* [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
@ 2014-11-12 10:02 Yangfei (Felix)
2014-11-12 10:55 ` Jiong Wang
2014-11-12 16:06 ` Richard Henderson
0 siblings, 2 replies; 14+ messages in thread
From: Yangfei (Felix) @ 2014-11-12 10:02 UTC (permalink / raw)
To: Yangfei (Felix), gcc-patches, Marcus Shawcroft, Richard Earnshaw
Hello,
Any comments on this patch? Thanks.
> Hi,
> This patch adds doloop_end standard pattern for AArch64 port so that
> -fmodulo-sched can do its job.
> Reg-tested for aarch64-linux-gnu with QEMU. OK for the trunk?
>
>
> Index: gcc/ChangeLog
> =============================================================
> ======
> --- gcc/ChangeLog (revision 216838)
> +++ gcc/ChangeLog (working copy)
> @@ -1,3 +1,7 @@
> +2014-11-10 Felix Yang <felix.yang@huawei.com>
> +
> + * config/aarch64/aarch64.md (doloop_end): New pattern.
> +
> 2014-10-29 Richard Sandiford <richard.sandiford@arm.com>
>
> * addresses.h, alias.c, asan.c, auto-inc-dec.c, bt-load.c, builtins.c,
> Index: gcc/config/aarch64/aarch64.md
> =============================================================
> ======
> --- gcc/config/aarch64/aarch64.md (revision 216838)
> +++ gcc/config/aarch64/aarch64.md (working copy)
> @@ -4087,6 +4087,48 @@
> [(set_attr "type" "mrs")])
>
>
> +;; Define the subtract-one-and-jump insns so loop.c ;; knows what to
> +generate.
> +(define_expand "doloop_end"
> + [(use (match_operand 0 "" "")) ; loop pseudo
> + (use (match_operand 1 "" ""))] ; label
> + ""
> + "
> +{
> + /* Currently SMS relies on the do-loop pattern to recognize loops
> + where (1) the control part consists of all insns defining and/or
> + using a certain 'count' register and (2) the loop count can be
> + adjusted by modifying this register prior to the loop.
> + ??? The possible introduction of a new block to initialize the
> + new IV can potentially affect branch optimizations. */
> + if (optimize > 0 && flag_modulo_sched)
> + {
> + rtx s0;
> + rtx bcomp;
> + rtx loc_ref;
> + rtx cc_reg;
> + rtx insn;
> + rtx cmp;
> +
> + if (GET_MODE (operands[0]) != DImode)
> + FAIL;
> +
> + s0 = operands [0];
> + insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
> +
> + cmp = XVECEXP (PATTERN (insn), 0, 0);
> + cc_reg = SET_DEST (cmp);
> + bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
> + loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
> + emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
> + gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
> + loc_ref, pc_rtx)));
> + DONE;
> + }
> + else
> + FAIL;
> +}")
> +
> ;; AdvSIMD Stuff
> (include "aarch64-simd.md")
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-12 10:02 [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched Yangfei (Felix)
@ 2014-11-12 10:55 ` Jiong Wang
2014-11-12 15:59 ` Andrew Pinski
2014-11-12 16:06 ` Richard Henderson
1 sibling, 1 reply; 14+ messages in thread
From: Jiong Wang @ 2014-11-12 10:55 UTC (permalink / raw)
To: Yangfei (Felix), gcc-patches, Marcus Shawcroft, Richard Earnshaw
On 12/11/14 10:01, Yangfei (Felix) wrote:
> + if (GET_MODE (operands[0]) != DImode)
> + FAIL;
shouldn't SImode be considered for -mabi=ilp32 ?
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-12 10:55 ` Jiong Wang
@ 2014-11-12 15:59 ` Andrew Pinski
0 siblings, 0 replies; 14+ messages in thread
From: Andrew Pinski @ 2014-11-12 15:59 UTC (permalink / raw)
To: Jiong Wang
Cc: Yangfei (Felix), gcc-patches, Marcus Shawcroft, Richard Earnshaw
On Wed, Nov 12, 2014 at 2:54 AM, Jiong Wang <jiong.wang@arm.com> wrote:
> On 12/11/14 10:01, Yangfei (Felix) wrote:
>
>> + if (GET_MODE (operands[0]) != DImode)
>> + FAIL;
>
>
> shouldn't SImode be considered for -mabi=ilp32 ?
