PR target/103773: Fix wrong-code with -Oz from pop to memory.

  My apologies for the inconvenience.  The new support for -Oz using
push/pop for small integer constants on x86_64 is only a win/correct
for loading registers.  Fixed by adding !MEM_P tests in the appropriate
locations.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures.  Ok for mainline?


2021-12-21  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	PR target/103773
	* config/i386/i386.md (*movdi_internal): Only use short
	push/pop sequence for register (non-memory) destinations.
	(*movsi_internal): Likewise.

gcc/testsuite/ChangeLog
	PR target/103773
	* gcc.target/i386/pr103773.c: New test case.

Roger
--
  

On Tue, Dec 21, 2021 at 1:27 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> My apologies for the inconvenience.  The new support for -Oz using
> push/pop for small integer constants on x86_64 is only a win/correct
> for loading registers.  Fixed by adding !MEM_P tests in the appropriate
> locations.
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures.  Ok for mainline?
>
>
> 2021-12-21  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
>         PR target/103773
>         * config/i386/i386.md (*movdi_internal): Only use short
>         push/pop sequence for register (non-memory) destinations.
>         (*movsi_internal): Likewise.
>
> gcc/testsuite/ChangeLog
>         PR target/103773
>         * gcc.target/i386/pr103773.c: New test case.

Ouch, as pointed out in the PR, this approach clobbers the red zone.

Please revert the original patch.

Thanks,
Uros.

>
> Roger
> --
>
  

On Wed, Dec 22, 2021 at 9:10 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Tue, Dec 21, 2021 at 1:27 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
> >
> >
> > My apologies for the inconvenience.  The new support for -Oz using
> > push/pop for small integer constants on x86_64 is only a win/correct
> > for loading registers.  Fixed by adding !MEM_P tests in the appropriate
> > locations.
> >
> > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> > and make -k check with no new failures.  Ok for mainline?
> >
> >
> > 2021-12-21  Roger Sayle  <roger@nextmovesoftware.com>
> >
> > gcc/ChangeLog
> >         PR target/103773
> >         * config/i386/i386.md (*movdi_internal): Only use short
> >         push/pop sequence for register (non-memory) destinations.
> >         (*movsi_internal): Likewise.
> >
> > gcc/testsuite/ChangeLog
> >         PR target/103773
> >         * gcc.target/i386/pr103773.c: New test case.
>
> Ouch, as pointed out in the PR, this approach clobbers the red zone.
>
> Please revert the original patch.

*Maybe* we can use frame->red_zone_size here, but the frame is
recalculated several times during the compilation. I think it is just
too dangerous to use push/pop w.r.t. red zone clobbering.

Uros.
  

Hi Uros,
Would you consider the following variant that disables this optimization when a
red zone is used by the current function?  You're right that cfun's red_zone_size is
recalculated dynamically, but ix86_red_zone_used should be a better "gate" given
that this logic resides very late during compilation, in the output templates, where
whether or not a red zone is used is known.

On CSiBE, disabling this optimization in non-leaf functions that use a red zone costs
219 bytes, but remains a significant win over -Os.  (Alas the absolute numbers aren't
comparable as this testing included the 0/-1 write to memory changes).

Tested (overnight) on x86_64-pc-linux-gnu with make bootstrap and make -k check
with no new failures.

2021-12-22  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	PR target/103773
	* config/i386/i386.md (*movdi_internal): Only use short
	push/pop sequence for register (non-memory) destinations
	when the current function doesn't make use of a red zone.
	(*movsi_internal): Likewise.

gcc/testsuite/ChangeLog
	PR target/103773
	* gcc.target/i386/pr103773.c: New test case.

Please let me know what you think.  I'll revert, if this tweak doesn't address
your concerns.
Roger
--

> -----Original Message-----
> From: Uros Bizjak <ubizjak@gmail.com>
> Sent: 22 December 2021 08:20
> To: Roger Sayle <roger@nextmovesoftware.com>
> Cc: GCC Patches <gcc-patches@gcc.gnu.org>
> Subject: Re: [PATCH] PR target/103773: Fix wrong-code with -Oz from pop to
> memory.
> 
> On Wed, Dec 22, 2021 at 9:10 AM Uros Bizjak <ubizjak@gmail.com> wrote:
> >
> > On Tue, Dec 21, 2021 at 1:27 PM Roger Sayle
> <roger@nextmovesoftware.com> wrote:
> > >
> > >
> > > My apologies for the inconvenience.  The new support for -Oz using
> > > push/pop for small integer constants on x86_64 is only a win/correct
> > > for loading registers.  Fixed by adding !MEM_P tests in the
> > > appropriate locations.
> > >
> > > This patch has been tested on x86_64-pc-linux-gnu with make
> > > bootstrap and make -k check with no new failures.  Ok for mainline?
> > >
> > >
> > > 2021-12-21  Roger Sayle  <roger@nextmovesoftware.com>
> > >
> > > gcc/ChangeLog
> > >         PR target/103773
> > >         * config/i386/i386.md (*movdi_internal): Only use short
> > >         push/pop sequence for register (non-memory) destinations.
> > >         (*movsi_internal): Likewise.
> > >
> > > gcc/testsuite/ChangeLog
> > >         PR target/103773
> > >         * gcc.target/i386/pr103773.c: New test case.
> >
> > Ouch, as pointed out in the PR, this approach clobbers the red zone.
> >
> > Please revert the original patch.
> 
> *Maybe* we can use frame->red_zone_size here, but the frame is recalculated
> several times during the compilation. I think it is just too dangerous to use
> push/pop w.r.t. red zone clobbering.
> 
> Uros.
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index d25453f..489cede 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -2217,7 +2217,9 @@
 	  if (optimize_size > 1
 	      && TARGET_64BIT
 	      && CONST_INT_P (operands[1])
-	      && IN_RANGE (INTVAL (operands[1]), -128, 127))
+	      && IN_RANGE (INTVAL (operands[1]), -128, 127)
+	      && !MEM_P (operands[0])
+	      && !ix86_red_zone_used)
 	    return "push{q}\t%1\n\tpop{q}\t%0";
 	  return "mov{l}\t{%k1, %k0|%k0, %k1}";
 	}
@@ -2440,7 +2442,9 @@
 	return "lea{l}\t{%E1, %0|%0, %E1}";
       else if (optimize_size > 1
 	       && CONST_INT_P (operands[1])
-	       && IN_RANGE (INTVAL (operands[1]), -128, 127))
+	       && IN_RANGE (INTVAL (operands[1]), -128, 127)
+	       && !MEM_P (operands[0])
+	       && !ix86_red_zone_used)
 	{
 	  if (TARGET_64BIT)
 	    return "push{q}\t%1\n\tpop{q}\t%q0";
  

