:XO 0 "vsx_register_operand" "=wa")
+	(unspec:XO [(match_operand:TDO 1 "dmr_operand" "wD")
+		    (match_operand 2 "const_0_to_1_operand" "n")]
+		   UNSPEC_DM_EXTRACT512))]
+  "TARGET_DENSE_MATH"
+  "dmxxextfdmr512 %0,%Y0,%1,%2"
+  [(set_attr "type" "mma")])
+
+;; Reload DMR registers from memory
+(define_insn_and_split "reload_dmr_from_memory"
+  [(set (match_operand:TDO 0 "dmr_operand" "=wD")
+	(unspec:TDO [(match_operand:TDO 1 "memory_operand" "m")]
+		    UNSPEC_DMR_RELOAD_FROM_MEMORY))
+   (clobber (match_operand:XO 2 "vsx_register_operand" "=wa"))]
+  "TARGET_DENSE_MATH"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+{
+  rtx dest = operands[0];
+  rtx src = operands[1];
+  rtx tmp = operands[2];
+  rtx mem_upper = adjust_address (src, XOmode, BYTES_BIG_ENDIAN ? 0 : 64);
+  rtx mem_lower = adjust_address (src, XOmode, BYTES_BIG_ENDIAN ? 64 : 0);
+
+  emit_move_insn (tmp, mem_upper);
+  emit_insn (gen_movtdo_insert512_upper (dest, tmp));
+
+  emit_move_insn (tmp, mem_lower);
+  emit_insn (gen_movtdo_insert512_lower (dest, dest, tmp));
+  DONE;
+}
+  [(set_attr "length" "16")
+   (set_attr "max_prefixed_insns" "2")
+   (set_attr "type" "vecload")])
+
+;; Reload dense math registers to memory
+(define_insn_and_split "reload_dmr_to_memory"
+  [(set (match_operand:TDO 0 "memory_operand" "=m")
+	(unspec:TDO [(match_operand:TDO 1 "dmr_operand" "wD")]
+		    UNSPEC_DMR_RELOAD_TO_MEMORY))
+   (clobber (match_operand:XO 2 "vsx_register_operand" "=wa"))]
+  "TARGET_DENSE_MATH"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+{
+  rtx dest = operands[0];
+  rtx src = operands[1];
+  rtx tmp = operands[2];
+  rtx mem_upper = adjust_address (dest, XOmode, BYTES_BIG_ENDIAN ? 0 : 64);
+  rtx mem_lower = adjust_address (dest, XOmode, BYTES_BIG_ENDIAN ? 64 : 0);
+
+  emit_insn (gen_movtdo_extract512 (tmp, src, const0_rtx));
+  emit_move_insn (mem_upper, tmp);
+
+  emit_insn (gen_movtdo_extract512 (tmp, src, const1_rtx));
+  emit_move_insn (mem_lower, tmp);
+  DONE;
+}
+  [(set_attr "length" "16")
+   (set_attr "max_prefixed_insns" "2")])
diff --git a/gcc/config/rs6000/rs6000-builtin.cc b/gcc/config/rs6000/rs6000-builtin.cc
index cf96ec6a869..976a42a74cd 100644
--- a/gcc/config/rs6000/rs6000-builtin.cc
+++ b/gcc/config/rs6000/rs6000-builtin.cc
@@ -495,6 +495,8 @@ const char *rs6000_type_string (tree type_node)
     return "__vector_pair";
   else if (type_node == vector_quad_type_node)
     return "__vector_quad";
+  else if (type_node == dmr_type_node)
+    return "__dmr";
 
   return "unknown";
 }
@@ -781,6 +783,21 @@ rs6000_init_builtins (void)
   t = build_qualified_type (vector_quad_type_node, TYPE_QUAL_CONST);
   ptr_vector_quad_type_node = build_pointer_type (t);
 