I don't think it is needed for -mabi=ilp32 since word_mode is still
DImode for -mabi=ilp32.
Thanks,
Andrew
>
>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-12 10:02 [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched Yangfei (Felix)
2014-11-12 10:55 ` Jiong Wang
@ 2014-11-12 16:06 ` Richard Henderson
2014-11-13 10:45 ` [PING ^ 2][PATCH, " Yangfei (Felix)
2014-11-17 8:57 ` [PING ^ 3][PATCH, " Yangfei (Felix)
1 sibling, 2 replies; 14+ messages in thread
From: Richard Henderson @ 2014-11-12 16:06 UTC (permalink / raw)
To: Yangfei (Felix), gcc-patches, Marcus Shawcroft, Richard Earnshaw
On 11/12/2014 11:01 AM, Yangfei (Felix) wrote:
> +(define_expand "doloop_end"
>> + [(use (match_operand 0 "" "")) ; loop pseudo
>> + (use (match_operand 1 "" ""))] ; label
>> + ""
>> + "
>> +{
Drop the "" surrounding the { }.
r~
^ permalink raw reply [flat|nested] 14+ messages in thread
* [PING ^ 2][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-12 16:06 ` Richard Henderson
@ 2014-11-13 10:45 ` Yangfei (Felix)
2014-11-17 8:57 ` [PING ^ 3][PATCH, " Yangfei (Felix)
1 sibling, 0 replies; 14+ messages in thread
From: Yangfei (Felix) @ 2014-11-13 10:45 UTC (permalink / raw)
To: Richard Henderson, gcc-patches, Marcus Shawcroft, Richard Earnshaw
[-- Attachment #1: Type: text/plain, Size: 2480 bytes --]
> On 11/12/2014 11:01 AM, Yangfei (Felix) wrote:
> > +(define_expand "doloop_end"
> >> + [(use (match_operand 0 "" "")) ; loop pseudo
> >> + (use (match_operand 1 "" ""))] ; label
> >> + ""
> >> + "
> >> +{
>
> Drop the "" surrounding the { }.
>
>
> r~
Hello,
I updated the patch with the redundant "" removed. Is it OK for trunk now?
I hope this patch can catch up with stage 1 of GCC-5.0. Thanks.
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,7 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.md (doloop_end): New pattern.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,47 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ ""
+{
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+ if (optimize > 0 && flag_modulo_sched)
+ {
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+ }
+ else
+ FAIL;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
[-- Attachment #2: aarch64-doloop-v2.diff --]
[-- Type: application/octet-stream, Size: 1966 bytes --]
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,7 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.md (doloop_end): New pattern.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,47 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ ""
+{
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+ if (optimize > 0 && flag_modulo_sched)
+ {
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+ }
+ else
+ FAIL;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
^ permalink raw reply [flat|nested] 14+ messages in thread
* [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-12 16:06 ` Richard Henderson
2014-11-13 10:45 ` [PING ^ 2][PATCH, " Yangfei (Felix)
@ 2014-11-17 8:57 ` Yangfei (Felix)
2014-11-18 10:18 ` Marcus Shawcroft
1 sibling, 1 reply; 14+ messages in thread
From: Yangfei (Felix) @ 2014-11-17 8:57 UTC (permalink / raw)
To: gcc-patches, Marcus Shawcroft, marcus.shawcroft, Richard Earnshaw
PING? Is it OK for me to apply this patch? Thanks.
>
> > On 11/12/2014 11:01 AM, Yangfei (Felix) wrote:
> > > +(define_expand "doloop_end"
> > >> + [(use (match_operand 0 "" "")) ; loop pseudo
> > >> + (use (match_operand 1 "" ""))] ; label
> > >> + ""
> > >> + "
> > >> +{
> >
> > Drop the "" surrounding the { }.
> >
> >
> > r~
>
>
> Hello,
> I updated the patch with the redundant "" removed. Is it OK for trunk
> now?
