diff mbox

[2/2,ARM/AArch64] Add a new Cortex-A53 scheduling model

Message ID 20151001094126.GA23247@arm.com
State Accepted
Commit cdc1afa3c04e6ddf2d2119a4144754cabcfe419a
Headers show

Commit Message

James Greenhalgh Oct. 1, 2015, 9:41 a.m. UTC
On Thu, Oct 01, 2015 at 09:33:07AM +0100, Marcus Shawcroft wrote:
> On 25/09/15 08:59, James Greenhalgh wrote:
> >
> > Hi,
> >
> > This patch introduces a new scheduling model for Cortex-A53.
> >
> > Bootstrapped and tested on arm-none-linux-gnueabi and aarch64-none-linux-gnu
> > and checked with a variety of popular benchmarking and microbenchmarking
> > suites to show a benefit.
> >
> > OK?
> >
> > Thanks,
> > James
> >
> > ---
> > 2015-09-25  James Greenhalgh  <james.greenhalgh@arm.com>
> >
> > 	* config/arm/aarch-common-protos.h
> > 	(aarch_accumulator_forwarding): New.
> > 	(aarch_forward_to_shift_is_not_shifted_reg): Likewise.
> > 	* config/arm/aarch-common.c (aarch_accumulator_forwarding): New.
> > 	(aarch_forward_to_shift_is_not_shifted_reg): Liekwise.
> > 	* config/arm/cortex-a53.md: Rewrite.
> >
> 
> OK aarch64 with Kyrill's comments fixed.
> /M

Thanks,

I had to rebase this over Evandro's recent patch adding neon_ldp/neon_ldp_q
types to the old scheduling model. The rebase was obvious to resolve, and
while I was there I also added the neon_stp/neon_stp_q types which were
missing.

I've attached what I ultimately committed as revision 228324. I messed up
fixing the ChangeLog typo before commit, so that is revision 228325.

Thanks,
James

Comments

Christophe Lyon Oct. 5, 2015, 10:07 a.m. UTC | #1
On 1 October 2015 at 11:41, James Greenhalgh <james.greenhalgh@arm.com> wrote:
> On Thu, Oct 01, 2015 at 09:33:07AM +0100, Marcus Shawcroft wrote:
>> On 25/09/15 08:59, James Greenhalgh wrote:
>> >
>> > Hi,
>> >
>> > This patch introduces a new scheduling model for Cortex-A53.
>> >
>> > Bootstrapped and tested on arm-none-linux-gnueabi and aarch64-none-linux-gnu
>> > and checked with a variety of popular benchmarking and microbenchmarking
>> > suites to show a benefit.
>> >
>> > OK?
>> >
>> > Thanks,
>> > James
>> >
>> > ---
>> > 2015-09-25  James Greenhalgh  <james.greenhalgh@arm.com>
>> >
>> >     * config/arm/aarch-common-protos.h
>> >     (aarch_accumulator_forwarding): New.
>> >     (aarch_forward_to_shift_is_not_shifted_reg): Likewise.
>> >     * config/arm/aarch-common.c (aarch_accumulator_forwarding): New.
>> >     (aarch_forward_to_shift_is_not_shifted_reg): Liekwise.
>> >     * config/arm/cortex-a53.md: Rewrite.
>> >
>>
>> OK aarch64 with Kyrill's comments fixed.
>> /M
>
> Thanks,
>
> I had to rebase this over Evandro's recent patch adding neon_ldp/neon_ldp_q
> types to the old scheduling model. The rebase was obvious to resolve, and
> while I was there I also added the neon_stp/neon_stp_q types which were
> missing.
>
> I've attached what I ultimately committed as revision 228324. I messed up
> fixing the ChangeLog typo before commit, so that is revision 228325.
>

Hi James,

Since this commit I can see
gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c fail at -O2
/aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c:
In function 'exec_vst1_lane':
/aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c:96:1:
internal compiler error: output_operand: invalid %-code
0x78f79e output_operand_lossage(char const*, ...)
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3417
0x7934f3 output_asm_insn(char const*, rtx_def**)
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3782
0x793d77 final_scan_insn(rtx_insn*, _IO_FILE*, int, int, int*)
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3029
0x794b3a final(rtx_insn*, _IO_FILE*, int)
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:2058
0x7956fb rest_of_handle_final
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:4449
0x7956fb execute
        /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:4524
Please submit a full bug report,

on aarch64_be.

I haven't looked at it in more details though.


> Thanks,
> James
>
James Greenhalgh Oct. 5, 2015, 10:36 a.m. UTC | #2
On Mon, Oct 05, 2015 at 11:07:45AM +0100, Christophe Lyon wrote:
> On 1 October 2015 at 11:41, James Greenhalgh <james.greenhalgh@arm.com> wrote:
> > On Thu, Oct 01, 2015 at 09:33:07AM +0100, Marcus Shawcroft wrote:
> >> On 25/09/15 08:59, James Greenhalgh wrote:
> >> >
> >> > Hi,
> >> >
> >> > This patch introduces a new scheduling model for Cortex-A53.
> >> >
> >> > Bootstrapped and tested on arm-none-linux-gnueabi and aarch64-none-linux-gnu
> >> > and checked with a variety of popular benchmarking and microbenchmarking
> >> > suites to show a benefit.
> >> >
> >> > OK?
> >> >
> >> > Thanks,
> >> > James
> >> >
> >> > ---
> >> > 2015-09-25  James Greenhalgh  <james.greenhalgh@arm.com>
> >> >
> >> >     * config/arm/aarch-common-protos.h
> >> >     (aarch_accumulator_forwarding): New.
> >> >     (aarch_forward_to_shift_is_not_shifted_reg): Likewise.
> >> >     * config/arm/aarch-common.c (aarch_accumulator_forwarding): New.
> >> >     (aarch_forward_to_shift_is_not_shifted_reg): Liekwise.
> >> >     * config/arm/cortex-a53.md: Rewrite.
> >> >
> >>
> >> OK aarch64 with Kyrill's comments fixed.
> >> /M
> >
> > Thanks,
> >
> > I had to rebase this over Evandro's recent patch adding neon_ldp/neon_ldp_q
> > types to the old scheduling model. The rebase was obvious to resolve, and
> > while I was there I also added the neon_stp/neon_stp_q types which were
> > missing.
> >
> > I've attached what I ultimately committed as revision 228324. I messed up
> > fixing the ChangeLog typo before commit, so that is revision 228325.
> >
> 
> Hi James,
> 
> Since this commit I can see
> gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c fail at -O2
> /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c:
> In function 'exec_vst1_lane':
> /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vst1_lane.c:96:1:
> internal compiler error: output_operand: invalid %-code
> 0x78f79e output_operand_lossage(char const*, ...)
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3417
> 0x7934f3 output_asm_insn(char const*, rtx_def**)
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3782
> 0x793d77 final_scan_insn(rtx_insn*, _IO_FILE*, int, int, int*)
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:3029
> 0x794b3a final(rtx_insn*, _IO_FILE*, int)
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:2058
> 0x7956fb rest_of_handle_final
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:4449
> 0x7956fb execute
>         /aci-gcc-fsf/sources/gcc-fsf/gccsrc/gcc/final.c:4524
> Please submit a full bug report,
> 
> on aarch64_be.
> 
> I haven't looked at it in more details though.

