[Bug target/49781] New: Unnecessary lea in x32 mode

[Bug target/49781] New: Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

           Summary: Unnecessary lea in x32 mode
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hjl.tools@gmail.com


[hjl@gnu-6 pr47383]$ cat x.i
static int heap[2*(256 +1+29)+1];
static int heap_len;
static int heap_max;
void 
__attribute__ ((noinline))
foo (int elems)
{
  int n, m;
  int max_code = -1;
  int node = elems;
  heap_len = 0, heap_max = (2*(256 +1+29)+1);
  for (n = 0; n < elems; n++)
    heap[++heap_len] = max_code = n;
  do {
    n = heap[1];
    heap[1] = heap[heap_len--];
    m = heap[1];
    heap[--heap_max] = n;
    heap[--heap_max] = m;
  } while (heap_len >= 2);
}

int
main ()
{
  foo (286);
  return 0;
}
[hjl@gnu-6 pr47383]$ make x.s
/export/build/gnu/gcc-x32/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc-x32/build-x86_64-linux/gcc/ -S -o x.s -mx32 -O2 -fPIC 
x.i
[hjl@gnu-6 pr47383]$ cat x.s
    .file    "x.i"
    .text
    .p2align 4,,15
    .globl    foo
    .type    foo, @function
foo:
.LFB0:
    .cfi_startproc
    xorl    %eax, %eax
    testl    %edi, %edi
    pushq    %rbx
    .cfi_def_cfa_offset 16
    .cfi_offset 3, -16
    leal    4+heap(%rip), %ecx
    jle    .L10
    .p2align 4,,10
    .p2align 3
.L7:
    leal    (%rcx,%rax,4), %edx
    movl    %eax, (%rdx)
    addl    $1, %eax
    cmpl    %edi, %eax
    jne    .L7
    leal    -1(%rdi), %ecx
.L3:
    leal    -1(%rdi), %eax
    movl    $573, %esi
    leal    2296+heap(%rip), %r11d
    leal    2292+heap(%rip), %r10d
    leal    heap(%rip), %ebx
    jmp    .L5
    .p2align 4,,10
    .p2align 3
.L11:
    leal    (%rbx,%rax,4), %edx
    subl    $1, %eax
    movl    (%rdx), %ecx
.L5:
    movl    %eax, %edx
    movl    4+heap(%rip), %r9d
    subl    $2, %esi
    subl    %edi, %edx
    movl    %ecx, 4+heap(%rip)
    sall    $3, %edx
    leal    (%r11,%rdx), %r8d
    addl    %r10d, %edx
    cmpl    $1, %eax
    movl    %r9d, (%r8)
    movl    %ecx, (%rdx)
    jg    .L11
    movl    %eax, heap_len(%rip)
    movl    %esi, heap_max(%rip)
    popq    %rbx
    .cfi_remember_state
    .cfi_def_cfa_offset 8
    ret
.L10:
    .cfi_restore_state
    movl    heap(%rip), %ecx
    xorl    %edi, %edi
    jmp    .L3
    .cfi_endproc
.LFE0:
    .size    foo, .-foo
    .section    .text.startup,"ax",@progbits
    .p2align 4,,15
    .globl    main
    .type    main, @function
main:
.LFB1:
    .cfi_startproc
    subq    $8, %rsp
    .cfi_def_cfa_offset 16
    movl    $286, %edi
    call    foo@PLT
    xorl    %eax, %eax
    addq    $8, %rsp
    .cfi_def_cfa_offset 8
    ret
    .cfi_endproc
.LFE1:
    .size    main, .-main
    .local    heap_len
    .comm    heap_len,4,4
    .local    heap_max
    .comm    heap_max,4,4
    .local    heap
    .comm    heap,2292,32
    .ident    "GCC: (GNU) 4.7.0 20110718 (experimental)"
    .section    .note.GNU-stack,"",@progbits
[hjl@gnu-6 pr47383]$ 

Many lea insns can be combined with the load/store insn followed.

[Bug target/49781] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #1 from Uros Bizjak <ubizjak at gmail dot com> 2011-07-19 06:45:56 UTC ---
(In reply to comment #0)

> Many lea insns can be combined with the load/store insn followed.

No, these are leal insns, they have inherent zero_extend from SImode and they
clear upper 32bits before address is passed to load/store insn.

This is what you got on x32.

[Bug target/49781] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> 2011-07-19 13:06:19 UTC ---
(In reply to comment #1)
> (In reply to comment #0)
> 
> > Many lea insns can be combined with the load/store insn followed.
> 
> No, these are leal insns, they have inherent zero_extend from SImode and they
> clear upper 32bits before address is passed to load/store insn.