> I hope this patch can catch up with stage 1 of GCC-5.0. Thanks.
>
>
> Index: gcc/ChangeLog
> =============================================================
> ======
> --- gcc/ChangeLog (revision 217394)
> +++ gcc/ChangeLog (working copy)
> @@ -1,3 +1,7 @@
> +2014-11-13 Felix Yang <felix.yang@huawei.com>
> +
> + * config/aarch64/aarch64.md (doloop_end): New pattern.
> +
> 2014-11-11 Andrew Pinski <apinski@cavium.com>
>
> Bug target/61997
> Index: gcc/config/aarch64/aarch64.md
> =============================================================
> ======
> --- gcc/config/aarch64/aarch64.md (revision 217394)
> +++ gcc/config/aarch64/aarch64.md (working copy)
> @@ -4087,6 +4087,47 @@
> [(set_attr "type" "mrs")])
>
>
> +;; Define the subtract-one-and-jump insns so loop.c ;; knows what to
> +generate.
> +(define_expand "doloop_end"
> + [(use (match_operand 0 "" "")) ; loop pseudo
> + (use (match_operand 1 "" ""))] ; label
> + ""
> +{
> + /* Currently SMS relies on the do-loop pattern to recognize loops
> + where (1) the control part consists of all insns defining and/or
> + using a certain 'count' register and (2) the loop count can be
> + adjusted by modifying this register prior to the loop.
> + ??? The possible introduction of a new block to initialize the
> + new IV can potentially affect branch optimizations. */
> + if (optimize > 0 && flag_modulo_sched)
> + {
> + rtx s0;
> + rtx bcomp;
> + rtx loc_ref;
> + rtx cc_reg;
> + rtx insn;
> + rtx cmp;
> +
> + if (GET_MODE (operands[0]) != DImode)
> + FAIL;
> +
> + s0 = operands [0];
> + insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
> +
> + cmp = XVECEXP (PATTERN (insn), 0, 0);
> + cc_reg = SET_DEST (cmp);
> + bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
> + loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
> + emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
> + gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
> + loc_ref, pc_rtx)));
> + DONE;
> + }
> + else
> + FAIL;
> +})
> +
> ;; AdvSIMD Stuff
> (include "aarch64-simd.md")
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-17 8:57 ` [PING ^ 3][PATCH, " Yangfei (Felix)
@ 2014-11-18 10:18 ` Marcus Shawcroft
2014-11-18 10:56 ` Yangfei (Felix)
0 siblings, 1 reply; 14+ messages in thread
From: Marcus Shawcroft @ 2014-11-18 10:18 UTC (permalink / raw)
To: Yangfei (Felix); +Cc: gcc-patches, marcus.shawcroft, Richard Earnshaw
On 17 November 2014 07:59, Yangfei (Felix) <felix.yang@huawei.com> wrote:
>> +2014-11-13 Felix Yang <felix.yang@huawei.com>
>> +
>> + * config/aarch64/aarch64.md (doloop_end): New pattern.
>> +
This looks like a straight copy of the ARM code, but without the
TARGET_CAN_USE_DOLOOP_P definition. If the reason for including this
code is for the benefit of module-sched then should the hook be
defined to limit the use of this pattern to inner most loops only?
Cheers
/Marcus
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 10:18 ` Marcus Shawcroft
@ 2014-11-18 10:56 ` Yangfei (Felix)
2014-11-18 11:29 ` Richard Henderson
0 siblings, 1 reply; 14+ messages in thread
From: Yangfei (Felix) @ 2014-11-18 10:56 UTC (permalink / raw)
To: Marcus Shawcroft; +Cc: gcc-patches, marcus.shawcroft, Richard Earnshaw
[-- Attachment #1: Type: text/plain, Size: 3540 bytes --]
Yes, major code is borrowed from ARM port with the expected mode of loop pseudo changed to DImode.
The function loop_canon_p called in sms_schedule can make sure that SMS is only performed for innermost loops.
But I think it's a good idea to implement the TARGET_CAN_USE_DOLOOP_P hook here.
See my updated patch below. How about this one?
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,8 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.c (doloop_end): New pattern.
+ * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,47 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ ""
+{
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+ if (optimize > 0 && flag_modulo_sched)
+ {
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+ }
+ else
+ FAIL;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
Index: gcc/config/aarch64/aarch64.c
===================================================================
--- gcc/config/aarch64/aarch64.c (revision 217394)
+++ gcc/config/aarch64/aarch64.c (working copy)
@@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p (unsigned i
#define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
aarch64_use_by_pieces_infrastructure_p
+#undef TARGET_CAN_USE_DOLOOP_P
+#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
+
struct gcc_target targetm = TARGET_INITIALIZER;
#include "gt-aarch64.h"
>
> On 17 November 2014 07:59, Yangfei (Felix) <felix.yang@huawei.com> wrote:
>
> >> +2014-11-13 Felix Yang <felix.yang@huawei.com>
> >> +
> >> + * config/aarch64/aarch64.md (doloop_end): New pattern.
> >> +
>
> This looks like a straight copy of the ARM code, but without the
> TARGET_CAN_USE_DOLOOP_P definition. If the reason for including this code
> is for the benefit of module-sched then should the hook be defined to limit the
> use of this pattern to inner most loops only?
>
> Cheers
> /Marcus
[-- Attachment #2: aarch64-doloop-v4.diff --]
[-- Type: application/octet-stream, Size: 2575 bytes --]
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,8 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.c (doloop_end): New pattern.
+ * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,47 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ ""
+{
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+ if (optimize > 0 && flag_modulo_sched)
+ {
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+ }
+ else
+ FAIL;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
Index: gcc/config/aarch64/aarch64.c
===================================================================
--- gcc/config/aarch64/aarch64.c (revision 217394)
+++ gcc/config/aarch64/aarch64.c (working copy)
@@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p (unsigned i
#define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
aarch64_use_by_pieces_infrastructure_p
+#undef TARGET_CAN_USE_DOLOOP_P
+#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
+
struct gcc_target targetm = TARGET_INITIALIZER;
#include "gt-aarch64.h"
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 10:56 ` Yangfei (Felix)
@ 2014-11-18 11:29 ` Richard Henderson
2014-11-18 11:51 ` Yangfei (Felix)
0 siblings, 1 reply; 14+ messages in thread
From: Richard Henderson @ 2014-11-18 11:29 UTC (permalink / raw)
To: Yangfei (Felix), Marcus Shawcroft
Cc: gcc-patches, marcus.shawcroft, Richard Earnshaw
On 11/18/2014 11:48 AM, Yangfei (Felix) wrote:
> +(define_expand "doloop_end"
> + [(use (match_operand 0 "" "")) ; loop pseudo
> + (use (match_operand 1 "" ""))] ; label
> + ""
> +{
> + /* Currently SMS relies on the do-loop pattern to recognize loops
> + where (1) the control part consists of all insns defining and/or
> + using a certain 'count' register and (2) the loop count can be
> + adjusted by modifying this register prior to the loop.
> + ??? The possible introduction of a new block to initialize the
> + new IV can potentially affect branch optimizations. */
> + if (optimize > 0 && flag_modulo_sched)
You'd be better off moving this condition into the expansion predicate (which
is currently "").
This short-circuits a lot of unnecessary work. See pass_rtl_doloop::gate.
r~
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 11:29 ` Richard Henderson
@ 2014-11-18 11:51 ` Yangfei (Felix)
2014-11-18 11:59 ` Richard Henderson
` (2 more replies)
0 siblings, 3 replies; 14+ messages in thread
From: Yangfei (Felix) @ 2014-11-18 11:51 UTC (permalink / raw)
To: Richard Henderson, Marcus Shawcroft
Cc: gcc-patches, marcus.shawcroft, Richard Earnshaw
[-- Attachment #1: Type: text/plain, Size: 3516 bytes --]
> On 11/18/2014 11:48 AM, Yangfei (Felix) wrote:
> > +(define_expand "doloop_end"
> > + [(use (match_operand 0 "" "")) ; loop pseudo
> > + (use (match_operand 1 "" ""))] ; label
> > + ""
> > +{
> > + /* Currently SMS relies on the do-loop pattern to recognize loops
> > + where (1) the control part consists of all insns defining and/or
> > + using a certain 'count' register and (2) the loop count can be
> > + adjusted by modifying this register prior to the loop.