+  /* For TDOmode (1,024 bit dense math accumulators), don't use an alignment of
+     1,024, use 512.  TDOmode loads and stores are always broken up into 2
+     vector pair loads or stores.  In addition, we don't have support for
+     aligning the stack to 1,024 bits.  */
+  dmr_type_node = make_node (OPAQUE_TYPE);
+  SET_TYPE_MODE (dmr_type_node, TDOmode);
+  TYPE_SIZE (dmr_type_node) = bitsize_int (GET_MODE_BITSIZE (TDOmode));
+  TYPE_PRECISION (dmr_type_node) = GET_MODE_BITSIZE (TDOmode);
+  TYPE_SIZE_UNIT (dmr_type_node) = size_int (GET_MODE_SIZE (TDOmode));
+  SET_TYPE_ALIGN (dmr_type_node, 512);
+  TYPE_USER_ALIGN (dmr_type_node) = 0;
+  lang_hooks.types.register_builtin_type (dmr_type_node, "__dmr");
+  t = build_qualified_type (dmr_type_node, TYPE_QUAL_CONST);
+  ptr_dmr_type_node = build_pointer_type (t);
+
   tdecl = add_builtin_type ("__bool char", bool_char_type_node);
   TYPE_NAME (bool_char_type_node) = tdecl;
 
diff --git a/gcc/config/rs6000/rs6000-call.cc b/gcc/config/rs6000/rs6000-call.cc
index 8c590903c86..5cda8375902 100644
--- a/gcc/config/rs6000/rs6000-call.cc
+++ b/gcc/config/rs6000/rs6000-call.cc
@@ -437,14 +437,15 @@ rs6000_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
   if (cfun
       && !cfun->machine->mma_return_type_error
       && TREE_TYPE (cfun->decl) == fntype
-      && (TYPE_MODE (type) == OOmode || TYPE_MODE (type) == XOmode))
+      && OPAQUE_MODE_P (TYPE_MODE (type)))
     {
       /* Record we have now handled function CFUN, so the next time we
 	 are called, we do not re-report the same error.  */
       cfun->machine->mma_return_type_error = true;
       if (TYPE_CANONICAL (type) != NULL_TREE)
 	type = TYPE_CANONICAL (type);
-      error ("invalid use of MMA type %qs as a function return value",
+      error ("invalid use of %s type %qs as a function return value",
+	     (TYPE_MODE (type) == TDOmode) ? "dense math" : "MMA",
 	     IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))));
     }
 
@@ -1632,11 +1633,12 @@ rs6000_function_arg (cumulative_args_t cum_v, const function_arg_info &arg)
   int n_elts;
 
   /* We do not allow MMA types being used as function arguments.  */
-  if (mode == OOmode || mode == XOmode)
+  if (OPAQUE_MODE_P (mode))
     {
       if (TYPE_CANONICAL (type) != NULL_TREE)
 	type = TYPE_CANONICAL (type);
-      error ("invalid use of MMA operand of type %qs as a function parameter",
+      error ("invalid use of %s operand of type %qs as a function parameter",
+	     (mode == TDOmode) ? "dense math" : "MMA",
 	     IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (type))));
       return NULL_RTX;
     }
diff --git a/gcc/config/rs6000/rs6000-modes.def b/gcc/config/rs6000/rs6000-modes.def
index 094b246c834..43d839bf30c 100644
--- a/gcc/config/rs6000/rs6000-modes.def
+++ b/gcc/config/rs6000/rs6000-modes.def
@@ -86,3 +86,7 @@ PARTIAL_INT_MODE (TI, 128, PTI);
 /* Modes used by __vector_pair and __vector_quad.  */
 OPAQUE_MODE (OO, 32);
 OPAQUE_MODE (XO, 64);
+
+/* Mode used by __dmr.  */
+OPAQUE_MODE (TDO, 128);
+
diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
index 050aa2930d2..3af3c0b989b 100644
--- a/gcc/config/rs6000/rs6000.cc
+++ b/gcc/config/rs6000/rs6000.cc
@@ -1836,7 +1836,8 @@ rs6000_hard_regno_nregs_internal (int regno, machine_mode mode)
      128-bit floating point that can go in vector registers, which has VSX
      memory addressing.  */
   if (FP_REGNO_P (regno))
-    reg_size = (VECTOR_MEM_VSX_P (mode) || VECTOR_ALIGNMENT_P (mode)
+    reg_size = (VECTOR_MEM_VSX_P (mode)
+		|| VECTOR_ALIGNMENT_P (mode)
 		? UNITS_PER_VSX_WORD
 		: UNITS_PER_FP_WORD);
 