Hi Christophe,

Thanks for the report, I'd be surprised if that was to do with the
scheduling model. I can reproduce the failure, and expect that Kugan's
patch at https://gcc.gnu.org/ml/gcc-patches/2015-10/msg00129.html ought
to do the job of fixing the ICE.

Thanks,
James
diff mbox

Patch

Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog	(revision 228323)
+++ gcc/ChangeLog	(working copy)
@@ -1,3 +1,12 @@ 
+2015-10-01  James Greenhalgh  <james.greenhalgh@arm.com>
+
+	* config/arm/aarch-common-protos.h
+	(aarch_accumulator_forwarding): New.
+	(aarch_forward_to_shift_is_not_shifted_reg): Likewise.
+	* config/arm/aarch-common.c (aarch_accumulator_forwarding): New.
+	(aarch_forward_to_shift_is_not_shifted_reg): Liekwise.
+	* config/arm/cortex-a53.md: Rewrite.
+
 2015-10-01  Richard Biener  <rguenther@suse.de>
 
 	* gimple-match.h (mprts_hook): Declare.
Index: gcc/config/arm/aarch-common-protos.h
===================================================================
--- gcc/config/arm/aarch-common-protos.h	(revision 228323)
+++ gcc/config/arm/aarch-common-protos.h	(working copy)
@@ -23,7 +23,9 @@ 
 #ifndef GCC_AARCH_COMMON_PROTOS_H
 #define GCC_AARCH_COMMON_PROTOS_H
 
+extern int aarch_accumulator_forwarding (rtx_insn *, rtx_insn *);
 extern int aarch_crypto_can_dual_issue (rtx_insn *, rtx_insn *);
+extern int aarch_forward_to_shift_is_not_shifted_reg (rtx_insn *, rtx_insn *);
 extern bool aarch_rev16_p (rtx);
 extern bool aarch_rev16_shleft_mask_imm_p (rtx, machine_mode);
 extern bool aarch_rev16_shright_mask_imm_p (rtx, machine_mode);
Index: gcc/config/arm/aarch-common.c
===================================================================
--- gcc/config/arm/aarch-common.c	(revision 228323)
+++ gcc/config/arm/aarch-common.c	(working copy)
@@ -394,6 +394,112 @@ 
           && !reg_overlap_mentioned_p (result, op1));
 }
 
+/* Return non-zero if the destination of PRODUCER feeds the accumulator
+   operand of an MLA-like operation.  */
+
+int
+aarch_accumulator_forwarding (rtx_insn *producer, rtx_insn *consumer)
+{
+  rtx producer_set = single_set (producer);
+  rtx consumer_set = single_set (consumer);
+
+  /* We are looking for a SET feeding a SET.  */
+  if (!producer_set || !consumer_set)
+    return 0;
+
+  rtx dest = SET_DEST (producer_set);
+  rtx mla = SET_SRC (consumer_set);
+
+  /* We're looking for a register SET.  */
+  if (!REG_P (dest))
+    return 0;
+
+  rtx accumulator;
+
+  /* Strip a zero_extend.  */
+  if (GET_CODE (mla) == ZERO_EXTEND)
+    mla = XEXP (mla, 0);
+
+  switch (GET_CODE (mla))
+    {
+    case PLUS:
+      /* Possibly an MADD.  */
+      if (GET_CODE (XEXP (mla, 0)) == MULT)
+	accumulator = XEXP (mla, 1);
+      else
+	return 0;
+      break;
+    case MINUS:
+      /* Possibly an MSUB.  */
+      if (GET_CODE (XEXP (mla, 1)) == MULT)
+	accumulator = XEXP (mla, 0);
+      else
+	return 0;
+      break;
+    case FMA:
+	{
+	  /* Possibly an FMADD/FMSUB/FNMADD/FNMSUB.  */
+	  if (REG_P (XEXP (mla, 1))
+	      && REG_P (XEXP (mla, 2))
+	      && (REG_P (XEXP (mla, 0))
+		  || GET_CODE (XEXP (mla, 0)) == NEG))
+
+	    {
+	      /* FMADD/FMSUB.  */
+	      accumulator = XEXP (mla, 2);
+	    }
+	  else if (REG_P (XEXP (mla, 1))
+		   && GET_CODE (XEXP (mla, 2)) == NEG
+		   && (REG_P (XEXP (mla, 0))
+		       || GET_CODE (XEXP (mla, 0)) == NEG))
+	    {
+	      /* FNMADD/FNMSUB.  */
+	      accumulator = XEXP (XEXP (mla, 2), 0);
+	    }
+	  else
+	    return 0;
+	  break;
+	}
+      default:
+	/* Not an MLA-like operation.  */
+	return 0;
+    }
+
+  return (REGNO (dest) == REGNO (accumulator));
+}
+
+/* Return nonzero if the CONSUMER instruction is some sort of
+   arithmetic or logic + shift operation, and the register we are
+   writing in PRODUCER is not used in a register shift by register
+   operation.  */
+
+int
+aarch_forward_to_shift_is_not_shifted_reg (rtx_insn *producer,
+					   rtx_insn *consumer)
+{
+  rtx value, op;
+  rtx early_op;
+
+  if (!arm_get_set_operands (producer, consumer, &value, &op))
+    return 0;
+
+  if ((early_op = arm_find_shift_sub_rtx (op)))
+    {
+      if (REG_P (early_op))
+	early_op = op;
+
+      /* Any other canonicalisation of a shift is a shift-by-constant
+	 so we don't care.  */
+      if (GET_CODE (early_op) == ASHIFT)
+	return (!REG_P (XEXP (early_op, 0))
+		|| !REG_P (XEXP (early_op, 1)));
+      else
+	return 1;
+    }
+
+  return 0;
+}
+
 /* Return non-zero if the consumer (a multiply-accumulate instruction)
    has an accumulator dependency on the result of the producer (a
    multiplication instruction) and no other dependency on that result.  */
Index: gcc/config/arm/cortex-a53.md
===================================================================
--- gcc/config/arm/cortex-a53.md	(revision 228323)
+++ gcc/config/arm/cortex-a53.md	(working copy)
@@ -22,345 +22,700 @@ 
 (define_automaton "cortex_a53")
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; Functional units.
+;; General-purpose functional units.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; There are two main integer execution pipelines, described as
-;; slot 0 and issue slot 1.
+;; We use slot0 and slot1 to model constraints on which instructions may
+;; dual-issue.
 