> > + ??? The possible introduction of a new block to initialize the
> > + new IV can potentially affect branch optimizations. */
> > + if (optimize > 0 && flag_modulo_sched)
>
> You'd be better off moving this condition into the expansion predicate (which is
> currently "").
>
> This short-circuits a lot of unnecessary work. See pass_rtl_doloop::gate.
>
>
> r~
Yeah, that's a good idea. See my updated patch :-)
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,8 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.c (doloop_end): New pattern.
+ * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,43 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ "optimize > 0 && flag_modulo_sched"
+{
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
Index: gcc/config/aarch64/aarch64.c
===================================================================
--- gcc/config/aarch64/aarch64.c (revision 217394)
+++ gcc/config/aarch64/aarch64.c (working copy)
@@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p (unsigned i
#define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
aarch64_use_by_pieces_infrastructure_p
+#undef TARGET_CAN_USE_DOLOOP_P
+#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
+
struct gcc_target targetm = TARGET_INITIALIZER;
#include "gt-aarch64.h"
[-- Attachment #2: aarch64-doloop-v5.diff --]
[-- Type: application/octet-stream, Size: 2484 bytes --]
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 217394)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,8 @@
+2014-11-13 Felix Yang <felix.yang@huawei.com>
+
+ * config/aarch64/aarch64.c (doloop_end): New pattern.
+ * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
+
2014-11-11 Andrew Pinski <apinski@cavium.com>
Bug target/61997
Index: gcc/config/aarch64/aarch64.md
===================================================================
--- gcc/config/aarch64/aarch64.md (revision 217394)
+++ gcc/config/aarch64/aarch64.md (working copy)
@@ -4087,6 +4087,43 @@
[(set_attr "type" "mrs")])
+;; Define the subtract-one-and-jump insns so loop.c
+;; knows what to generate.
+(define_expand "doloop_end"
+ [(use (match_operand 0 "" "")) ; loop pseudo
+ (use (match_operand 1 "" ""))] ; label
+ "optimize > 0 && flag_modulo_sched"
+{
+ rtx s0;
+ rtx bcomp;
+ rtx loc_ref;
+ rtx cc_reg;
+ rtx insn;
+ rtx cmp;
+
+ /* Currently SMS relies on the do-loop pattern to recognize loops
+ where (1) the control part consists of all insns defining and/or
+ using a certain 'count' register and (2) the loop count can be
+ adjusted by modifying this register prior to the loop.
+ ??? The possible introduction of a new block to initialize the
+ new IV can potentially affect branch optimizations. */
+
+ if (GET_MODE (operands[0]) != DImode)
+ FAIL;
+
+ s0 = operands [0];
+ insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
+
+ cmp = XVECEXP (PATTERN (insn), 0, 0);
+ cc_reg = SET_DEST (cmp);
+ bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
+ loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
+ emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
+ gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
+ loc_ref, pc_rtx)));
+ DONE;
+})
+
;; AdvSIMD Stuff
(include "aarch64-simd.md")
Index: gcc/config/aarch64/aarch64.c
===================================================================
--- gcc/config/aarch64/aarch64.c (revision 217394)
+++ gcc/config/aarch64/aarch64.c (working copy)
@@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p (unsigned i
#define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
aarch64_use_by_pieces_infrastructure_p
+#undef TARGET_CAN_USE_DOLOOP_P
+#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
+
struct gcc_target targetm = TARGET_INITIALIZER;
#include "gt-aarch64.h"
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 11:51 ` Yangfei (Felix)
@ 2014-11-18 11:59 ` Richard Henderson
2014-11-18 12:02 ` Marcus Shawcroft
2014-12-09 14:35 ` Tejas Belagod
2 siblings, 0 replies; 14+ messages in thread
From: Richard Henderson @ 2014-11-18 11:59 UTC (permalink / raw)
To: Yangfei (Felix), Marcus Shawcroft
Cc: gcc-patches, marcus.shawcroft, Richard Earnshaw
On 11/18/2014 12:28 PM, Yangfei (Felix) wrote:
> +2014-11-13 Felix Yang <felix.yang@huawei.com>
> +
> + * config/aarch64/aarch64.c (doloop_end): New pattern.
> + * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
Looks good to me. I'll leave it for aarch64 maintainers for final approval.
r~
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 11:51 ` Yangfei (Felix)
2014-11-18 11:59 ` Richard Henderson
@ 2014-11-18 12:02 ` Marcus Shawcroft
2014-12-09 14:35 ` Tejas Belagod
2 siblings, 0 replies; 14+ messages in thread
From: Marcus Shawcroft @ 2014-11-18 12:02 UTC (permalink / raw)
To: Yangfei (Felix)
Cc: Richard Henderson, gcc-patches, marcus.shawcroft, Richard Earnshaw
On 18 November 2014 11:28, Yangfei (Felix) <felix.yang@huawei.com> wrote:
> Yeah, that's a good idea. See my updated patch :-)
>
>
> Index: gcc/ChangeLog
> ===================================================================
> --- gcc/ChangeLog (revision 217394)
> +++ gcc/ChangeLog (working copy)
> @@ -1,3 +1,8 @@
> +2014-11-13 Felix Yang <felix.yang@huawei.com>
> +
> + * config/aarch64/aarch64.c (doloop_end): New pattern.
> + * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
> +
OK with me /Marcus
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-11-18 11:51 ` Yangfei (Felix)
2014-11-18 11:59 ` Richard Henderson
2014-11-18 12:02 ` Marcus Shawcroft
@ 2014-12-09 14:35 ` Tejas Belagod
2014-12-10 8:23 ` Yangfei (Felix)
2 siblings, 1 reply; 14+ messages in thread
From: Tejas Belagod @ 2014-12-09 14:35 UTC (permalink / raw)
To: Yangfei (Felix)
Cc: Richard Henderson, Marcus Shawcroft, gcc-patches,
Marcus Shawcroft, Richard Earnshaw
On 18/11/14 11:28, Yangfei (Felix) wrote:
>> On 11/18/2014 11:48 AM, Yangfei (Felix) wrote:
>>> +(define_expand "doloop_end"
>>> + [(use (match_operand 0 "" "")) ; loop pseudo
>>> + (use (match_operand 1 "" ""))] ; label
>>> + ""
>>> +{
>>> + /* Currently SMS relies on the do-loop pattern to recognize loops
>>> + where (1) the control part consists of all insns defining and/or
>>> + using a certain 'count' register and (2) the loop count can be
>>> + adjusted by modifying this register prior to the loop.
>>> + ??? The possible introduction of a new block to initialize the
>>> + new IV can potentially affect branch optimizations. */
>>> + if (optimize > 0 && flag_modulo_sched)
>>
>> You'd be better off moving this condition into the expansion predicate (which is
>> currently "").
>>
>> This short-circuits a lot of unnecessary work. See pass_rtl_doloop::gate.
>>
>>
>> r~
>
>
> Yeah, that's a good idea. See my updated patch :-)
>
>
> Index: gcc/ChangeLog
> ===================================================================
> --- gcc/ChangeLog (revision 217394)
> +++ gcc/ChangeLog (working copy)
> @@ -1,3 +1,8 @@
> +2014-11-13 Felix Yang <felix.yang@huawei.com>
> +
> + * config/aarch64/aarch64.c (doloop_end): New pattern.
> + * config/aarch64/aarch64.md (TARGET_CAN_USE_DOLOOP_P): Implement.
> +
> 2014-11-11 Andrew Pinski <apinski@cavium.com>
>
> Bug target/61997
> Index: gcc/config/aarch64/aarch64.md
> ===================================================================
> --- gcc/config/aarch64/aarch64.md (revision 217394)
> +++ gcc/config/aarch64/aarch64.md (working copy)
> @@ -4087,6 +4087,43 @@
> [(set_attr "type" "mrs")])
>
>
> +;; Define the subtract-one-and-jump insns so loop.c
> +;; knows what to generate.
