[PATCH] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[PATCH] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

Hello,

This enhances location description generation so that the generated
opcodes for integer literals are as space-efficient when HOST_WIDE_INT
is 64-bits large than when it's 32-bits large. In particular, this
reduces the size of the opcodes generated to produce big unsigned
literals using small literal integers instead.

Bootstrapped and regtested on x86-linux (no regression).  I also checked
that the new testcase fails with mainline.  Ok to commit?  Thank you in
advance!

gcc/

	* dwarf2out.c (int_loc_descriptor): Generate opcodes for another
	equivalent 32-bit constant (modulo 2**32) when that yields
	smaller instructions.
	(size_of_int_loc_descriptor): Update accordingly.
---
 gcc/dwarf2out.c                  | 31 ++++++++++++++++++++++++++-----
 gcc/testsuite/gnat.dg/debug8.adb | 29 +++++++++++++++++++++++++++++
 2 files changed, 55 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gnat.dg/debug8.adb

diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c
index 0fdab9a..f175ea1 100644
--- a/gcc/dwarf2out.c
+++ b/gcc/dwarf2out.c
@@ -11954,20 +11954,36 @@ int_loc_descriptor (HOST_WIDE_INT i)
 	/* DW_OP_const1u X DW_OP_litY DW_OP_shl takes just 4 bytes,
 	   while DW_OP_const4u is 5 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 8);
+
+      else if (HOST_BITS_PER_WIDE_INT > 32
+	       && DWARF2_ADDR_SIZE == 4 && i > 0x7fffffff
+	       && size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i)
+		  <= 4)
+	{
+	  /* As i >= 2**31, the double cast above will yield a negative number.
+	     Since wrapping is defined in DWARF expressions we can output big
+	     positive integers as small negative ones, regardless of the size
+	     of host wide ints.
+
+	     Here, since the evaluator will handle 32-bit values and since i >=
+	     2**31, we know it's going to be interpreted as a negative literal:
+	     store it this way if we can do better than 5 bytes this way.  */
+	  return int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i);
+	}
       else if (HOST_BITS_PER_WIDE_INT == 32 || i <= 0xffffffff)
 	op = DW_OP_const4u;
+
+      /* Past this point, i >= 0x100000000 and thus DW_OP_constu will take at
+	 least 6 bytes: see if we can do better before falling back to it.  */
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 8
 	       && clz + 8 + 255 >= HOST_BITS_PER_WIDE_INT)
-	/* DW_OP_const1u X DW_OP_const1u Y DW_OP_shl takes just 5 bytes,
-	   while DW_OP_constu of constant >= 0x100000000 takes at least
-	   6 bytes.  */
+	/* DW_OP_const1u X DW_OP_const1u Y DW_OP_shl takes just 5 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 8);
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 16
 	       && clz + 16 + (size_of_uleb128 (i) > 5 ? 255 : 31)
 		  >= HOST_BITS_PER_WIDE_INT)
 	/* DW_OP_const2u X DW_OP_litY DW_OP_shl takes just 5 bytes,
-	   DW_OP_const2u X DW_OP_const1u Y DW_OP_shl takes 6 bytes,
-	   while DW_OP_constu takes in this case at least 6 bytes.  */
+	   DW_OP_const2u X DW_OP_const1u Y DW_OP_shl takes 6 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 16);
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 32
 	       && clz + 32 + 31 >= HOST_BITS_PER_WIDE_INT
@@ -12192,6 +12208,11 @@ size_of_int_loc_descriptor (HOST_WIDE_INT i)
 	       && clz + 8 + 31 >= HOST_BITS_PER_WIDE_INT)
 	return size_of_int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT
 						    - clz - 8);
+      else if (HOST_BITS_PER_WIDE_INT > 32
+	       && DWARF2_ADDR_SIZE == 4 && i > 0x7fffffff
+	       && size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i)
+		  <= 4)
+	return size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i);
       else if (HOST_BITS_PER_WIDE_INT == 32 || i <= 0xffffffff)
 	return 5;
       s = size_of_uleb128 ((unsigned HOST_WIDE_INT) i);
diff --git a/gcc/testsuite/gnat.dg/debug8.adb b/gcc/testsuite/gnat.dg/debug8.adb
new file mode 100644
index 0000000..fabcc22
--- /dev/null
+++ b/gcc/testsuite/gnat.dg/debug8.adb
@@ -0,0 +1,29 @@
+-- { dg-do compile }
+-- { dg-options "-cargs -g -fgnat-encodings=minimal -dA" }
+-- { dg-final { scan-assembler-not "DW_OP_const4u" } }
+-- { dg-final { scan-assembler-not "DW_OP_const8u" } }
+
+--  The DW_AT_byte_size attribute DWARF expression for the
+--  DW_TAG_structure_type DIE that describes Rec_Type contains the -4u literal.
+--  Check that it is not created using an inefficient encoding (DW_OP_const1s
+--  is expected).
+
+procedure Debug8 is
+
+   type Rec_Type (I : Integer) is record
+      B : Boolean;
+      case I is
+         when 0 =>
+            null;
+         when 1 .. 10 =>
+            C : Character;
+         when others =>
+            N : Natural;
+      end case;
+   end record;
+
+   R : access Rec_Type := null;
+
+begin
+   null;
+end Debug8;
-- 
2.9.3