On Wed, Dec 22, 2021 at 10:26 AM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> Hi Uros,
> Would you consider the following variant that disables this optimization when a
> red zone is used by the current function?  You're right that cfun's red_zone_size is
> recalculated dynamically, but ix86_red_zone_used should be a better "gate" given
> that this logic resides very late during compilation, in the output templates, where
> whether or not a red zone is used is known.
>
> On CSiBE, disabling this optimization in non-leaf functions that use a red zone costs
> 219 bytes, but remains a significant win over -Os.  (Alas the absolute numbers aren't
> comparable as this testing included the 0/-1 write to memory changes).
>
> Tested (overnight) on x86_64-pc-linux-gnu with make bootstrap and make -k check
> with no new failures.
>
> 2021-12-22  Roger Sayle  <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
>         PR target/103773
>         * config/i386/i386.md (*movdi_internal): Only use short
>         push/pop sequence for register (non-memory) destinations
>         when the current function doesn't make use of a red zone.
>         (*movsi_internal): Likewise.
>
> gcc/testsuite/ChangeLog
>         PR target/103773
>         * gcc.target/i386/pr103773.c: New test case.
>
> Please let me know what you think.  I'll revert, if this tweak doesn't address
> your concerns.

Yes, using ix86_red_zone_used looks safe.

OTOH, is there a reason the transformation is not implemented via
peephole2 pass? IIRC, frame is stable after pro_and_epilogue_pass, and
peephole2 pass is instanced well after register allocation.

Uros.
  

On Wed, Dec 22, 2021 at 11:26 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Wed, Dec 22, 2021 at 10:26 AM Roger Sayle <roger@nextmovesoftware.com> wrote:
> >
> >
> > Hi Uros,
> > Would you consider the following variant that disables this optimization when a
> > red zone is used by the current function?  You're right that cfun's red_zone_size is
> > recalculated dynamically, but ix86_red_zone_used should be a better "gate" given
> > that this logic resides very late during compilation, in the output templates, where
> > whether or not a red zone is used is known.
> >
> > On CSiBE, disabling this optimization in non-leaf functions that use a red zone costs
> > 219 bytes, but remains a significant win over -Os.  (Alas the absolute numbers aren't
> > comparable as this testing included the 0/-1 write to memory changes).
> >
> > Tested (overnight) on x86_64-pc-linux-gnu with make bootstrap and make -k check
> > with no new failures.
> >
> > 2021-12-22  Roger Sayle  <roger@nextmovesoftware.com>
> >
> > gcc/ChangeLog
> >         PR target/103773
> >         * config/i386/i386.md (*movdi_internal): Only use short
> >         push/pop sequence for register (non-memory) destinations
> >         when the current function doesn't make use of a red zone.
> >         (*movsi_internal): Likewise.
> >
> > gcc/testsuite/ChangeLog
> >         PR target/103773
> >         * gcc.target/i386/pr103773.c: New test case.
> >
> > Please let me know what you think.  I'll revert, if this tweak doesn't address
> > your concerns.
>
> Yes, using ix86_red_zone_used looks safe.
>
> OTOH, is there a reason the transformation is not implemented via
> peephole2 pass? IIRC, frame is stable after pro_and_epilogue_pass, and
> peephole2 pass is instanced well after register allocation.

Something like the attached patch.

Uros.
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index 58b10643fcb..e5d603f0025 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -2514,6 +2514,24 @@
 	   ]
 	   (symbol_ref "true")))])
 
+(define_peephole2
+  [(set (match_operand:SWI48 0 "general_reg_operand")
+	(match_operand:SWI48 1 "const_int_operand"))]
+  "optimize_insn_for_size_p () && optimize_size > 1
+   && IN_RANGE (INTVAL (operands[1]), -128, 127)
+   && !ix86_red_zone_used"
+  [(set (match_dup 2) (match_dup 1))
+   (set (match_dup 0) (match_dup 3))]
+{
+  if (GET_MODE (operands[0]) != word_mode)
+    operands[0] = gen_rtx_REG (word_mode, REGNO (operands[0]));
+
+  operands[2] = gen_rtx_MEM (word_mode,
+			     gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx));
+  operands[3] = gen_rtx_MEM (word_mode,
+			     gen_rtx_POST_INC (Pmode, stack_pointer_rtx));
+})
+
 (define_insn "*movhi_internal"
   [(set (match_operand:HI 0 "nonimmediate_operand"
     "=r,r,r,m ,*k,*k ,r ,m ,*k ,?r,?*v,*v,*v,*v,m")