This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]
Other format:	[Raw text]

Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched

From: Tejas Belagod <tejas dot belagod at arm dot com>
To: "Yangfei (Felix)" <felix dot yang at huawei dot com>
Cc: Richard Henderson <rth at redhat dot com>, Marcus Shawcroft <marcus dot shawcroft at gmail dot com>, "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>, Marcus Shawcroft <Marcus dot Shawcroft at arm dot com>, Richard Earnshaw <Richard dot Earnshaw at arm dot com>
Date: Tue, 09 Dec 2014 14:35:24 +0000
Subject: Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
Authentication-results: sourceware.org; auth=none
References: <DA41BE1DDCA941489001C7FBD7A8820E555559C9 at szxema507-mbx dot china dot huawei dot com> <54638488 dot 5070808 at redhat dot com> <DA41BE1DDCA941489001C7FBD7A8820E555584B0 at szxema507-mbx dot china dot huawei dot com> <CAFqB+PyR=Ga3afj+b1M0r_GArEe_4n=U2z56YYwaeXYYpZE+BQ at mail dot gmail dot com> <DA41BE1DDCA941489001C7FBD7A8820E555589A3 at szxema507-mbx dot china dot huawei dot com> <546B25E3 dot 8030907 at redhat dot com> <DA41BE1DDCA941489001C7FBD7A8820E555589E6 at szxema507-mbx dot china dot huawei dot com>

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=64240when sms-3 is tested with -fPIC. It runs fine when I reverse this patch out.


Please could you have a look?

Thanks,
Tejas.

Follow-Ups:
- Re: [PING ^ 3][PATCH, AArch64] Add doloop_end pattern for -fmodulo-sched
  - From: Yangfei (Felix)

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]