@@ -1877,12 +1878,13 @@ rs6000_hard_regno_mode_ok_uncached (int regno, machine_mode mode)
      have part of the register with register numberss less than 32 (i.e. the
      traditional floating point registers) and part above 32 (i.e. the
      traditional Altivec registers).  */
-  if (mode == XOmode)
+  if (mode == XOmode || mode == TDOmode)
     {
       if (TARGET_MMA)
 	{
 	  if (!TARGET_DENSE_MATH)
-	    return (FP_REGNO_P (regno)
+	    return (mode == XOmode
+		    && FP_REGNO_P (regno)
 		    && FP_REGNO_P (last_regno)
 		    && (regno % 3) == 0);
 
@@ -1899,7 +1901,7 @@ rs6000_hard_regno_mode_ok_uncached (int regno, machine_mode mode)
       return 0;
     }
 
-  /* No other types other than XOmode can go in DMRs.  */
+  /* No other types other than XOmode or TDOmode can go in DMRs.  */
   if (DMR_REGNO_P (regno))
     return 0;
 
@@ -2007,9 +2009,11 @@ rs6000_hard_regno_mode_ok (unsigned int regno, machine_mode mode)
    GPR registers, and TImode can go in any GPR as well as VSX registers (PR
    57744).
 
-   Similarly, don't allow OOmode (vector pair, restricted to even VSX
-   registers) or XOmode (vector quad, restricted to FPR registers divisible
-   by 4) to tie with other modes.
+   Similarly, don't allow OOmode (vector pair), XOmode (vector quad), or
+   TDOmode (dmr register) to pair with anything else.  Vector pairs are
+   restricted to even/odd VSX registers.  Without dense math, vector quads are
+   limited to FPR registers divisible by 4.  With dense math, vector quads are
+   limited to even VSX registers or DMR registers.
 
    Altivec/VSX vector tests were moved ahead of scalar float mode, so that IEEE
    128-bit floating point on VSX systems ties with other vectors.  */
@@ -2018,7 +2022,8 @@ static bool
 rs6000_modes_tieable_p (machine_mode mode1, machine_mode mode2)
 {
   if (mode1 == PTImode || mode1 == OOmode || mode1 == XOmode
-      || mode2 == PTImode || mode2 == OOmode || mode2 == XOmode)
+      || mode1 == TDOmode || mode2 == PTImode || mode2 == OOmode
+      || mode2 == XOmode || mode2 == TDOmode)
     return mode1 == mode2;
 
   if (ALTIVEC_OR_VSX_VECTOR_MODE (mode1))
@@ -2309,6 +2314,7 @@ rs6000_debug_reg_global (void)
     V4DFmode,
     OOmode,
     XOmode,
+    TDOmode,
     CCmode,
     CCUNSmode,
     CCEQmode,
@@ -2674,7 +2680,7 @@ rs6000_setup_reg_addr_masks (void)
 	  /* Special case DMR registers.  */
 	  if (rc == RELOAD_REG_DMR)
 	    {
-	      if (TARGET_DENSE_MATH && m2 == XOmode)
+	      if (TARGET_DENSE_MATH && (m2 == XOmode || m2 == TDOmode))
 		{
 		  addr_mask = RELOAD_REG_VALID;
 		  reg_addr[m].addr_mask[rc] = addr_mask;
@@ -2781,10 +2787,10 @@ rs6000_setup_reg_addr_masks (void)
 
