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: [PATCH] Rewrite pic.md to improve medany and pic code size.

From: Palmer Dabbelt <palmer at sifive dot com>
To: Jim Wilson <jimw at sifive dot com>
Cc: gcc-patches at gcc dot gnu dot org
Date: Wed, 29 Aug 2018 09:22:27 -0700 (PDT)
Subject: Re: [PATCH] Rewrite pic.md to improve medany and pic code size.

On Tue, 28 Aug 2018 19:21:23 PDT (-0700), Jim Wilson wrote:

The pic.md file has patterns used only for the medany code model and for pic
code.  They match an unsplit 2-instruction address load pattern followed by
a load or store instruction, and emit an assembler macro that expands to two
instructions.  This replaces 3 instructions with 2.  Unfortunately, there
are a lot of broken patterns in the file.  It has things like
   (sign_extend:ANYI (mem:ANYI ...)
which can't work, as the sign_extend needs to be a larger mode than the mem.
There are also a lot of missing patterns.  It has patterns for sign and zero
extending loads, but not regular loads for instance.  Fixing these problems
required a major rewrite of the file.  While doing this, I noticed a case
that results in worse code, and had to change address costs to fix that.
I also ended up adding a number of new mode attributes and mode iterators
to support the new pic.md file.

Thanks Jim -- I'm afraid at least part of this was my mess, as I had to go addin the ZERO_EXTEND_LOAD hackery to work around some bug in this file that Icouldn't figure out how to fix in a better way.

This was tested with cross rv{32,64}-{elf,linux} builds and checks using the
medany code model.  There were no regressions.  I'm seeing about a 0.3% to 0.4%
code size reduction for newlib/glibc libc.a/libc.so.  This was also tested
with a cross kernel build and boot on qemu to verify that I didn't break kernel
builds.

IIRC the issue I found was in booting a Linux kernel that was built with"-mcmodel=medany". If I'm reading our current kernel port correctly that's thedefault for 64-bit systems, but I'm not sure when we flipped over so I thinkit's worth checking -- I know at the time I fixed the bug it wasn't the default:). You should be able to do a "make V=1" inside Linux to just see thecompiler command lines.

Committed.


Thanks for the fix!


Jim

	gcc/
	* config/riscv/pic.md: Rewrite.
	* config/riscv/riscv.c (riscv_address_insns): Return cost of 3 for
	invalid address.
	* config/riscv/riscv.md (ZERO_EXTEND_LOAD): Delete.
	(SOFTF, default_load, softload, softstore): New.
---
 gcc/config/riscv/pic.md   | 113 ++++++++++++++++++++++++--------------
 gcc/config/riscv/riscv.c  |   8 ++-
 gcc/config/riscv/riscv.md |  16 +++++-
 3 files changed, 91 insertions(+), 46 deletions(-)

diff --git a/gcc/config/riscv/pic.md b/gcc/config/riscv/pic.md
index a4a9732656c..942502058e0 100644
--- a/gcc/config/riscv/pic.md
+++ b/gcc/config/riscv/pic.md
@@ -22,71 +22,100 @@
 ;; Simplify PIC loads to static variables.
 ;; These should go away once we figure out how to emit auipc discretely.

-(define_insn "*local_pic_load_s<mode>"
+(define_insn "*local_pic_load<mode>"
   [(set (match_operand:ANYI 0 "register_operand" "=r")
-	(sign_extend:ANYI (mem:ANYI (match_operand 1 "absolute_symbolic_operand" ""))))]
+	(mem:ANYI (match_operand 1 "absolute_symbolic_operand" "")))]
+  "USE_LOAD_ADDRESS_MACRO (operands[1])"
+  "<default_load>\t%0,%1"
+  [(set (attr "length") (const_int 8))])
+
+(define_insn "*local_pic_load_s<mode>"
+  [(set (match_operand:SUPERQI 0 "register_operand" "=r")
+	(sign_extend:SUPERQI (mem:SUBX (match_operand 1 "absolute_symbolic_operand" ""))))]
   "USE_LOAD_ADDRESS_MACRO (operands[1])"
-  "<load>\t%0,%1"
+  "<SUBX:load>\t%0,%1"
   [(set (attr "length") (const_int 8))])

