PATCH: PR target/45946: ICE: in extract_insn, at recog.c:2127 when using _Decimal128 with -Os -fno-omit-frame-pointer

On Thu, Oct 21, 2010 at 8:32 AM, Uros Bizjak <ubizjak@gmail.com> wrote:
> On Wed, Oct 20, 2010 at 8:02 PM, H.J. Lu <hongjiu.lu@intel.com> wrote:
>
>> pushtd1 is missing.  This patch adds it. OK for 4.4/4.5/trunk?
>
> Actually, it looks that something is missing, so generic part of gcc
> doesn't push the value in TImode. TDmode values are moved as TImode,
> so they should also be pushed via equivalent integer mode (TImode).
>
> Looking at the .md files, TImode push (pushti1) is also missing. I
> will investigate this a bit more.

Attached patch produces substantially better code for the testcase.

The approach with expander generates:

0000000000000000 <foo>:
   0:	55                   	push   %rbp
   1:	48 ba 00 00 00 00 00 	mov    $0x3040000000000000,%rdx
   8:	00 40 30
   b:	0f 28 3d 00 00 00 00 	movaps 0x0(%rip),%xmm7        # 12 <foo+0x12>
  12:	48 89 e5             	mov    %rsp,%rbp
  15:	50                   	push   %rax
  16:	0f 28 f7             	movaps %xmm7,%xmm6
  19:	0f 28 ef             	movaps %xmm7,%xmm5
  1c:	50                   	push   %rax
  1d:	0f 28 e7             	movaps %xmm7,%xmm4
  20:	0f 28 df             	movaps %xmm7,%xmm3
  23:	0f 28 d7             	movaps %xmm7,%xmm2
  26:	0f 28 cf             	movaps %xmm7,%xmm1
  29:	0f 28 c7             	movaps %xmm7,%xmm0
  2c:	48 c7 04 24 00 00 00 	movq   $0x0,(%rsp)
  33:	00
  34:	48 89 54 24 08       	mov    %rdx,0x8(%rsp)
  39:	e8 00 00 00 00       	callq  3e <foo+0x3e>
  3e:	59                   	pop    %rcx
  3f:	5e                   	pop    %rsi
  40:	c9                   	leaveq
  41:	c3                   	retq

and the new approach with a post-reload splitter:

0000000000000000 <foo>:
   0:	55                   	push   %rbp
   1:	48 b8 00 00 00 00 00 	mov    $0x3040000000000000,%rax
   8:	00 40 30
   b:	0f 28 3d 00 00 00 00 	movaps 0x0(%rip),%xmm7        # 12 <foo+0x12>
  12:	48 89 e5             	mov    %rsp,%rbp
  15:	50                   	push   %rax
  16:	0f 28 f7             	movaps %xmm7,%xmm6
  19:	0f 28 ef             	movaps %xmm7,%xmm5
  1c:	6a 00                	pushq  $0x0
  1e:	0f 28 e7             	movaps %xmm7,%xmm4
  21:	0f 28 df             	movaps %xmm7,%xmm3
  24:	0f 28 d7             	movaps %xmm7,%xmm2
  27:	0f 28 cf             	movaps %xmm7,%xmm1
  2a:	0f 28 c7             	movaps %xmm7,%xmm0
  2d:	e8 00 00 00 00       	callq  32 <foo+0x32>
  32:	5a                   	pop    %rdx
  33:	59                   	pop    %rcx
  34:	c9                   	leaveq
  35:	c3                   	retq

So, 2 insns less and 12 bytes smaller code.

2010-10-21  Uros Bizjak  <ubizjak@gmail.com>

	PR target/45946
	* config/i386/i386.md (*pushti2): New insn pattern.
	(pushti2 splitter): New insn splitter.
	(*push<mode>2): Macroize insn pattern from *push{di,ti}2 using
	DWI mode iterator.

testsuite/ChangeLog:

2010-10-21  Uros Bizjak  <ubizjak@gmail.com>
	    H.J. Lu  <hongjiu.lu@intel.com>

	PR target/45946
	* gcc.target/i386/pr45946.c: New test.


Patch is currently in regression test on x86_64-pc-linux-gnu {,-m32}.
Will be committed to SVN mainline and backported to 4.4/4.5 after a
few days.

Uros.