 	  /* Vector pairs can do both indexed and offset loads if the
 	     instructions are enabled, otherwise they can only do offset loads
-	     since it will be broken into two vector moves.  Vector quads can
-	     only do offset loads.  */
+	     since it will be broken into two vector moves.  Vector quads and
+	     dense math types can only do offset loads.  */
 	  else if ((addr_mask != 0) && TARGET_MMA
-		   && (m2 == OOmode || m2 == XOmode))
+		   && (m2 == OOmode || m2 == XOmode || m2 == TDOmode))
 	    {
 	      addr_mask |= RELOAD_REG_OFFSET;
 	      if (rc == RELOAD_REG_FPR || rc == RELOAD_REG_VMX)
@@ -3012,6 +3018,14 @@ rs6000_init_hard_regno_mode_ok (bool global_init_p)
       rs6000_vector_align[XOmode] = 512;
     }
 
+  /* Add support for 1,024 bit DMR registers.  */
+  if (TARGET_DENSE_MATH)
+    {
+      rs6000_vector_unit[TDOmode] = VECTOR_NONE;
+      rs6000_vector_mem[TDOmode] = VECTOR_VSX;
+      rs6000_vector_align[TDOmode] = 512;
+    }
+
   /* Register class constraints for the constraints that depend on compile
      switches. When the VSX code was added, different constraints were added
      based on the type (DFmode, V2DFmode, V4SFmode).  For the vector types, all
@@ -3224,6 +3238,12 @@ rs6000_init_hard_regno_mode_ok (bool global_init_p)
 	}
     }
 
+  if (TARGET_DENSE_MATH)
+    {
+      reg_addr[TDOmode].reload_load = CODE_FOR_reload_dmr_from_memory;
+      reg_addr[TDOmode].reload_store = CODE_FOR_reload_dmr_to_memory;
+    }
+
   /* Precalculate HARD_REGNO_NREGS.  */
   for (r = 0; HARD_REGISTER_NUM_P (r); ++r)
     for (m = 0; m < NUM_MACHINE_MODES; ++m)
@@ -8699,12 +8719,15 @@ reg_offset_addressing_ok_p (machine_mode mode)
 	return mode_supports_dq_form (mode);
       break;
 
-      /* The vector pair/quad types support offset addressing if the
-	 underlying vectors support offset addressing.  */
+      /* The vector pair/quad types and the dense math types support offset
+	 addressing if the underlying vectors support offset addressing.  */
     case E_OOmode:
     case E_XOmode:
       return TARGET_MMA;
 
+    case E_TDOmode:
+      return TARGET_DENSE_MATH;
+
     case E_SDmode:
       /* If we can do direct load/stores of SDmode, restrict it to reg+reg
 	 addressing for the LFIWZX and STFIWX instructions.  */
@@ -11253,6 +11276,12 @@ rs6000_emit_move (rtx dest, rtx source, machine_mode mode)
 	       (mode == OOmode) ? "__vector_pair" : "__vector_quad");
       break;
 
+    case E_TDOmode:
+      if (CONST_INT_P (operands[1]))
+	error ("%qs is an opaque type, and you cannot set it to constants",
+	       "__dmr");
+      break;
+
     case E_SImode:
     case E_DImode:
       /* Use default pattern for address of ELF small data */
@@ -12716,7 +12745,7 @@ rs6000_secondary_reload_simple_move (enum rs6000_reg_type to_type,
 
   /* We can transfer between VSX registers and DMR registers without needing
      extra registers.  */
-  if (TARGET_DENSE_MATH && mode == XOmode
+  if (TARGET_DENSE_MATH && (mode == XOmode || mode == TDOmode)
       && ((to_type == DMR_REG_TYPE && from_type == VSX_REG_TYPE)
 	  || (to_type == VSX_REG_TYPE && from_type == DMR_REG_TYPE)))
     return true;
@@ -13517,6 +13546,9 @@ rs6000_preferred_reload_class (rtx x, enum reg_class rclass)
       if (mode == XOmode)
 	return TARGET_MMA_DENSE_MATH ? VSX_REGS : FLOAT_REGS;
 