 (define_cpu_unit "cortex_a53_slot0" "cortex_a53")
 (define_cpu_unit "cortex_a53_slot1" "cortex_a53")
 
-(define_reservation "cortex_a53_slot_any" "cortex_a53_slot0|cortex_a53_slot1")
-(define_reservation "cortex_a53_single_issue" "cortex_a53_slot0+cortex_a53_slot1")
+(define_reservation "cortex_a53_slot_any"
+		    "cortex_a53_slot0\
+		     |cortex_a53_slot1")
 
-;; The load/store pipeline.  Load/store instructions can dual-issue from
-;; either pipeline, but two load/stores cannot simultaneously issue.
+(define_reservation "cortex_a53_single_issue"
+		    "cortex_a53_slot0\
+		     +cortex_a53_slot1")
 
-(define_cpu_unit "cortex_a53_ls" "cortex_a53")
+;; Used to model load and store pipelines.  Load/store instructions
+;; can dual-issue with other instructions, but two load/stores cannot
+;; simultaneously issue.
 
-;; The store pipeline.  Shared between both execution pipelines.
-
 (define_cpu_unit "cortex_a53_store" "cortex_a53")
+(define_cpu_unit "cortex_a53_load" "cortex_a53")
+(define_cpu_unit "cortex_a53_ls_agen" "cortex_a53")
 
-;; The branch pipeline.  Branches can dual-issue with other instructions
-;; (except when those instructions take multiple cycles to issue).
+;; Used to model a branch pipeline.  Branches can dual-issue with other
+;; instructions (except when those instructions take multiple cycles
+;; to issue).
 
 (define_cpu_unit "cortex_a53_branch" "cortex_a53")
 
-;; The integer divider.
+;; Used to model an integer divide pipeline.
 
 (define_cpu_unit "cortex_a53_idiv" "cortex_a53")
 
-;; The floating-point add pipeline used to model the usage
-;; of the add pipeline by fmac instructions.
+;; Used to model an integer multiply/multiply-accumulate pipeline.
 
-(define_cpu_unit "cortex_a53_fpadd_pipe" "cortex_a53")
+(define_cpu_unit "cortex_a53_imul" "cortex_a53")
 
-;; Floating-point div/sqrt (long latency, out-of-order completion).
+;; Model general structural hazards, for wherever we need them.
 
-(define_cpu_unit "cortex_a53_fp_div_sqrt" "cortex_a53")
+(define_cpu_unit "cortex_a53_hazard" "cortex_a53")
 
-;; The Advanced SIMD pipelines.
-
-(define_cpu_unit "cortex_a53_simd0" "cortex_a53")
-(define_cpu_unit "cortex_a53_simd1" "cortex_a53")
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; ALU instructions.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn_reservation "cortex_a53_alu" 2
+(define_insn_reservation "cortex_a53_shift" 2
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "alu_imm,alus_imm,logic_imm,logics_imm,\
-                        alu_sreg,alus_sreg,logic_reg,logics_reg,\
-                        adc_imm,adcs_imm,adc_reg,adcs_reg,\
-                        adr,bfm,csel,clz,rbit,rev,alu_dsp_reg,\
-                        rotate_imm,shift_imm,shift_reg,\
-                        mov_imm,mov_reg,mvn_imm,mvn_reg,\
-                        mrs,multiple,no_insn"))
+       (eq_attr "type" "adr,shift_imm,shift_reg,mov_imm,mvn_imm"))
   "cortex_a53_slot_any")
 
-(define_insn_reservation "cortex_a53_alu_shift" 2
+(define_insn_reservation "cortex_a53_alu_rotate_imm" 2
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "alu_shift_imm,alus_shift_imm,\
-                        crc,logic_shift_imm,logics_shift_imm,\
-                        alu_ext,alus_ext,alu_shift_reg,alus_shift_reg,\
-                        logic_shift_reg,logics_shift_reg,\
-                        extend,mov_shift,mov_shift_reg,\
-                        mvn_shift,mvn_shift_reg"))
+       (eq_attr "type" "rotate_imm"))
+  "(cortex_a53_slot1)
+   | (cortex_a53_single_issue)")
+
+(define_insn_reservation "cortex_a53_alu" 3
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "type" "alu_imm,alus_imm,logic_imm,logics_imm,
+			alu_sreg,alus_sreg,logic_reg,logics_reg,
+			adc_imm,adcs_imm,adc_reg,adcs_reg,
+			bfm,csel,clz,rbit,rev,alu_dsp_reg,
+			mov_reg,mvn_reg,
+			mrs,multiple,no_insn"))
   "cortex_a53_slot_any")
 
-;; Forwarding path for unshifted operands.
+(define_insn_reservation "cortex_a53_alu_shift" 3
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "type" "alu_shift_imm,alus_shift_imm,
+			crc,logic_shift_imm,logics_shift_imm,
+			alu_ext,alus_ext,
+			extend,mov_shift,mvn_shift"))
+  "cortex_a53_slot_any")
 
-(define_bypass 1 "cortex_a53_alu,cortex_a53_alu_shift"
-  "cortex_a53_alu")
+(define_insn_reservation "cortex_a53_alu_shift_reg" 3
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "type" "alu_shift_reg,alus_shift_reg,
+			logic_shift_reg,logics_shift_reg,
+			mov_shift_reg,mvn_shift_reg"))
+  "cortex_a53_slot_any+cortex_a53_hazard")
 
-(define_bypass 1 "cortex_a53_alu,cortex_a53_alu_shift"
-  "cortex_a53_alu_shift"
-  "arm_no_early_alu_shift_dep")
-
-;; The multiplier pipeline can forward results so there's no need to specify
-;; bypasses. Multiplies can only single-issue currently.
-
 (define_insn_reservation "cortex_a53_mul" 3
   (and (eq_attr "tune" "cortexa53")
        (ior (eq_attr "mul32" "yes")
-            (eq_attr "mul64" "yes")))
-  "cortex_a53_single_issue")
+	    (eq_attr "mul64" "yes")))
+  "cortex_a53_slot_any+cortex_a53_imul")
 
-;; A multiply with a single-register result or an MLA, followed by an
-;; MLA with an accumulator dependency, has its result forwarded so two
-;; such instructions can issue back-to-back.
+;; From the perspective of the GCC scheduling state machine, if we wish to
+;; model an instruction as serialising other instructions, we are best to do
+;; so by modelling it as taking very few cycles.  Scheduling many other
+;; instructions underneath it at the cost of freedom to pick from the
+;; ready list is likely to hurt us more than it helps.  However, we do
+;; want to model some resource and latency cost for divide instructions in
+;; order to avoid divides ending up too lumpy.
 
