diff mbox

[RFC,ARM] Add pipeline description for Cortex-M4 -- try 2

Message ID 4C8F3FF1.5020300@codesourcery.com
State New
Headers show

Commit Message

Jie Zhang Sept. 14, 2010, 9:27 a.m. UTC 
Since there is no Cortex-M4F processor, I updated my patch accordingly.

Comments are welcome!

Comments

Ramana Radhakrishnan Sept. 14, 2010, 1:17 p.m. UTC | #1 
On Tue, Sep 14, 2010 at 2:57 PM, Jie Zhang <jie@codesourcery.com> wrote:
> Since there is no Cortex-M4F processor, I updated my patch accordingly.
>
> Comments are welcome!
>

I've had a quick read through and an obvious query I had was whether
there was any difference in cycle counts on the M4 for 16 and 32 bit
multiplications / multiply and accumulate operations ?

cheers
Ramana
> --
> Jie Zhang
> CodeSourcery
>
Jie Zhang Sept. 14, 2010, 1:28 p.m. UTC | #2 
On 09/14/2010 09:17 PM, Ramana Radhakrishnan wrote:
> On Tue, Sep 14, 2010 at 2:57 PM, Jie Zhang<jie@codesourcery.com>  wrote:
>> Since there is no Cortex-M4F processor, I updated my patch accordingly.
>>
>> Comments are welcome!
>>
>
> I've had a quick read through and an obvious query I had was whether
> there was any difference in cycle counts on the M4 for 16 and 32 bit
> multiplications / multiply and accumulate operations ?
>
According to the latest Cortex-M4 TRM, no. All multiply and MAC are 1 cycle.
Paul Brook Sept. 20, 2010, 8:59 a.m. UTC | #3 
> Since there is no Cortex-M4F processor, I updated my patch accordingly.

> +;; ALU and multiply is one cycle.
> +(define_insn_reservation "cortex_m4_alu" 1

I think it's worth a note here that the default latency assumes no stall and 
the "bypass" increases the latency.

In my experience it's normal for the default latency to be the maximum 
latency.  I don't think this is a problem, just worthy of a comment.

Other than that, ok.

Paul
Jie Zhang Sept. 20, 2010, 9:16 a.m. UTC | #4 
On 09/20/2010 04:59 PM, Paul Brook wrote:
>> Since there is no Cortex-M4F processor, I updated my patch accordingly.
>
>> +;; ALU and multiply is one cycle.
>> +(define_insn_reservation "cortex_m4_alu" 1
>
> I think it's worth a note here that the default latency assumes no stall and
> the "bypass" increases the latency.
>
> In my experience it's normal for the default latency to be the maximum
> latency.  I don't think this is a problem, just worthy of a comment.
>
> Other than that, ok.
>
Thanks for review. The last example in

http://gcc.gnu.org/onlinedocs/gccint/Processor-pipeline-description.html#Processor-pipeline-description

      (define_insn_reservation "float" 3 (eq_attr "type" "float")
                               "f_pipeline, nothing, (port0 | port1))

      (define_bypass 4 "float" "simple,mult,div")

shows the former way might also be normal.
Paul Brook Sept. 20, 2010, 9:23 a.m. UTC | #5 
> On 09/20/2010 04:59 PM, Paul Brook wrote:
> >> Since there is no Cortex-M4F processor, I updated my patch accordingly.
> >> 
> >> +;; ALU and multiply is one cycle.
> >> +(define_insn_reservation "cortex_m4_alu" 1
> > 
> > I think it's worth a note here that the default latency assumes no stall
> > and the "bypass" increases the latency.
> > 
> > In my experience it's normal for the default latency to be the maximum
> > latency.  I don't think this is a problem, just worthy of a comment.
> > 
> > Other than that, ok.
> 
> Thanks for review. The last example in
> 
> http://gcc.gnu.org/onlinedocs/gccint/Processor-pipeline-description.html#Pr
> ocessor-pipeline-description
> 
>       (define_insn_reservation "float" 3 (eq_attr "type" "float")
>                                "f_pipeline, nothing, (port0 | port1))
> 
>       (define_bypass 4 "float" "simple,mult,div")
> 
> shows the former way might also be normal.

Maybe it's just me then.  Patch OK.