> +(define_expand "doloop_end"
> + [(use (match_operand 0 "" "")) ; loop pseudo
> + (use (match_operand 1 "" ""))] ; label
> + "optimize > 0 && flag_modulo_sched"
> +{
> + rtx s0;
> + rtx bcomp;
> + rtx loc_ref;
> + rtx cc_reg;
> + rtx insn;
> + rtx cmp;
> +
> + /* Currently SMS relies on the do-loop pattern to recognize loops
> + where (1) the control part consists of all insns defining and/or
> + using a certain 'count' register and (2) the loop count can be
> + adjusted by modifying this register prior to the loop.
> + ??? The possible introduction of a new block to initialize the
> + new IV can potentially affect branch optimizations. */
> +
> + if (GET_MODE (operands[0]) != DImode)
> + FAIL;
> +
> + s0 = operands [0];
> + insn = emit_insn (gen_adddi3_compare0 (s0, s0, GEN_INT (-1)));
> +
> + cmp = XVECEXP (PATTERN (insn), 0, 0);
> + cc_reg = SET_DEST (cmp);
> + bcomp = gen_rtx_NE (VOIDmode, cc_reg, const0_rtx);
> + loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands [1]);
> + emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx,
> + gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp,
> + loc_ref, pc_rtx)));
> + DONE;
> +})
> +
> ;; AdvSIMD Stuff
> (include "aarch64-simd.md")
>
> Index: gcc/config/aarch64/aarch64.c
> ===================================================================
> --- gcc/config/aarch64/aarch64.c (revision 217394)
> +++ gcc/config/aarch64/aarch64.c (working copy)
> @@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p (unsigned i
> #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
> aarch64_use_by_pieces_infrastructure_p
>
> +#undef TARGET_CAN_USE_DOLOOP_P
> +#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
> +
> struct gcc_target targetm = TARGET_INITIALIZER;
>
> #include "gt-aarch64.h"
>
Hi Felix,
This patch causes https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64240
when sms-3 is tested with -fPIC. It runs fine when I reverse this patch out.
Please could you have a look?
Thanks,
Tejas.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
2014-12-09 14:35 ` Tejas Belagod
@ 2014-12-10 8:23 ` Yangfei (Felix)
0 siblings, 0 replies; 14+ messages in thread
From: Yangfei (Felix) @ 2014-12-10 8:23 UTC (permalink / raw)
To: Tejas Belagod; +Cc: gcc-patches, Marcus Shawcroft, Richard Earnshaw
> > --- gcc/config/aarch64/aarch64.c (revision 217394)
> > +++ gcc/config/aarch64/aarch64.c (working copy)
> > @@ -10224,6 +10224,9 @@ aarch64_use_by_pieces_infrastructure_p
> (unsigned i
> > #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
> > aarch64_use_by_pieces_infrastructure_p
> >
> > +#undef TARGET_CAN_USE_DOLOOP_P
> > +#define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
> > +
> > struct gcc_target targetm = TARGET_INITIALIZER;
> >
> > #include "gt-aarch64.h"
> >
>
> Hi Felix,
>
> This patch causes https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64240
> when sms-3 is tested with -fPIC. It runs fine when I reverse this patch out.
>
> Please could you have a look?
>
> Thanks,
> Tejas.
OK, I have reproduced with -fPIC option. Will take a look.
^ permalink raw reply [flat|nested] 14+ messages in thread
end of thread, other threads:[~2014-12-10 8:23 UTC | newest]
Thread overview: 14+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2014-11-12 10:02 [PING][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched Yangfei (Felix)
2014-11-12 10:55 ` Jiong Wang
2014-11-12 15:59 ` Andrew Pinski
2014-11-12 16:06 ` Richard Henderson
2014-11-13 10:45 ` [PING ^ 2][PATCH, " Yangfei (Felix)
2014-11-17 8:57 ` [PING ^ 3][PATCH, " Yangfei (Felix)
2014-11-18 10:18 ` Marcus Shawcroft
2014-11-18 10:56 ` Yangfei (Felix)
2014-11-18 11:29 ` Richard Henderson
2014-11-18 11:51 ` Yangfei (Felix)
2014-11-18 11:59 ` Richard Henderson
2014-11-18 12:02 ` Marcus Shawcroft
2014-12-09 14:35 ` Tejas Belagod
2014-12-10 8:23 ` Yangfei (Felix)
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).