-(define_bypass 1 "cortex_a53_mul"
-               "cortex_a53_mul"
-               "arm_mac_accumulator_is_mul_result")
-
-;; Punt with a high enough latency for divides.
-(define_insn_reservation "cortex_a53_udiv" 8
+(define_insn_reservation "cortex_a53_div" 4
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "udiv"))
-  "(cortex_a53_slot0+cortex_a53_idiv),cortex_a53_idiv*7")
+       (eq_attr "type" "udiv,sdiv"))
+  "cortex_a53_slot0,cortex_a53_idiv*2")
 
-(define_insn_reservation "cortex_a53_sdiv" 9
-  (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "sdiv"))
-  "(cortex_a53_slot0+cortex_a53_idiv),cortex_a53_idiv*8")
-
-
-(define_bypass 2 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv"
-               "cortex_a53_alu")
-(define_bypass 2 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv"
-               "cortex_a53_alu_shift"
-               "arm_no_early_alu_shift_dep")
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; Load/store instructions.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; Address-generation happens in the issue stage.
+;; TODO: load<n> is not prescriptive about how much data is to be loaded.
+;; This is most obvious for LDRD from AArch32 and LDP (X register) from
+;; AArch64, both are tagged load2 but LDP will load 128-bits compared to
+;; LDRD which is 64-bits.
+;;
+;; For the below, we assume AArch64 X-registers for load2, and AArch32
+;; registers for load3/load4.
 
-(define_insn_reservation "cortex_a53_load1" 3
+(define_insn_reservation "cortex_a53_load1" 4
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "load_byte,load1,load_acq"))
-  "cortex_a53_slot_any+cortex_a53_ls")
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_load")
 
 (define_insn_reservation "cortex_a53_store1" 2
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "store1,store_rel"))
-  "cortex_a53_slot_any+cortex_a53_ls+cortex_a53_store")
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_store")
 
-(define_insn_reservation "cortex_a53_load2" 3
+;; Model AArch64-sized LDP Xm, Xn, [Xa]
+
+(define_insn_reservation "cortex_a53_load2" 4
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "load2"))
-  "cortex_a53_single_issue+cortex_a53_ls")
+  "cortex_a53_single_issue+cortex_a53_ls_agen,
+   cortex_a53_load+cortex_a53_slot0,
+   cortex_a53_load")
 
 (define_insn_reservation "cortex_a53_store2" 2
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "store2"))
-  "cortex_a53_single_issue+cortex_a53_ls+cortex_a53_store")
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_store")
 
-(define_insn_reservation "cortex_a53_load3plus" 4
+;; Model AArch32-sized LDM Ra, {Rm, Rn, Ro}
+
+(define_insn_reservation "cortex_a53_load3plus" 6
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "load3,load4"))
-  "(cortex_a53_single_issue+cortex_a53_ls)*2")
+  "cortex_a53_single_issue+cortex_a53_ls_agen,
+   cortex_a53_load+cortex_a53_slot0,
+   cortex_a53_load")
 
-(define_insn_reservation "cortex_a53_store3plus" 3
+(define_insn_reservation "cortex_a53_store3plus" 2
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "store3,store4"))
-  "(cortex_a53_single_issue+cortex_a53_ls+cortex_a53_store)*2")
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_store+cortex_a53_slot0,
+   cortex_a53_store")
 
-;; Load/store addresses are required early in Issue.
-(define_bypass 3 "cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus,cortex_a53_alu,cortex_a53_alu_shift"
-                 "cortex_a53_load*"
-                 "arm_early_load_addr_dep")
-(define_bypass 3 "cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus,cortex_a53_alu,cortex_a53_alu_shift"
-                 "cortex_a53_store*"
-                 "arm_early_store_addr_dep")
-
-;; Load data can forward in the ALU pipeline
-(define_bypass 2 "cortex_a53_load1,cortex_a53_load2"
-               "cortex_a53_alu")
-(define_bypass 2 "cortex_a53_load1,cortex_a53_load2"
-               "cortex_a53_alu_shift"
-               "arm_no_early_alu_shift_dep")
-
-;; ALU ops can forward to stores.
-(define_bypass 0 "cortex_a53_alu,cortex_a53_alu_shift"
-                 "cortex_a53_store1,cortex_a53_store2,cortex_a53_store3plus"
-                 "arm_no_early_store_addr_dep")
-
-(define_bypass 1 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv,cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus"
-                 "cortex_a53_store1,cortex_a53_store2,cortex_a53_store3plus"
-                 "arm_no_early_store_addr_dep")
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; Branches.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; Currently models all branches as dual-issuable from either execution
-;; slot, which isn't true for all cases. We still need to model indirect
-;; branches.
+;; Model all branches as dual-issuable from either execution, which
+;; is not strictly true for all cases (indirect branches).
 