+      if (mode == TDOmode)
+	return VSX_REGS;
+
       if (GET_MODE_CLASS (mode) == MODE_INT)
 	return GENERAL_REGS;
     }
@@ -20733,6 +20765,8 @@ rs6000_mangle_type (const_tree type)
     return "u13__vector_pair";
   if (type == vector_quad_type_node)
     return "u13__vector_quad";
+  if (type == dmr_type_node)
+    return "u5__dmr";
 
   /* For all other types, use the default mangling.  */
   return NULL;
@@ -22858,6 +22892,10 @@ rs6000_dmr_register_move_cost (machine_mode mode, reg_class_t rclass)
       if (mode == XOmode)
 	return reg_move_base;
 
+      /* __dmr (i.e. TDOmode) is transferred in 2 instructions.  */
+      else if (mode == TDOmode)
+	return reg_move_base * 2;
+
       else
 	return reg_move_base * 2 * hard_regno_nregs (FIRST_DMR_REGNO, mode);
     }
@@ -27545,9 +27583,10 @@ rs6000_split_multireg_move (rtx dst, rtx src)
   mode = GET_MODE (dst);
   nregs = hard_regno_nregs (reg, mode);
 
-  /* If we have a vector quad register for MMA, and this is a load or store,
-     see if we can use vector paired load/stores.  */
-  if (mode == XOmode && TARGET_MMA
+  /* If we have a vector quad register for MMA or DMR register for dense math,
+     and this is a load or store, see if we can use vector paired
+     load/stores.  */
+  if ((mode == XOmode || mode == TDOmode) && TARGET_MMA
       && (MEM_P (dst) || MEM_P (src)))
     {
       reg_mode = OOmode;
@@ -27555,7 +27594,7 @@ rs6000_split_multireg_move (rtx dst, rtx src)
     }
   /* If we have a vector pair/quad mode, split it into two/four separate
      vectors.  */
-  else if (mode == OOmode || mode == XOmode)
+  else if (mode == OOmode || mode == XOmode || mode == TDOmode)
     reg_mode = V1TImode;
   else if (FP_REGNO_P (reg))
     reg_mode = DECIMAL_FLOAT_MODE_P (mode) ? DDmode :
@@ -27601,13 +27640,13 @@ rs6000_split_multireg_move (rtx dst, rtx src)
       return;
     }
 
-  /* The __vector_pair and __vector_quad modes are multi-register
-     modes, so if we have to load or store the registers, we have to be
-     careful to properly swap them if we're in little endian mode
-     below.  This means the last register gets the first memory
-     location.  We also need to be careful of using the right register
-     numbers if we are splitting XO to OO.  */
-  if (mode == OOmode || mode == XOmode)
+  /* The __vector_pair, __vector_quad, and __dmr modes are multi-register
+     modes, so if we have to load or store the registers, we have to be careful
+     to properly swap them if we're in little endian mode below.  This means
+     the last register gets the first memory location.  We also need to be
+     careful of using the right register numbers if we are splitting XO to
+     OO.  */
+  if (mode == OOmode || mode == XOmode || mode == TDOmode)
     {
       nregs = hard_regno_nregs (reg, mode);
       int reg_mode_nregs = hard_regno_nregs (reg, reg_mode);
@@ -27744,7 +27783,7 @@ rs6000_split_multireg_move (rtx dst, rtx src)
 	 overlap.  */
       int i;
       /* XO/OO are opaque so cannot use subregs. */
-      if (mode == OOmode || mode == XOmode )
+      if (mode == OOmode || mode == XOmode || mode == TDOmode)
 	{
 	  for (i = nregs - 1; i >= 0; i--)
 	    {
@@ -27918,7 +27957,7 @@ rs6000_split_multireg_move (rtx dst, rtx src)
 	    continue;
 