Re: [PATCH] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Trevor Saunders]

On Thu, Aug 18, 2016 at 11:33:06AM +0200, Pierre-Marie de Rodat wrote:
> Hello,
> 
> This enhances location description generation so that the generated
> opcodes for integer literals are as space-efficient when HOST_WIDE_INT
> is 64-bits large than when it's 32-bits large. In particular, this
> reduces the size of the opcodes generated to produce big unsigned
> literals using small literal integers instead.
> 
> Bootstrapped and regtested on x86-linux (no regression).  I also checked
> that the new testcase fails with mainline.  Ok to commit?  Thank you in
> advance!
> 
> gcc/
> 
> 	* dwarf2out.c (int_loc_descriptor): Generate opcodes for another
> 	equivalent 32-bit constant (modulo 2**32) when that yields
> 	smaller instructions.
> 	(size_of_int_loc_descriptor): Update accordingly.
> ---
>  gcc/dwarf2out.c                  | 31 ++++++++++++++++++++++++++-----
>  gcc/testsuite/gnat.dg/debug8.adb | 29 +++++++++++++++++++++++++++++
>  2 files changed, 55 insertions(+), 5 deletions(-)
>  create mode 100644 gcc/testsuite/gnat.dg/debug8.adb
> 
> diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c
> index 0fdab9a..f175ea1 100644
> --- a/gcc/dwarf2out.c
> +++ b/gcc/dwarf2out.c
> @@ -11954,20 +11954,36 @@ int_loc_descriptor (HOST_WIDE_INT i)
>  	/* DW_OP_const1u X DW_OP_litY DW_OP_shl takes just 4 bytes,
>  	   while DW_OP_const4u is 5 bytes.  */
>  	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 8);
> +
> +      else if (HOST_BITS_PER_WIDE_INT > 32

HOST_BITS_PER_WIDE_INT is now always 64, so you don't need that check.

Trev

Re: [PATCH] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

On 08/18/2016 01:29 PM, Trevor Saunders wrote:
> HOST_BITS_PER_WIDE_INT is now always 64, so you don't need that check.

Good point, thank you! I’ll adapt the patch.

-- 
Pierre-Marie de Rodat

[PATCH, PING] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

[-- Attachment #1: Type: text/plain, Size: 190 bytes --]

Hello,

Ping for the patch submitted at 
<https://gcc.gnu.org/ml/gcc-patches/2016-08/msg01314.html>. Also, here’s 
the update that Trevor suggested.

Thanks!

-- 
Pierre-Marie de Rodat

[-- Attachment #2: 0001-DWARF-space-optimize-loc.-descr.-for-integer-literal.patch --]
[-- Type: text/x-diff, Size: 4831 bytes --]

From da6753ad98127778e8923db63ce51b52a853ed17 Mon Sep 17 00:00:00 2001
From: Pierre-Marie de Rodat <derodat@adacore.com>
Date: Fri, 24 Jul 2015 15:48:33 +0200
Subject: [PATCH] DWARF: space-optimize loc. descr. for integer literals on
 32-bit targets

This enhances location description generation so that the generated
opcodes for integer literals are as space-efficient when HOST_WIDE_INT
is 64-bits large than when it's 32-bits large. In particular, this
reduces the size of the opcodes generated to produce big unsigned
literals using small literal integers instead.

gcc/

	* dwarf2out.c (int_loc_descriptor): Generate opcodes for another
	equivalent 32-bit constant (modulo 2**32) when that yields
	smaller instructions.
	(size_of_int_loc_descriptor): Update accordingly.
---
 gcc/dwarf2out.c                  | 29 ++++++++++++++++++++++++-----
 gcc/testsuite/gnat.dg/debug8.adb | 29 +++++++++++++++++++++++++++++
 2 files changed, 53 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gnat.dg/debug8.adb

diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c
index 0fdab9a..2591f89 100644
--- a/gcc/dwarf2out.c
+++ b/gcc/dwarf2out.c
@@ -11954,20 +11954,35 @@ int_loc_descriptor (HOST_WIDE_INT i)
 	/* DW_OP_const1u X DW_OP_litY DW_OP_shl takes just 4 bytes,
 	   while DW_OP_const4u is 5 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 8);
+
+      else if (DWARF2_ADDR_SIZE == 4 && i > 0x7fffffff
+	       && size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i)
+		  <= 4)
+	{
+	  /* As i >= 2**31, the double cast above will yield a negative number.
+	     Since wrapping is defined in DWARF expressions we can output big
+	     positive integers as small negative ones, regardless of the size
+	     of host wide ints.
+
+	     Here, since the evaluator will handle 32-bit values and since i >=
+	     2**31, we know it's going to be interpreted as a negative literal:
+	     store it this way if we can do better than 5 bytes this way.  */
+	  return int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i);
+	}
       else if (HOST_BITS_PER_WIDE_INT == 32 || i <= 0xffffffff)
 	op = DW_OP_const4u;
+
+      /* Past this point, i >= 0x100000000 and thus DW_OP_constu will take at
+	 least 6 bytes: see if we can do better before falling back to it.  */
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 8
 	       && clz + 8 + 255 >= HOST_BITS_PER_WIDE_INT)
-	/* DW_OP_const1u X DW_OP_const1u Y DW_OP_shl takes just 5 bytes,
-	   while DW_OP_constu of constant >= 0x100000000 takes at least
-	   6 bytes.  */
+	/* DW_OP_const1u X DW_OP_const1u Y DW_OP_shl takes just 5 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 8);
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 16
 	       && clz + 16 + (size_of_uleb128 (i) > 5 ? 255 : 31)
 		  >= HOST_BITS_PER_WIDE_INT)
 	/* DW_OP_const2u X DW_OP_litY DW_OP_shl takes just 5 bytes,
-	   DW_OP_const2u X DW_OP_const1u Y DW_OP_shl takes 6 bytes,
-	   while DW_OP_constu takes in this case at least 6 bytes.  */
+	   DW_OP_const2u X DW_OP_const1u Y DW_OP_shl takes 6 bytes.  */
 	return int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT - clz - 16);
       else if (clz + ctz >= HOST_BITS_PER_WIDE_INT - 32
 	       && clz + 32 + 31 >= HOST_BITS_PER_WIDE_INT
@@ -12192,6 +12207,10 @@ size_of_int_loc_descriptor (HOST_WIDE_INT i)
 	       && clz + 8 + 31 >= HOST_BITS_PER_WIDE_INT)
 	return size_of_int_shift_loc_descriptor (i, HOST_BITS_PER_WIDE_INT
 						    - clz - 8);
+      else if (DWARF2_ADDR_SIZE == 4 && i > 0x7fffffff
+	       && size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i)
+		  <= 4)
+	return size_of_int_loc_descriptor ((HOST_WIDE_INT) (int32_t) i);
       else if (HOST_BITS_PER_WIDE_INT == 32 || i <= 0xffffffff)
 	return 5;
       s = size_of_uleb128 ((unsigned HOST_WIDE_INT) i);
diff --git a/gcc/testsuite/gnat.dg/debug8.adb b/gcc/testsuite/gnat.dg/debug8.adb
new file mode 100644
index 0000000..fabcc22
--- /dev/null
+++ b/gcc/testsuite/gnat.dg/debug8.adb
@@ -0,0 +1,29 @@
+-- { dg-do compile }
+-- { dg-options "-cargs -g -fgnat-encodings=minimal -dA" }
+-- { dg-final { scan-assembler-not "DW_OP_const4u" } }
+-- { dg-final { scan-assembler-not "DW_OP_const8u" } }
+
+--  The DW_AT_byte_size attribute DWARF expression for the
+--  DW_TAG_structure_type DIE that describes Rec_Type contains the -4u literal.
+--  Check that it is not created using an inefficient encoding (DW_OP_const1s
+--  is expected).
+
+procedure Debug8 is
+
+   type Rec_Type (I : Integer) is record
+      B : Boolean;
+      case I is
+         when 0 =>
+            null;
+         when 1 .. 10 =>
+            C : Character;
+         when others =>
+            N : Natural;
+      end case;
+   end record;
+
+   R : access Rec_Type := null;
+
+begin
+   null;
+end Debug8;
-- 
1.8.2.1

[PATCH, PING*2] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

Hello,

Ping for the patch submitted at 
<https://gcc.gnu.org/ml/gcc-patches/2016-08/msg02117.html>. Thanks!

-- 
Pierre-Marie de Rodat

[PATCH, PING*3] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

Hello,

Ping for the patch submitted at 
<https://gcc.gnu.org/ml/gcc-patches/2016-08/msg02117.html>. Thanks!

-- 
Pierre-Marie de Rodat

Re: [PATCH, PING] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Richard Biener]

On Wed, Aug 31, 2016 at 1:41 PM, Pierre-Marie de Rodat
<derodat@adacore.com> wrote:
> Hello,
>
> Ping for the patch submitted at
> <https://gcc.gnu.org/ml/gcc-patches/2016-08/msg01314.html>. Also, here’s the
> update that Trevor suggested.

Ok.

Thanks,
Richard.

> Thanks!
>
> --
> Pierre-Marie de Rodat

Re: [PATCH, PING] DWARF: space-optimize loc. descr. for integer literals on 32-bit targets

[Pierre-Marie de Rodat]

On 10/12/2016 10:13 AM, Richard Biener wrote:
> Ok.
>
> Thanks,

Committed. Thank you!

-- 
Pierre-Marie de Rodat