 (define_insn_reservation "cortex_a53_branch" 0
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "branch,call"))
-  "cortex_a53_slot_any+cortex_a53_branch")
+  "cortex_a53_slot_any,cortex_a53_branch")
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; General-purpose register bypasses
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; Model bypasses for unshifted operands to ALU instructions.
+
+(define_bypass 1 "cortex_a53_shift"
+		 "cortex_a53_shift")
+
+(define_bypass 1 "cortex_a53_alu,
+		  cortex_a53_alu_shift*,
+		  cortex_a53_alu_rotate_imm,
+		  cortex_a53_shift"
+		 "cortex_a53_alu")
+
+(define_bypass 2 "cortex_a53_alu,
+		  cortex_a53_alu_shift*"
+		 "cortex_a53_alu_shift*"
+		 "aarch_forward_to_shift_is_not_shifted_reg")
+
+;; In our model, we allow any general-purpose register operation to
+;; bypass to the accumulator operand of an integer MADD-like operation.
+
+(define_bypass 1 "cortex_a53_alu*,
+		  cortex_a53_load*,
+		  cortex_a53_mul"
+		 "cortex_a53_mul"
+		 "aarch_accumulator_forwarding")
+
+;; Model a bypass from MLA/MUL to many ALU instructions.
+
+(define_bypass 2 "cortex_a53_mul"
+		 "cortex_a53_alu,
+		  cortex_a53_alu_shift*")
+
+;; We get neater schedules by allowing an MLA/MUL to feed an
+;; early load address dependency to a load.
+
+(define_bypass 2 "cortex_a53_mul"
+		 "cortex_a53_load*"
+		 "arm_early_load_addr_dep")
+
+;; Model bypasses for loads which are to be consumed by the ALU.
+
+(define_bypass 2 "cortex_a53_load1"
+		 "cortex_a53_alu")
+
+(define_bypass 3 "cortex_a53_load1"
+		 "cortex_a53_alu_shift*")
+
+;; Model a bypass for ALU instructions feeding stores.
+
+(define_bypass 1 "cortex_a53_alu*"
+		 "cortex_a53_store1,
+		  cortex_a53_store2,
+		  cortex_a53_store3plus"
+		 "arm_no_early_store_addr_dep")
+
+;; Model a bypass for load and multiply instructions feeding stores.
+
+(define_bypass 2 "cortex_a53_mul,
+		  cortex_a53_load1,
+		  cortex_a53_load2,
+		  cortex_a53_load3plus"
+		 "cortex_a53_store1,
+		  cortex_a53_store2,
+		  cortex_a53_store3plus"
+		 "arm_no_early_store_addr_dep")
+
+;; Model a GP->FP register move as similar to stores.
+
+(define_bypass 1 "cortex_a53_alu*"
+		 "cortex_a53_r2f")
+
+(define_bypass 2 "cortex_a53_mul,
+		  cortex_a53_load1,
+		  cortex_a53_load2,
+		  cortex_a53_load3plus"
+		 "cortex_a53_r2f")
+
+;; Shifts feeding Load/Store addresses may not be ready in time.
+
+(define_bypass 3 "cortex_a53_shift"
+		 "cortex_a53_load*"
+		 "arm_early_load_addr_dep")
+
+(define_bypass 3 "cortex_a53_shift"
+		 "cortex_a53_store*"
+		 "arm_early_store_addr_dep")
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Floating-point/Advanced SIMD.
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(define_automaton "cortex_a53_advsimd")
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Broad Advanced SIMD type categorisation
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(define_attr "cortex_a53_advsimd_type"
+  "advsimd_alu, advsimd_alu_q,
+   advsimd_mul, advsimd_mul_q,
+   advsimd_div_s, advsimd_div_s_q,
+   advsimd_div_d, advsimd_div_d_q,
+   advsimd_load_64, advsimd_store_64,
+   advsimd_load_128, advsimd_store_128,
+   advsimd_load_lots, advsimd_store_lots,
+   unknown"
+  (cond [
+    (eq_attr "type" "neon_add, neon_qadd, neon_add_halve, neon_sub, neon_qsub,\
+		     neon_sub_halve, neon_abs, neon_neg, neon_qneg,\
+		     neon_qabs, neon_abd, neon_minmax, neon_compare,\
+		     neon_compare_zero, neon_arith_acc, neon_reduc_add,\
+		     neon_reduc_add_acc, neon_reduc_minmax,\
+		     neon_logic, neon_tst, neon_shift_imm,\
+		     neon_shift_reg, neon_shift_acc, neon_sat_shift_imm,\
+		     neon_sat_shift_reg, neon_ins, neon_move,\
+		     neon_permute, neon_zip, neon_tbl1,\
+		     neon_tbl2, neon_tbl3, neon_tbl4, neon_bsl,\
+		     neon_cls, neon_cnt, neon_dup,\
+		     neon_ext, neon_rbit, neon_rev,\
+		     neon_fp_abd_s, neon_fp_abd_d,\
+		     neon_fp_abs_s, neon_fp_abs_d,\
+		     neon_fp_addsub_s, neon_fp_addsub_d, neon_fp_compare_s,\
+		     neon_fp_compare_d, neon_fp_minmax_s,\
+		     neon_fp_minmax_d, neon_fp_neg_s, neon_fp_neg_d,\
+		     neon_fp_reduc_add_s, neon_fp_reduc_add_d,\
+		     neon_fp_reduc_minmax_s, neon_fp_reduc_minmax_d,\
+		     neon_fp_cvt_widen_h, neon_fp_to_int_s,neon_fp_to_int_d,\
+		     neon_int_to_fp_s, neon_int_to_fp_d, neon_fp_round_s,\
+		     neon_fp_recpe_s, neon_fp_recpe_d, neon_fp_recps_s,\
+		     neon_fp_recps_d, neon_fp_recpx_s, neon_fp_recpx_d,\
+		     neon_fp_rsqrte_s, neon_fp_rsqrte_d, neon_fp_rsqrts_s,\
+		     neon_fp_rsqrts_d")
+      (const_string "advsimd_alu")
+    (eq_attr "type" "neon_add_q, neon_add_widen, neon_add_long,\
+		     neon_qadd_q, neon_add_halve_q, neon_add_halve_narrow_q,\
+		     neon_sub_q, neon_sub_widen, neon_sub_long,\
+		     neon_qsub_q, neon_sub_halve_q, neon_sub_halve_narrow_q,\
+		     neon_abs_q, neon_neg_q, neon_qneg_q, neon_qabs_q,\
+		     neon_abd_q, neon_abd_long, neon_minmax_q,\
+		     neon_compare_q, neon_compare_zero_q,\
+		     neon_arith_acc_q, neon_reduc_add_q,\
+		     neon_reduc_add_long, neon_reduc_add_acc_q,\
+		     neon_reduc_minmax_q, neon_logic_q, neon_tst_q,\
+		     neon_shift_imm_q, neon_shift_imm_narrow_q,\
+		     neon_shift_imm_long, neon_shift_reg_q,\
+		     neon_shift_acc_q, neon_sat_shift_imm_q,\
+		     neon_sat_shift_imm_narrow_q, neon_sat_shift_reg_q,\
+		     neon_ins_q, neon_move_q, neon_move_narrow_q,\
+		     neon_permute_q, neon_zip_q,\
+		     neon_tbl1_q, neon_tbl2_q, neon_tbl3_q,\
+		     neon_tbl4_q, neon_bsl_q, neon_cls_q, neon_cnt_q,\
+		     neon_dup_q, neon_ext_q, neon_rbit_q,\
+		     neon_rev_q, neon_fp_abd_s_q, neon_fp_abd_d_q,\
+		     neon_fp_abs_s_q, neon_fp_abs_d_q,\
+		     neon_fp_addsub_s_q, neon_fp_addsub_d_q,\
+		     neon_fp_compare_s_q, neon_fp_compare_d_q,\
+		     neon_fp_minmax_s_q, neon_fp_minmax_d_q,\
+		     neon_fp_cvt_widen_s, neon_fp_neg_s_q, neon_fp_neg_d_q,\
+		     neon_fp_reduc_add_s_q, neon_fp_reduc_add_d_q,\
+		     neon_fp_reduc_minmax_s_q, neon_fp_reduc_minmax_d_q,\
+		     neon_fp_cvt_narrow_s_q, neon_fp_cvt_narrow_d_q,\
+		     neon_fp_to_int_s_q, neon_fp_to_int_d_q,\
+		     neon_int_to_fp_s_q, neon_int_to_fp_d_q,\
+		     neon_fp_round_s_q,\
+		     neon_fp_recpe_s_q, neon_fp_recpe_d_q,\
+		     neon_fp_recps_s_q, neon_fp_recps_d_q,\
+		     neon_fp_recpx_s_q, neon_fp_recpx_d_q,\