 (define_insn "*local_pic_load_u<mode>"
-  [(set (match_operand:ZERO_EXTEND_LOAD 0 "register_operand" "=r")
-	(zero_extend:ZERO_EXTEND_LOAD (mem:ZERO_EXTEND_LOAD (match_operand 1 "absolute_symbolic_operand" ""))))]
+  [(set (match_operand:SUPERQI 0 "register_operand" "=r")
+	(zero_extend:SUPERQI (mem:SUBX (match_operand 1 "absolute_symbolic_operand" ""))))]
   "USE_LOAD_ADDRESS_MACRO (operands[1])"
-  "<load>u\t%0,%1"
+  "<SUBX:load>u\t%0,%1"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_load<mode>"
-  [(set (match_operand:ANYF 0 "register_operand" "=f")
+;; We can support ANYF loads into X register if there is no double support
+;; or if the target is 64-bit.
+
+(define_insn "*local_pic_load<ANYF:mode>"
+  [(set (match_operand:ANYF 0 "register_operand" "=f,*r")
 	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
-   (clobber (match_scratch:DI 2 "=r"))]
-  "TARGET_HARD_FLOAT && TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[1])"
-  "<load>\t%0,%1,%2"
+   (clobber (match_scratch:P 2 "=r,X"))]
+  "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
+   && (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
+  "@
+   <ANYF:load>\t%0,%1,%2
+   <softload>\t%0,%1"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_load<mode>"
+;; ??? For a 32-bit target with double float, a DF load into a X reg isn't
+;; supported.  ld is not valid in that case.  Punt for now.  Maybe add a split
+;; for this later.
+
+(define_insn "*local_pic_load_32d<ANYF:mode>"
   [(set (match_operand:ANYF 0 "register_operand" "=f")
 	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
-   (clobber (match_scratch:SI 2 "=r"))]
-  "TARGET_HARD_FLOAT && !TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[1])"
-  "<load>\t%0,%1,%2"
+   (clobber (match_scratch:P 2 "=r"))]
+  "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
+   && (TARGET_DOUBLE_FLOAT && !TARGET_64BIT)"
+  "<ANYF:load>\t%0,%1,%2"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_loadu<mode>"
-  [(set (match_operand:SUPERQI 0 "register_operand" "=r")
-	(zero_extend:SUPERQI (mem:SUBX (match_operand 1 "absolute_symbolic_operand" ""))))]
-  "USE_LOAD_ADDRESS_MACRO (operands[1])"
-  "<load>u\t%0,%1"
+(define_insn "*local_pic_load_sf<mode>"
+  [(set (match_operand:SOFTF 0 "register_operand" "=r")
+	(mem:SOFTF (match_operand 1 "absolute_symbolic_operand" "")))]
+  "!TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])"
+  "<softload>\t%0,%1"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_storedi<mode>"
-  [(set (mem:ANYI (match_operand 0 "absolute_symbolic_operand" ""))
-	(match_operand:ANYI 1 "reg_or_0_operand" "rJ"))
-   (clobber (match_scratch:DI 2 "=&r"))]
-  "TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[0])"
-  "<store>\t%z1,%0,%2"
-  [(set (attr "length") (const_int 8))])
+;; Simplify PIC stores to static variables.
+;; These should go away once we figure out how to emit auipc discretely.

-(define_insn "*local_pic_storesi<mode>"
+(define_insn "*local_pic_store<ANYI:mode>"
   [(set (mem:ANYI (match_operand 0 "absolute_symbolic_operand" ""))
 	(match_operand:ANYI 1 "reg_or_0_operand" "rJ"))
-   (clobber (match_scratch:SI 2 "=&r"))]
-  "!TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[0])"
-  "<store>\t%z1,%0,%2"
+   (clobber (match_scratch:P 2 "=&r"))]
+  "USE_LOAD_ADDRESS_MACRO (operands[0])"
+  "<ANYI:store>\t%z1,%0,%2"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_storedi<mode>"
+(define_insn "*local_pic_store<ANYF:mode>"
   [(set (mem:ANYF (match_operand 0 "absolute_symbolic_operand" ""))
-	(match_operand:ANYF 1 "register_operand" "f"))
-   (clobber (match_scratch:DI 2 "=r"))]
-  "TARGET_HARD_FLOAT && TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[0])"
-  "<store>\t%1,%0,%2"
+	(match_operand:ANYF 1 "register_operand" "f,*r"))
+   (clobber (match_scratch:P 2 "=r,&r"))]
+  "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[0])
+   && (!TARGET_DOUBLE_FLOAT || TARGET_64BIT)"
+  "@
+   <ANYF:store>\t%1,%0,%2
+   <softstore>\t%1,%0,%2"
   [(set (attr "length") (const_int 8))])