Paul
Jie Zhang Sept. 20, 2010, 4:01 p.m. UTC | #6 
On 09/20/2010 05:23 PM, Paul Brook wrote:
>> On 09/20/2010 04:59 PM, Paul Brook wrote:
>>>> Since there is no Cortex-M4F processor, I updated my patch accordingly.
>>>>
>>>> +;; ALU and multiply is one cycle.
>>>> +(define_insn_reservation "cortex_m4_alu" 1
>>>
>>> I think it's worth a note here that the default latency assumes no stall
>>> and the "bypass" increases the latency.
>>>
>>> In my experience it's normal for the default latency to be the maximum
>>> latency.  I don't think this is a problem, just worthy of a comment.
>>>
>>> Other than that, ok.
>>
>> Thanks for review. The last example in
>>
>> http://gcc.gnu.org/onlinedocs/gccint/Processor-pipeline-description.html#Pr
>> ocessor-pipeline-description
>>
>>        (define_insn_reservation "float" 3 (eq_attr "type" "float")
>>                                 "f_pipeline, nothing, (port0 | port1))
>>
>>        (define_bypass 4 "float" "simple,mult,div")
>>
>> shows the former way might also be normal.
>
> Maybe it's just me then.  Patch OK.
>
Committed on trunk. Thanks.
diff mbox

Patch


	* config/arm/arm.c (arm_address_offset_is_imm): New.
	(arm_early_store_addr_dep): New.
	(arm_early_load_addr_dep): New.
	* config/arm/arm-protos.h (arm_early_store_addr_dep): Declare.
	(arm_early_load_addr_dep): Declare.
	(arm_address_offset_is_imm): Declare.
	* config/arm/cortex-m4.md: New file.
	* config/arm/cortex-m4-fpu.md: New file.
	* config/arm/arm.md: Include cortex-m4.md and cortex-m4-fpu.md.
	(attr generic_sched): Exclude cortexm4.
	(attr generic_vfp): Exclude cortexm4.


Index: config/arm/arm.c
===================================================================
--- config/arm/arm.c	(revision 164242)
+++ config/arm/arm.c	(working copy)
@@ -13333,6 +13333,34 @@  arm_attr_length_move_neon (rtx insn)
     return 4;
 }
 
+/* Return nonzero if the offset in the address is an immediate.  Otherwise,
+   return zero.  */
+
+int
+arm_address_offset_is_imm (rtx insn)
+{
+  rtx mem, addr;
+
+  extract_insn_cached (insn);
+
+  if (REG_P (recog_data.operand[0]))
+    return 0;
+
+  mem = recog_data.operand[0];
+
+  gcc_assert (MEM_P (mem));
+
+  addr = XEXP (mem, 0);
+
+  if (GET_CODE (addr) == REG
+      || (GET_CODE (addr) == PLUS
+	  && GET_CODE (XEXP (addr, 0)) == REG
+	  && GET_CODE (XEXP (addr, 1)) == CONST_INT))
+    return 1;
+  else
+    return 0;
+}
+
 /* Output an ADD r, s, #n where n may be too big for one instruction.
    If adding zero to one register, output nothing.  */
 const char *
@@ -21491,6 +21519,38 @@  arm_no_early_store_addr_dep (rtx produce
   return !reg_overlap_mentioned_p (value, addr);
 }
 