+		     neon_fp_rsqrte_s_q, neon_fp_rsqrte_d_q,\
+		     neon_fp_rsqrts_s_q, neon_fp_rsqrts_d_q")
+      (const_string "advsimd_alu_q")
+    (eq_attr "type" "neon_mul_b, neon_mul_h, neon_mul_s,\
+		     neon_mul_h_scalar, neon_mul_s_scalar,\
+		     neon_sat_mul_b, neon_sat_mul_h, neon_sat_mul_s,\
+		     neon_sat_mul_h_scalar, neon_sat_mul_s_scalar,\
+		     neon_mla_b, neon_mla_h, neon_mla_s,\
+		     neon_mla_h_scalar, neon_mla_s_scalar,\
+		     neon_fp_mul_s, neon_fp_mul_s_scalar,\
+		     neon_fp_mul_d, neon_fp_mla_s,\
+		     neon_fp_mla_s_scalar, neon_fp_mla_d")
+      (const_string "advsimd_mul")
+    (eq_attr "type" "neon_mul_b_q, neon_mul_h_q, neon_mul_s_q,\
+		     neon_mul_b_long, neon_mul_h_long, neon_mul_s_long,\
+		     neon_mul_d_long, neon_mul_h_scalar_q,\
+		     neon_mul_s_scalar_q, neon_mul_h_scalar_long,\
+		     neon_mul_s_scalar_long, neon_sat_mul_b_q,\
+		     neon_sat_mul_h_q, neon_sat_mul_s_q,\
+		     neon_sat_mul_b_long, neon_sat_mul_h_long,\
+		     neon_sat_mul_s_long, neon_sat_mul_h_scalar_q,\
+		     neon_sat_mul_s_scalar_q, neon_sat_mul_h_scalar_long,\
+		     neon_sat_mul_s_scalar_long, neon_mla_b_q,\
+		     neon_mla_h_q, neon_mla_s_q, neon_mla_b_long,\
+		     neon_mla_h_long, neon_mla_s_long,\
+		     neon_mla_h_scalar_q, neon_mla_s_scalar_q,\
+		     neon_mla_h_scalar_long, neon_mla_s_scalar_long,\
+		     neon_sat_mla_b_long, neon_sat_mla_h_long,\
+		     neon_sat_mla_s_long, neon_sat_mla_h_scalar_long,\
+		     neon_sat_mla_s_scalar_long,\
+		     neon_fp_mul_s_q, neon_fp_mul_s_scalar_q,\
+		     neon_fp_mul_d_q, neon_fp_mul_d_scalar_q,\
+		     neon_fp_mla_s_q, neon_fp_mla_s_scalar_q,\
+		     neon_fp_mla_d_q, neon_fp_mla_d_scalar_q")
+      (const_string "advsimd_mul_q")
+    (eq_attr "type" "neon_fp_sqrt_s, neon_fp_div_s")
+      (const_string "advsimd_div_s")
+    (eq_attr "type" "neon_fp_sqrt_s_q, neon_fp_div_s_q")
+      (const_string "advsimd_div_s_q")
+    (eq_attr "type" "neon_fp_sqrt_d, neon_fp_div_d")
+      (const_string "advsimd_div_d")
+    (eq_attr "type" "neon_fp_sqrt_d_q, neon_fp_div_d_q")
+      (const_string "advsimd_div_d_q")
+    (eq_attr "type" "neon_ldr, neon_load1_1reg,\
+		     neon_load1_all_lanes, neon_load1_all_lanes_q,\
+		     neon_load1_one_lane, neon_load1_one_lane_q")
+      (const_string "advsimd_load_64")
+    (eq_attr "type" "neon_str, neon_store1_1reg,\
+		     neon_store1_one_lane,neon_store1_one_lane_q")
+      (const_string "advsimd_store_64")
+    (eq_attr "type" "neon_load1_1reg_q, neon_load1_2reg,\
+		     neon_load2_2reg,\
+		     neon_load2_all_lanes, neon_load2_all_lanes_q,\
+		     neon_load2_one_lane, neon_load2_one_lane_q")
+      (const_string "advsimd_load_128")
+    (eq_attr "type" "neon_store1_1reg_q, neon_store1_2reg,\
+		     neon_store2_2reg,\
+		     neon_store2_one_lane, neon_store2_one_lane_q")
+      (const_string "advsimd_store_128")
+    (eq_attr "type" "neon_load1_2reg_q, neon_load1_3reg, neon_load1_3reg_q,\
+		     neon_load1_4reg, neon_load1_4reg_q, \
+		     neon_load2_2reg_q, neon_load2_4reg,\
+		     neon_load2_4reg_q, neon_load3_3reg,\
+		     neon_load3_3reg_q, neon_load3_all_lanes,\
+		     neon_load3_all_lanes_q, neon_load3_one_lane,\
+		     neon_load3_one_lane_q, neon_load4_4reg,\
+		     neon_load4_4reg_q, neon_load4_all_lanes,\
+		     neon_load4_all_lanes_q, neon_load4_one_lane,\
+		     neon_load4_one_lane_q, neon_ldp, neon_ldp_q")
+      (const_string "advsimd_load_lots")
+    (eq_attr "type" "neon_store1_2reg_q, neon_store1_3reg,\
+		     neon_store1_3reg_q, neon_store1_4reg,\
+		     neon_store1_4reg_q, neon_store2_2reg_q,\
+		     neon_store2_4reg, neon_store2_4reg_q,\
+		     neon_store3_3reg, neon_store3_3reg_q,\
+		     neon_store3_one_lane, neon_store3_one_lane_q,\
+		     neon_store4_4reg, neon_store4_4reg_q,\
+		     neon_store4_one_lane, neon_store4_one_lane_q,\
+		     neon_stp, neon_stp_q")
+      (const_string "advsimd_store_lots")]
+      (const_string "unknown")))
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Floating-point/Advanced SIMD functional units.
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+;; We model the Advanced SIMD unit as two 64-bit units, each with three
+;; pipes, FP_ALU, FP_MUL, FP_DIV.  We also give convenient reservations
+;; for 128-bit Advanced SIMD instructions, which use both units.
+
+;; The floating-point/Advanced SIMD ALU pipelines.
+
+(define_cpu_unit "cortex_a53_fp_alu_lo,\
+		  cortex_a53_fp_alu_hi"
+		 "cortex_a53_advsimd")
+
+(define_reservation "cortex_a53_fp_alu"
+		    "cortex_a53_fp_alu_lo\
+		     |cortex_a53_fp_alu_hi")
+
+(define_reservation "cortex_a53_fp_alu_q"
+		    "cortex_a53_fp_alu_lo\
+		     +cortex_a53_fp_alu_hi")
+
+;; The floating-point/Advanced SIMD multiply/multiply-accumulate
+;; pipelines.
+
+(define_cpu_unit "cortex_a53_fp_mul_lo,\
+		  cortex_a53_fp_mul_hi"
+		 "cortex_a53_advsimd")
+
+(define_reservation "cortex_a53_fp_mul"
+		    "cortex_a53_fp_mul_lo\
+		     |cortex_a53_fp_mul_hi")
+
+(define_reservation "cortex_a53_fp_mul_q"
+		    "cortex_a53_fp_mul_lo\
+		     +cortex_a53_fp_mul_hi")
+
+;; Floating-point/Advanced SIMD divide/square root.
+
+(define_cpu_unit "cortex_a53_fp_div_lo,\
+		  cortex_a53_fp_div_hi"
+		 "cortex_a53_advsimd")
+
+;; Once we choose a pipe, stick with it for three simulated cycles.
+
+(define_reservation "cortex_a53_fp_div"
+		    "(cortex_a53_fp_div_lo*3)\
+		     |(cortex_a53_fp_div_hi*3)")
+
+(define_reservation "cortex_a53_fp_div_q"
+		    "(cortex_a53_fp_div_lo*3)\
+		     +(cortex_a53_fp_div_hi*3)")
+
+;; Cryptographic extensions
+
+(define_cpu_unit "cortex_a53_crypto"
+		 "cortex_a53_advsimd")
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; Floating-point arithmetic.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn_reservation "cortex_a53_fpalu" 4
+(define_insn_reservation "cortex_a53_fpalu" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "ffariths, fadds, ffarithd, faddd, fmov, fmuls,\
-                        f_cvt,f_cvtf2i,f_cvti2f,\
-                        fcmps, fcmpd, fcsel, f_rints, f_rintd, f_minmaxs,\
-                        f_minmaxd"))
-  "cortex_a53_slot0+cortex_a53_fpadd_pipe")
+	(eq_attr "type" "ffariths, fadds, ffarithd, faddd, fmov,
+			f_cvt, fcmps, fcmpd, fcsel, f_rints, f_rintd,
+			f_minmaxs, f_minmaxd"))
+  "cortex_a53_slot_any,cortex_a53_fp_alu")
 