 	  /* XO/OO are opaque so cannot use subregs. */
-	  if (mode == OOmode || mode == XOmode )
+	  if (mode == OOmode || mode == XOmode || mode == TDOmode)
 	    {
 	      rtx dst_i = gen_rtx_REG (reg_mode, REGNO (dst) + j);
 	      rtx src_i = gen_rtx_REG (reg_mode, REGNO (src) + j);
@@ -28900,7 +28939,8 @@ rs6000_invalid_conversion (const_tree fromtype, const_tree totype)
 
   if (frommode != tomode)
     {
-      /* Do not allow conversions to/from XOmode and OOmode types.  */
+      /* Do not allow conversions to/from XOmode, OOmode, and TDOmode
+	 types.  */
       if (frommode == XOmode)
 	return N_("invalid conversion from type %<__vector_quad%>");
       if (tomode == XOmode)
@@ -28909,6 +28949,10 @@ rs6000_invalid_conversion (const_tree fromtype, const_tree totype)
 	return N_("invalid conversion from type %<__vector_pair%>");
       if (tomode == OOmode)
 	return N_("invalid conversion to type %<__vector_pair%>");
+      if (frommode == TDOmode)
+	return N_("invalid conversion from type %<__dmr%>");
+      if (tomode == TDOmode)
+	return N_("invalid conversion to type %<__dmr%>");
     }
 
   /* Conversion allowed.  */
diff --git a/gcc/config/rs6000/rs6000.h b/gcc/config/rs6000/rs6000.h
index f5d144cbb12..67ef3d3a7d0 100644
--- a/gcc/config/rs6000/rs6000.h
+++ b/gcc/config/rs6000/rs6000.h
@@ -1009,7 +1009,7 @@ enum data_align { align_abi, align_opt, align_both };
 /* Modes that are not vectors, but require vector alignment.  Treat these like
    vectors in terms of loads and stores.  */
 #define VECTOR_ALIGNMENT_P(MODE)					\
-  (FLOAT128_VECTOR_P (MODE) || (MODE) == OOmode || (MODE) == XOmode)
+  (FLOAT128_VECTOR_P (MODE) || OPAQUE_MODE_P (MODE))
 
 #define ALTIVEC_VECTOR_MODE(MODE)					\
   ((MODE) == V16QImode							\
@@ -2298,6 +2298,7 @@ enum rs6000_builtin_type_index
   RS6000_BTI_const_str,		 /* pointer to const char * */
   RS6000_BTI_vector_pair,	 /* unsigned 256-bit types (vector pair).  */
   RS6000_BTI_vector_quad,	 /* unsigned 512-bit types (vector quad).  */
+  RS6000_BTI_dmr,		 /* unsigned 1,024-bit types (dmr).  */
   RS6000_BTI_const_ptr_void,     /* const pointer to void */
   RS6000_BTI_ptr_V16QI,
   RS6000_BTI_ptr_V1TI,
@@ -2336,6 +2337,7 @@ enum rs6000_builtin_type_index
   RS6000_BTI_ptr_dfloat128,
   RS6000_BTI_ptr_vector_pair,
   RS6000_BTI_ptr_vector_quad,
+  RS6000_BTI_ptr_dmr,
   RS6000_BTI_ptr_long_long,
   RS6000_BTI_ptr_long_long_unsigned,
   RS6000_BTI_MAX
@@ -2393,6 +2395,7 @@ enum rs6000_builtin_type_index
 #define const_str_type_node		 (rs6000_builtin_types[RS6000_BTI_const_str])
 #define vector_pair_type_node		 (rs6000_builtin_types[RS6000_BTI_vector_pair])
 #define vector_quad_type_node		 (rs6000_builtin_types[RS6000_BTI_vector_quad])
+#define dmr_type_node			 (rs6000_builtin_types[RS6000_BTI_dmr])
 #define pcvoid_type_node		 (rs6000_builtin_types[RS6000_BTI_const_ptr_void])
 #define ptr_V16QI_type_node		 (rs6000_builtin_types[RS6000_BTI_ptr_V16QI])
 #define ptr_V1TI_type_node		 (rs6000_builtin_types[RS6000_BTI_ptr_V1TI])
@@ -2431,6 +2434,7 @@ enum rs6000_builtin_type_index
 #define ptr_dfloat128_type_node		 (rs6000_builtin_types[RS6000_BTI_ptr_dfloat128])
 #define ptr_vector_pair_type_node	 (rs6000_builtin_types[RS6000_BTI_ptr_vector_pair])
 #define ptr_vector_quad_type_node	 (rs6000_builtin_types[RS6000_BTI_ptr_vector_quad])
+#define ptr_dmr_type_node		 (rs6000_builtin_types[RS6000_BTI_ptr_dmr])
 #define ptr_long_long_integer_type_node	 (rs6000_builtin_types[RS6000_BTI_ptr_long_long])
 #define ptr_long_long_unsigned_type_node (rs6000_builtin_types[RS6000_BTI_ptr_long_long_unsigned])
 