+/* Return nonzero if the CONSUMER instruction (a store) does need
+   PRODUCER's value to calculate the address.  */
+
+int
+arm_early_store_addr_dep (rtx producer, rtx consumer)
+{
+  return !arm_no_early_store_addr_dep (producer, consumer);
+}
+
+/* Return nonzero if the CONSUMER instruction (a load) does need
+   PRODUCER's value to calculate the address.  */
+
+int
+arm_early_load_addr_dep (rtx producer, rtx consumer)
+{
+  rtx value = PATTERN (producer);
+  rtx addr = PATTERN (consumer);
+
+  if (GET_CODE (value) == COND_EXEC)
+    value = COND_EXEC_CODE (value);
+  if (GET_CODE (value) == PARALLEL)
+    value = XVECEXP (value, 0, 0);
+  value = XEXP (value, 0);
+  if (GET_CODE (addr) == COND_EXEC)
+    addr = COND_EXEC_CODE (addr);
+  if (GET_CODE (addr) == PARALLEL)
+    addr = XVECEXP (addr, 0, 0);
+  addr = XEXP (addr, 1);
+
+  return reg_overlap_mentioned_p (value, addr);
+}
+
 /* Return nonzero if the CONSUMER instruction (an ALU op) does not
    have an early register shift value or amount dependency on the
    result of PRODUCER.  */
Index: config/arm/cortex-m4.md
===================================================================
--- config/arm/cortex-m4.md	(revision 0)
+++ config/arm/cortex-m4.md	(revision 0)
@@ -0,0 +1,111 @@ 
+;; ARM Cortex-M4 pipeline description
+;; Copyright (C) 2010 Free Software Foundation, Inc.
+;; Contributed by CodeSourcery.
+;;
+;; This file is part of GCC.
+;;
+;; GCC is free software; you can redistribute it and/or modify it
+;; under the terms of the GNU General Public License as published by
+;; the Free Software Foundation; either version 3, or (at your option)
+;; any later version.
+;;
+;; GCC is distributed in the hope that it will be useful, but
+;; WITHOUT ANY WARRANTY; without even the implied warranty of
+;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;; General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License
+;; along with GCC; see the file COPYING3.  If not see
+;; <http://www.gnu.org/licenses/>.
+
+(define_automaton "cortex_m4")
+
+;; We model the pipelining of LDR instructions by using two artificial units.
+
+(define_cpu_unit "cortex_m4_a" "cortex_m4")
+
+(define_cpu_unit "cortex_m4_b" "cortex_m4")
+
+(define_reservation "cortex_m4_ex" "cortex_m4_a+cortex_m4_b")
+
+;; ALU and multiply is one cycle.
+(define_insn_reservation "cortex_m4_alu" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "alu,alu_shift,alu_shift_reg,mult"))
+  "cortex_m4_ex")
+
+;; Byte, half-word and word load is two cycles.
+(define_insn_reservation "cortex_m4_load1" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "load_byte,load1"))
+  "cortex_m4_a, cortex_m4_b")
+
+;; str rx, [ry, #imm] is always one cycle.
+(define_insn_reservation "cortex_m4_store1_1" 1
+  (and (and (eq_attr "tune" "cortexm4")
+	    (eq_attr "type" "store1"))
+       (ne (symbol_ref ("arm_address_offset_is_imm (insn)")) (const_int 0)))
+  "cortex_m4_a")
+
+;; Other byte, half-word and word load is two cycles.
+(define_insn_reservation "cortex_m4_store1_2" 2
+  (and (and (eq_attr "tune" "cortexm4")
+	    (eq_attr "type" "store1"))
+       (eq (symbol_ref ("arm_address_offset_is_imm (insn)")) (const_int 0)))
+  "cortex_m4_a*2")
+
+(define_insn_reservation "cortex_m4_load2" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "load2"))
+  "cortex_m4_ex*3")
+
+(define_insn_reservation "cortex_m4_store2" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "store2"))
+  "cortex_m4_ex*3")
+
+(define_insn_reservation "cortex_m4_load3" 4
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "load3"))
+  "cortex_m4_ex*4")
+
+(define_insn_reservation "cortex_m4_store3" 4
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "store3"))
+  "cortex_m4_ex*4")
+
+(define_insn_reservation "cortex_m4_load4" 5
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "load4"))
+  "cortex_m4_ex*5")
+
+(define_insn_reservation "cortex_m4_store4" 5
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "store4"))
+  "cortex_m4_ex*5")
+
+;; If the address of load or store depends on the result of the preceding
+;; instruction, the latency is increased by one.
+
+(define_bypass 2 "cortex_m4_alu"
+		 "cortex_m4_load1"
+		 "arm_early_load_addr_dep")
+
+(define_bypass 2 "cortex_m4_alu"
+		 "cortex_m4_store1_1,cortex_m4_store1_2"
+		 "arm_early_store_addr_dep")
+
+(define_insn_reservation "cortex_m4_branch" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "branch"))
+  "cortex_m4_ex*3")
+
+(define_insn_reservation "cortex_m4_call" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "call"))
+  "cortex_m4_ex*3")
+
+(define_insn_reservation "cortex_m4_block" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "block"))
+  "cortex_m4_ex")
Index: config/arm/arm-protos.h
===================================================================
--- config/arm/arm-protos.h	(revision 164242)
+++ config/arm/arm-protos.h	(working copy)
@@ -88,6 +88,8 @@  extern int arm_coproc_mem_operand (rtx,
 extern int neon_vector_mem_operand (rtx, int);
 extern int neon_struct_mem_operand (rtx);
 extern int arm_no_early_store_addr_dep (rtx, rtx);
+extern int arm_early_store_addr_dep (rtx, rtx);
+extern int arm_early_load_addr_dep (rtx, rtx);
 extern int arm_no_early_alu_shift_dep (rtx, rtx);
 extern int arm_no_early_alu_shift_value_dep (rtx, rtx);
 extern int arm_no_early_mul_dep (rtx, rtx);
@@ -129,6 +131,7 @@  extern const char *output_move_quad (rtx
 extern const char *output_move_vfp (rtx *operands);
 extern const char *output_move_neon (rtx *operands);
 extern int arm_attr_length_move_neon (rtx);
+extern int arm_address_offset_is_imm (rtx);
 extern const char *output_add_immediate (rtx *);
 extern const char *arithmetic_instr (rtx, int);
 extern void output_ascii_pseudo_op (FILE *, const unsigned char *, int);
Index: config/arm/arm.md
===================================================================
--- config/arm/arm.md	(revision 164242)
+++ config/arm/arm.md	(working copy)
@@ -496,16 +496,16 @@ 
 ;; True if the generic scheduling description should be used.
 