-(define_insn_reservation "cortex_a53_fconst" 2
+(define_insn_reservation "cortex_a53_fconst" 3
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "fconsts,fconstd"))
-  "cortex_a53_slot0+cortex_a53_fpadd_pipe")
+  "cortex_a53_slot_any,cortex_a53_fp_alu")
 
-(define_insn_reservation "cortex_a53_fpmul" 4
+(define_insn_reservation "cortex_a53_fpmul" 5
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "fmuls,fmuld"))
-  "cortex_a53_slot0")
+  "cortex_a53_slot_any,cortex_a53_fp_mul")
 
-;; For single-precision multiply-accumulate, the add (accumulate) is issued after
-;; the multiply completes. Model that accordingly.
+;; For multiply-accumulate, model the add (accumulate) as being issued
+;; after the multiply completes.
 
 (define_insn_reservation "cortex_a53_fpmac" 8
   (and (eq_attr "tune" "cortexa53")
        (eq_attr "type" "fmacs,fmacd,ffmas,ffmad"))
-  "cortex_a53_slot0, nothing*3, cortex_a53_fpadd_pipe")
+  "cortex_a53_slot_any,cortex_a53_fp_mul,
+   nothing*3, cortex_a53_fp_alu")
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; Floating-point divide/square root instructions.
+;; Floating-point to/from core transfers.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; fsqrt really takes one cycle less, but that is not modelled.
 
-(define_insn_reservation "cortex_a53_fdivs" 14
+(define_insn_reservation "cortex_a53_r2f" 6
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "fdivs, fsqrts"))
-  "cortex_a53_slot0, cortex_a53_fp_div_sqrt * 5")
+       (eq_attr "type" "f_mcr,f_mcrr,f_cvti2f,
+			neon_from_gp, neon_from_gp_q"))
+  "cortex_a53_slot_any,cortex_a53_store,
+   nothing,cortex_a53_fp_alu")
 
-(define_insn_reservation "cortex_a53_fdivd" 29
+(define_insn_reservation "cortex_a53_f2r" 6
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "fdivd, fsqrtd"))
-  "cortex_a53_slot0, cortex_a53_fp_div_sqrt * 8")
+       (eq_attr "type" "f_mrc,f_mrrc,f_cvtf2i,
+			neon_to_gp, neon_to_gp_q"))
+  "cortex_a53_slot_any,cortex_a53_fp_alu,
+   nothing,cortex_a53_store")
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; ARMv8-A Cryptographic extensions.
+;; Floating-point flag transfer.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn_reservation "cortex_a53_crypto_aese" 2
+(define_insn_reservation "cortex_a53_f_flags" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "crypto_aese"))
-  "cortex_a53_simd0")
+       (eq_attr "type" "f_flag"))
+  "cortex_a53_slot_any")
 
-(define_insn_reservation "cortex_a53_crypto_aesmc" 2
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Floating-point load/store.
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(define_insn_reservation "cortex_a53_f_load_64" 4
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "crypto_aesmc"))
-  "cortex_a53_simd0 | cortex_a53_simd1")
+       (ior (eq_attr "type" "f_loads,f_loadd")
+	    (eq_attr "cortex_a53_advsimd_type"
+		     "advsimd_load_64")))
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_load")
 
-(define_insn_reservation "cortex_a53_crypto_sha1_fast" 2
+(define_insn_reservation "cortex_a53_f_load_many" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "crypto_sha1_fast, crypto_sha256_fast"))
-  "cortex_a53_simd0")
+       (eq_attr "cortex_a53_advsimd_type"
+		"advsimd_load_128,advsimd_load_lots"))
+  "cortex_a53_single_issue+cortex_a53_ls_agen,
+   cortex_a53_load+cortex_a53_slot0,
+   cortex_a53_load")
 
-(define_insn_reservation "cortex_a53_crypto_sha1_xor" 3
+(define_insn_reservation "cortex_a53_f_store_64" 0
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "crypto_sha1_xor"))
-  "cortex_a53_simd0")
+       (ior (eq_attr "type" "f_stores,f_stored")
+	    (eq_attr "cortex_a53_advsimd_type"
+		     "advsimd_store_64")))
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_store")
 
-(define_insn_reservation "cortex_a53_crypto_sha_slow" 5
+(define_insn_reservation "cortex_a53_f_store_many" 0
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "crypto_sha1_slow, crypto_sha256_slow"))
-  "cortex_a53_simd0")
+       (eq_attr "cortex_a53_advsimd_type"
+		"advsimd_store_128,advsimd_store_lots"))
+  "cortex_a53_slot_any+cortex_a53_ls_agen,
+   cortex_a53_store+cortex_a53_slot0,
+   cortex_a53_store")
 