diff --git a/gcc/testsuite/gcc.target/powerpc/dm-1024bit.c b/gcc/testsuite/gcc.target/powerpc/dm-1024bit.c
new file mode 100644
index 00000000000..0a9884ddf63
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/dm-1024bit.c
@@ -0,0 +1,63 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target powerpc_dense_math_ok } */
+/* { dg-options "-mdejagnu-cpu=future -O2" } */
+
+/* Test basic load/store for __dmr type.  */
+
+#ifndef CONSTRAINT
+#if defined(USE_D)
+#define CONSTRAINT "d"
+
+#elif defined(USE_V)
+#define CONSTRAINT "v"
+
+#elif defined(USE_WA)
+#define CONSTRAINT "wa"
+
+#else
+#define CONSTRAINT "wD"
+#endif
+#endif
+const char constraint[] = CONSTRAINT;
+
+void foo_mem_asm (__dmr *p, __dmr *q)
+{
+  /* 2 LXVP instructions.  */
+  __dmr vq = *p;
+
+  /* 2 DMXXINSTDMR512 instructions to transfer VSX to DMR.  */
+  __asm__ ("# foo (" CONSTRAINT ") %A0" : "+" CONSTRAINT (vq));
+  /* 2 DMXXEXTFDMR512 instructions to transfer DMR to VSX.  */
+
+  /* 2 STXVP instructions.  */
+  *q = vq;
+}
+
+void foo_mem_asm2 (__dmr *p, __dmr *q)
+{
+  /* 2 LXVP instructions.  */
+  __dmr vq = *p;
+  __dmr vq2;
+  __dmr vq3;
+
+  /* 2 DMXXINSTDMR512 instructions to transfer VSX to DMR.  */
+  __asm__ ("# foo1 (" CONSTRAINT ") %A0" : "+" CONSTRAINT (vq));
+  /* 2 DMXXEXTFDMR512 instructions to transfer DMR to VSX.  */
+
+  vq2 = vq;
+  __asm__ ("# foo2 (wa) %0" : "+wa" (vq2));
+
+  /* 2 STXVP instructions.  */
+  *q = vq2;
+}
+
+void foo_mem (__dmr *p, __dmr *q)
+{
+  /* 2 LXVP, 2 STXVP instructions, no DMR transfer.  */
+  *q = *p;
+}
+
+/* { dg-final { scan-assembler-times {\mdmxxextfdmr512\M}  4 } } */
+/* { dg-final { scan-assembler-times {\mdmxxinstdmr512\M}  4 } } */
+/* { dg-final { scan-assembler-times {\mlxvp\M}           12 } } */
+/* { dg-final { scan-assembler-times {\mstxvp\M}          12 } } */