-(define_insn "*local_pic_storesi<mode>"
-  [(set (mem:ANYF (match_operand 0 "absolute_symbolic_operand" ""))
-	(match_operand:ANYF 1 "register_operand" "f"))
-   (clobber (match_scratch:SI 2 "=r"))]
-  "TARGET_HARD_FLOAT && !TARGET_64BIT && USE_LOAD_ADDRESS_MACRO (operands[0])"
-  "<store>\t%1,%0,%2"
+;; ??? For a 32-bit target with double float, a DF store from a X reg isn't
+;; supported.  sd is not valid in that case.  Punt for now.  Maybe add a split
+;; for this later.
+
+(define_insn "*local_pic_store_32d<ANYF:mode>"
+  [(set (match_operand:ANYF 0 "register_operand" "=f")
+	(mem:ANYF (match_operand 1 "absolute_symbolic_operand" "")))
+   (clobber (match_scratch:P 2 "=r"))]
+  "TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[1])
+   && (TARGET_DOUBLE_FLOAT && !TARGET_64BIT)"
+  "<ANYF:store>\t%1,%0,%2"
+  [(set (attr "length") (const_int 8))])
+
+(define_insn "*local_pic_store_sf<SOFTF:mode>"
+  [(set (mem:SOFTF (match_operand 0 "absolute_symbolic_operand" ""))
+	(match_operand:SOFTF 1 "register_operand" "r"))
+   (clobber (match_scratch:P 2 "=&r"))]
+  "!TARGET_HARD_FLOAT && USE_LOAD_ADDRESS_MACRO (operands[0])"
+  "<softstore>\t%1,%0,%2"
   [(set (attr "length") (const_int 8))])
diff --git a/gcc/config/riscv/riscv.c b/gcc/config/riscv/riscv.c
index 69e70feaf33..9d6d981a42a 100644
--- a/gcc/config/riscv/riscv.c
+++ b/gcc/config/riscv/riscv.c
@@ -802,7 +802,13 @@ riscv_address_insns (rtx x, machine_mode mode, bool might_split_p)
   int n = 1;

   if (!riscv_classify_address (&addr, x, mode, false))
-    return 0;
+    {
+      /* This could be a pattern from the pic.md file.  In which case we want
+	 this address to always have a cost of 3 to make it as expensive as the
+	 most expensive symbol.  This prevents constant propagation from
+	 preferring symbols over register plus offset.  */
+      return 3;
+    }

   /* BLKmode is used for single unaligned loads and stores and should
      not count as a multiword mode. */
diff --git a/gcc/config/riscv/riscv.md b/gcc/config/riscv/riscv.md
index 613af9d79e4..95fbb282c7c 100644
--- a/gcc/config/riscv/riscv.md
+++ b/gcc/config/riscv/riscv.md
@@ -269,9 +269,6 @@
 ;; Iterator for QImode extension patterns.
 (define_mode_iterator SUPERQI [HI SI (DI "TARGET_64BIT")])

-;; Iterator for extending loads.
-(define_mode_iterator ZERO_EXTEND_LOAD [QI HI (SI "TARGET_64BIT")])
-
 ;; Iterator for hardware integer modes narrower than XLEN.
 (define_mode_iterator SUBX [QI HI (SI "TARGET_64BIT")])

@@ -282,6 +279,9 @@
 (define_mode_iterator ANYF [(SF "TARGET_HARD_FLOAT")
 			    (DF "TARGET_DOUBLE_FLOAT")])

+;; Iterator for floating-point modes that can be loaded into X registers.
+(define_mode_iterator SOFTF [SF (DF "TARGET_64BIT")])
+
 ;; This attribute gives the length suffix for a sign- or zero-extension
 ;; instruction.
 (define_mode_attr size [(QI "b") (HI "h")])
@@ -289,9 +289,19 @@
 ;; Mode attributes for loads.
 (define_mode_attr load [(QI "lb") (HI "lh") (SI "lw") (DI "ld") (SF "flw") (DF "fld")])

+;; Instruction names for integer loads that aren't explicitly sign or zero
+;; extended.  See riscv_output_move and LOAD_EXTEND_OP.
+(define_mode_attr default_load [(QI "lbu") (HI "lhu") (SI "lw") (DI "ld")])
+
+;; Mode attribute for FP loads into integer registers.
+(define_mode_attr softload [(SF "lw") (DF "ld")])
+
 ;; Instruction names for stores.
 (define_mode_attr store [(QI "sb") (HI "sh") (SI "sw") (DI "sd") (SF "fsw") (DF "fsd")])

+;; Instruction names for FP stores from integer registers.
+(define_mode_attr softstore [(SF "sw") (DF "sd")])
+
 ;; This attribute gives the best constraint to use for registers of
 ;; a given mode.
 (define_mode_attr reg [(SI "d") (DI "d") (CC "d")])

Follow-Ups:
- Re: [PATCH] Rewrite pic.md to improve medany and pic code size.
  - From: Jim Wilson

References:
- [PATCH] Rewrite pic.md to improve medany and pic code size.
  - From: Jim Wilson

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