-(define_bypass 0 "cortex_a53_crypto_aese"
-                 "cortex_a53_crypto_aesmc"
-                 "aarch_crypto_can_dual_issue")
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; VFP to/from core transfers.
+;; Advanced SIMD.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn_reservation "cortex_a53_r2f" 4
+;; Either we want to model use of the ALU pipe, the multiply pipe or the
+;; divide/sqrt pipe.  In all cases we need to check if we are a 64-bit
+;; operation (in which case we model dual-issue without penalty)
+;; or a 128-bit operation in which case we require in our model that we
+;; issue from slot 0.
+
+(define_insn_reservation "cortex_a53_advsimd_alu" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_mcr,f_mcrr"))
-  "cortex_a53_slot0")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_alu"))
+  "cortex_a53_slot_any,cortex_a53_fp_alu")
 
-(define_insn_reservation "cortex_a53_f2r" 2
+(define_insn_reservation "cortex_a53_advsimd_alu_q" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_mrc,f_mrrc"))
-  "cortex_a53_slot0")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_alu_q"))
+  "cortex_a53_slot0,cortex_a53_fp_alu_q")
 
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; VFP flag transfer.
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(define_insn_reservation "cortex_a53_advsimd_mul" 5
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_mul"))
+  "cortex_a53_slot_any,cortex_a53_fp_mul")
 
-(define_insn_reservation "cortex_a53_f_flags" 4
+(define_insn_reservation "cortex_a53_advsimd_mul_q" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_flag"))
-  "cortex_a53_slot0")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_mul_q"))
+  "cortex_a53_slot0,cortex_a53_fp_mul_q")
 
+;; SIMD Dividers.
+
+(define_insn_reservation "cortex_a53_advsimd_div_s" 14
+  (and (eq_attr "tune" "cortexa53")
+       (ior (eq_attr "type" "fdivs,fsqrts")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_div_s")))
+  "cortex_a53_slot0,cortex_a53_fp_mul,
+   cortex_a53_fp_div")
+
+(define_insn_reservation "cortex_a53_advsimd_div_d" 29
+  (and (eq_attr "tune" "cortexa53")
+       (ior (eq_attr "type" "fdivd,fsqrtd")
+	    (eq_attr "cortex_a53_advsimd_type" "advsimd_div_d")))
+  "cortex_a53_slot0,cortex_a53_fp_mul,
+   cortex_a53_fp_div")
+
+(define_insn_reservation "cortex_a53_advsimd_div_s_q" 14
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_div_s_q"))
+  "cortex_a53_single_issue,cortex_a53_fp_mul_q,
+   cortex_a53_fp_div_q")
+
+(define_insn_reservation "cortex_a53_advsimd_divd_q" 29
+  (and (eq_attr "tune" "cortexa53")
+       (eq_attr "cortex_a53_advsimd_type" "advsimd_div_d_q"))
+  "cortex_a53_single_issue,cortex_a53_fp_mul_q,
+   cortex_a53_fp_div_q")
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; VFP load/store.
+;; ARMv8-A Cryptographic extensions.
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn_reservation "cortex_a53_f_loads" 4
+;; We want AESE and AESMC to end up consecutive to one another.
+
+(define_insn_reservation "cortex_a53_crypto_aese" 3
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_loads"))
+       (eq_attr "type" "crypto_aese"))
   "cortex_a53_slot0")
 
-(define_insn_reservation "cortex_a53_f_loadd" 5
+(define_insn_reservation "cortex_a53_crypto_aesmc" 3
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_loadd"))
-  "cortex_a53_slot0")
+       (eq_attr "type" "crypto_aesmc"))
+  "cortex_a53_slot_any")
 
-(define_insn_reservation "cortex_a53_f_load_2reg" 5
+;; SHA1H
+
+(define_insn_reservation "cortex_a53_crypto_sha1_fast" 3
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "neon_ldp, neon_ldp_q, neon_load2_2reg_q"))
-  "(cortex_a53_slot_any+cortex_a53_ls)*2")
+       (eq_attr "type" "crypto_sha1_fast"))
+  "cortex_a53_slot_any,cortex_a53_crypto")
 
-(define_insn_reservation "cortex_a53_f_loadq" 5
+(define_insn_reservation "cortex_a53_crypto_sha256_fast" 3
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "neon_load1_1reg_q"))
-  "cortex_a53_slot_any+cortex_a53_ls")
+       (eq_attr "type" "crypto_sha256_fast"))
+  "cortex_a53_slot0,cortex_a53_crypto")
 
-(define_insn_reservation "cortex_a53_f_stores" 0
+(define_insn_reservation "cortex_a53_crypto_sha1_xor" 4
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_stores"))
-  "cortex_a53_slot0")
+       (eq_attr "type" "crypto_sha1_xor"))
+  "cortex_a53_slot0,cortex_a53_crypto")
 
-(define_insn_reservation "cortex_a53_f_stored" 0
+(define_insn_reservation "cortex_a53_crypto_sha_slow" 5
   (and (eq_attr "tune" "cortexa53")
-       (eq_attr "type" "f_stored"))
-  "cortex_a53_slot0")
+       (eq_attr "type" "crypto_sha1_slow, crypto_sha256_slow"))
+  "cortex_a53_slot0,cortex_a53_crypto")
 
-;; Load-to-use for floating-point values has a penalty of one cycle,
-;; i.e. a latency of two.
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Floating-point/Advanced SIMD register bypasses.
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_bypass 2 "cortex_a53_f_loads"
-                 "cortex_a53_fpalu, cortex_a53_fpmac, cortex_a53_fpmul,\
-		  cortex_a53_fdivs, cortex_a53_fdivd,\
-		  cortex_a53_f2r")
+;; Model the late use of the accumulator operand for floating-point
+;; multiply-accumulate operations as a bypass reducing the latency
+;; of producing instructions to near zero.
 
-(define_bypass 2 "cortex_a53_f_loadd"
-                 "cortex_a53_fpalu, cortex_a53_fpmac, cortex_a53_fpmul,\
-		  cortex_a53_fdivs, cortex_a53_fdivd,\
-		  cortex_a53_f2r")
+(define_bypass 1 "cortex_a53_fp*,
+		  cortex_a53_r2f,
+		  cortex_a53_f_load*"
+		 "cortex_a53_fpmac"
+		 "aarch_accumulator_forwarding")
 
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; Crude Advanced SIMD approximation.
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Model a bypass from the result of an FP operation to a use.
 
-(define_insn_reservation "cortex_a53_advsimd" 4
-  (and (eq_attr "tune" "cortexa53")
-       (eq_attr "is_neon_type" "yes"))
-  "cortex_a53_simd0")
+(define_bypass 4 "cortex_a53_fpalu,
+		  cortex_a53_fpmul"
+		 "cortex_a53_fpalu,
+		  cortex_a53_fpmul,
+		  cortex_a53_fpmac,
+		  cortex_a53_advsimd_div*")
+
+;; We want AESE and AESMC to end up consecutive to one another.
+
+(define_bypass 0 "cortex_a53_crypto_aese"
+		 "cortex_a53_crypto_aesmc"
+		 "aarch_crypto_can_dual_issue")
+