 (define_attr "generic_sched" "yes,no"
-  (const (if_then_else 
-          (ior (eq_attr "tune" "arm926ejs,arm1020e,arm1026ejs,arm1136js,arm1136jfs,cortexa5,cortexa8,cortexa9")
-	      (eq_attr "tune_cortexr4" "yes"))
+  (const (if_then_else
+          (ior (eq_attr "tune" "arm926ejs,arm1020e,arm1026ejs,arm1136js,arm1136jfs,cortexa5,cortexa8,cortexa9,cortexm4")
+	       (eq_attr "tune_cortexr4" "yes"))
           (const_string "no")
           (const_string "yes"))))
 
 (define_attr "generic_vfp" "yes,no"
   (const (if_then_else
 	  (and (eq_attr "fpu" "vfp")
-	       (eq_attr "tune" "!arm1020e,arm1022e,cortexa5,cortexa8,cortexa9")
+	       (eq_attr "tune" "!arm1020e,arm1022e,cortexa5,cortexa8,cortexa9,cortexm4")
 	       (eq_attr "tune_cortexr4" "no"))
 	  (const_string "yes")
 	  (const_string "no"))))
@@ -520,6 +520,8 @@ 
 (include "cortex-a9.md")
 (include "cortex-r4.md")
 (include "cortex-r4f.md")
+(include "cortex-m4.md")
+(include "cortex-m4-fpu.md")
 (include "vfp11.md")
 
 
Index: config/arm/cortex-m4-fpu.md
===================================================================
--- config/arm/cortex-m4-fpu.md	(revision 0)
+++ config/arm/cortex-m4-fpu.md	(revision 0)
@@ -0,0 +1,111 @@ 
+;; ARM Cortex-M4 FPU pipeline description
+;; Copyright (C) 2010 Free Software Foundation, Inc.
+;; Contributed by CodeSourcery.
+;;
+;; This file is part of GCC.
+;;
+;; GCC is free software; you can redistribute it and/or modify it
+;; under the terms of the GNU General Public License as published by
+;; the Free Software Foundation; either version 3, or (at your option)
+;; any later version.
+;;
+;; GCC is distributed in the hope that it will be useful, but
+;; WITHOUT ANY WARRANTY; without even the implied warranty of
+;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+;; General Public License for more details.
+;;
+;; You should have received a copy of the GNU General Public License
+;; along with GCC; see the file COPYING3.  If not see
+;; <http://www.gnu.org/licenses/>.
+
+;; Use an artifial unit to model FPU.
+(define_cpu_unit "cortex_m4_v" "cortex_m4")
+
+(define_reservation "cortex_m4_ex_v" "cortex_m4_ex+cortex_m4_v")
+
+;; Integer instructions following VDIV or VSQRT complete out-of-order.
+(define_insn_reservation "cortex_m4_fdivs" 15
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fdivs"))
+  "cortex_m4_ex_v,cortex_m4_v*13")
+
+(define_insn_reservation "cortex_m4_vmov_1" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fcpys,fconsts"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_vmov_2" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_2_r,r_2_f"))
+  "cortex_m4_ex_v*2")
+
+(define_insn_reservation "cortex_m4_fmuls" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fmuls"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_fmacs" 4
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fmacs"))
+  "cortex_m4_ex_v*3")
+
+(define_insn_reservation "cortex_m4_ffariths" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "ffariths"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_fadds" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fadds"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_fcmps" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "fcmps"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_f_flag" 1
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_flag"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_f_cvt" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_cvt"))
+  "cortex_m4_ex_v")
+
+(define_insn_reservation "cortex_m4_f_load" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_load"))
+  "cortex_m4_ex_v*2")
+
+(define_insn_reservation "cortex_m4_f_store" 2
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_store"))
+  "cortex_m4_ex_v*2")
+
+(define_insn_reservation "cortex_m4_f_loadd" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_loadd"))
+  "cortex_m4_ex_v*3")
+
+(define_insn_reservation "cortex_m4_f_stored" 3
+  (and (eq_attr "tune" "cortexm4")
+       (eq_attr "type" "f_stored"))
+  "cortex_m4_ex_v*3")
+
+;; MAC instructions consume their addend one cycle later. If the result
+;; of an arithmetic instruction is consumed as the addend of the following
+;; MAC instruction, the latency can be decreased by one.
+
+(define_bypass 1 "cortex_m4_fadds,cortex_m4_fmuls,cortex_m4_f_cvt"
+		 "cortex_m4_fmacs"
+		 "arm_no_early_mul_dep")
+
+(define_bypass 3 "cortex_m4_fmacs"
+		 "cortex_m4_fmacs"
+		 "arm_no_early_mul_dep")
+
+(define_bypass 14 "cortex_m4_fdivs"
+		  "cortex_m4_fmacs"
+		  "arm_no_early_mul_dep")