There are 2 ways to address this:

1. Use addr_32 prefix.
2. If address wrap-around is ignored, we can use the address as is.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #3 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 07:00:12 UTC ---
Basically, we should allow ZERO_EXTEND in address:

Trying 117 -> 118:
Failed to match this instruction:
(set (mem:SI (zero_extend:DI (plus:SI (mult:SI (reg/v:SI 150 [ n ])
                    (const_int 4 [0x4]))
                (reg/f:SI 189))) [2 MEM[symbol: heap, index: D.2768_17, offset:
4B]+0 S4 A32])
    (reg/v:SI 150 [ n ]))

We should emit addr32 in this case.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 10:24:17 UTC ---
(In reply to comment #0)

> Many lea insns can be combined with the load/store insn followed.

I have a patch that generates addr32 prefix. The result:

    .file    "pr49781.c"
    .text
    .p2align 4,,15
    .globl    foo
    .type    foo, @function
foo:
.LFB0:
    .cfi_startproc
    testl    %edi, %edi
    jle    .L9
    subl    $1, %edi
    leaq    4+heap(%rip), %rdx
    xorl    %eax, %eax
    mov    %edi, %ecx
    addq    $1, %rcx
    .p2align 4,,10
    .p2align 3
.L4:
    movl    %eax, (%rdx,%rax,4)
    addq    $1, %rax
    cmpq    %rcx, %rax
    jne    .L4
    movl    %edi, %ecx
.L3:
    leaq    2288+heap(%rip), %rax
    movl    $571, %edx
    jmp    .L6
    .p2align 4,,10
    .p2align 3
.L10:
    leaq    heap(%rip), %rcx
    movslq    %edi, %rsi
    subl    $1, %edi
    movl    (%rcx,%rsi,4), %ecx
.L6:
    movl    4+heap(%rip), %esi
    movl    %ecx, 4+heap(%rip)
    movl    %ecx, -4(%rax)
    movl    %esi, (%rax)
    movl    %edx, %esi
    subq    $8, %rax
    subl    $2, %edx
    cmpl    $1, %edi
    jg    .L10
    movl    %edi, heap_len(%rip)
    movl    %esi, heap_max(%rip)
    ret
.L9:
    movl    heap(%rip), %ecx
    movl    $-1, %edi
    jmp    .L3
    .cfi_endproc
.LFE0:
    .size    foo, .-foo
    .section    .text.startup,"ax",@progbits
    .p2align 4,,15
    .globl    main
    .type    main, @function
main:
.LFB1:
    .cfi_startproc
    subq    $8, %rsp
    .cfi_def_cfa_offset 16
    movl    $286, %edi
    call    foo@PLT
    xorl    %eax, %eax
    addq    $8, %rsp
    .cfi_def_cfa_offset 8
    ret
    .cfi_endproc
.LFE1:
    .size    main, .-main
    .local    heap_len
    .comm    heap_len,4,4
    .local    heap_max
    .comm    heap_max,4,4
    .local    heap
    .comm    heap,2292,32
    .ident    "GCC: (GNU) 4.7.0 20110801 (experimental) [trunk revision
176998]"
    .section    .note.GNU-stack,"",@progbits

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #5 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 10:27:35 UTC ---
(In reply to comment #4)
> (In reply to comment #0)
> 
> > Many lea insns can be combined with the load/store insn followed.
> 
> I have a patch that generates addr32 prefix. The result:

Er, wrong one (that was -m64). This one is -mx32:

    .file    "pr49781.c"
    .text
    .p2align 4,,15
    .globl    foo
    .type    foo, @function
foo:
.LFB0:
    .cfi_startproc
    xorl    %eax, %eax
    testl    %edi, %edi
    leal    4+heap(%rip), %edx
    jle    .L10
    .p2align 4,,10
    .p2align 3
.L7:
    movl    %eax, (%edx,%eax,4)
    addl    $1, %eax
    cmpl    %edi, %eax
    jne    .L7
    leal    -1(%rdi), %eax
    movl    %eax, %ecx
.L3:
    movl    $573, %esi
    leal    2296+heap(%rip), %r10d
    leal    2292+heap(%rip), %r9d
    leal    heap(%rip), %r11d
    jmp    .L5
    .p2align 4,,10
    .p2align 3
.L11:
    movl    (%r11d,%eax,4), %ecx
    subl    $1, %eax
.L5:
    movl    4+heap(%rip), %r8d
    movl    %eax, %edx
    subl    $2, %esi
    subl    %edi, %edx
    movl    %ecx, 4+heap(%rip)
    sall    $3, %edx
    cmpl    $1, %eax
    movl    %r8d, (%r10d,%edx)
    movl    %ecx, (%r9d,%edx)
    jg    .L11
    movl    %eax, heap_len(%rip)
    movl    %esi, heap_max(%rip)
    ret
.L10:
    movl    heap(%rip), %ecx
    xorl    %edi, %edi
    movl    $-1, %eax
    jmp    .L3
    .cfi_endproc
.LFE0:
    .size    foo, .-foo
    .section    .text.startup,"ax",@progbits
    .p2align 4,,15
    .globl    main
    .type    main, @function
main:
.LFB1:
    .cfi_startproc
    subq    $8, %rsp
    .cfi_def_cfa_offset 16
    movl    $286, %edi
    call    foo@PLT
    xorl    %eax, %eax
    addq    $8, %rsp
    .cfi_def_cfa_offset 8
    ret
    .cfi_endproc
.LFE1:
    .size    main, .-main
    .local    heap_len
    .comm    heap_len,4,4
    .local    heap_max
    .comm    heap_max,4,4
    .local    heap
    .comm    heap,2292,32
    .ident    "GCC: (GNU) 4.7.0 20110803 (experimental) [trunk revision
177229]"
    .section    .note.GNU-stack,"",@progbits

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #6 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 10:30:39 UTC ---
Created attachment 24899
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24899
Proposed patch that exploits addr32.

H.J., can you please test this patch on mx32.

The patch bootstraps and regression tests OK on x86_64-pc-linux-gnu {,-m32}.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #7 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 13:26:20 UTC ---
(In reply to comment #6)
> Created attachment 24899 [details]
> Proposed patch that exploits addr32.
> 
> H.J., can you please test this patch on mx32.
> 
> The patch bootstraps and regression tests OK on x86_64-pc-linux-gnu {,-m32}.

I started it now. Thanks.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #8 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 14:10:25 UTC ---
(In reply to comment #6)
> Created attachment 24899 [details]
> Proposed patch that exploits addr32.
> 
> H.J., can you please test this patch on mx32.
> 
> The patch bootstraps and regression tests OK on x86_64-pc-linux-gnu {,-m32}.

It failed the testcase for PR 47744, which I just checked in:

[hjl@gnu-33 ilp32-24]$
/export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/ -S -o x.s -mx32 -O3
-std=gnu99
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c: In
function \u2018matmul_i16\u2019:
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
error: insn does not satisfy its constraints:
(insn 146 66 67 4 (set (reg:TI 0 ax)
        (mem:TI (zero_extend:DI (plus:SI (reg:SI 4 si [orig:119 ivtmp.30 ]
[119])
                    (reg:SI 5 di [orig:102 dest_y ] [102]))) [6 MEM[base:
dest_y_18, index: ivtmp.30_63, offset: 0B]+0 S16 A128]))
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:34 60
{*movti_internal_rex64}
     (nil))
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
internal compiler error: in reload_cse_simplify_operands, at postreload.c:403
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
[hjl@gnu-33 ilp32-24]$

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #9 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 14:45:40 UTC ---
(In reply to comment #8)

> > Created attachment 24899 [details]
> > Proposed patch that exploits addr32.
> > 
> > H.J., can you please test this patch on mx32.
> > 
> > The patch bootstraps and regression tests OK on x86_64-pc-linux-gnu {,-m32}.
> 
> It failed the testcase for PR 47744, which I just checked in:
> 
> [hjl@gnu-33 ilp32-24]$
> /export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/xgcc
> -B/export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/ -S -o x.s -mx32 -O3
> -std=gnu99
> /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c
> /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c: In
> function \u2018matmul_i16\u2019:
> /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
> error: insn does not satisfy its constraints:
> (insn 146 66 67 4 (set (reg:TI 0 ax)
>         (mem:TI (zero_extend:DI (plus:SI (reg:SI 4 si [orig:119 ivtmp.30 ]
> [119])
>                     (reg:SI 5 di [orig:102 dest_y ] [102]))) [6 MEM[base:
> dest_y_18, index: ivtmp.30_63, offset: 0B]+0 S16 A128]))
> /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:34 60
> {*movti_internal_rex64}
>      (nil))
> /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
> internal compiler error: in reload_cse_simplify_operands, at postreload.c:403
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <http://gcc.gnu.org/bugs.html> for instructions.
> [hjl@gnu-33 ilp32-24]$

Hm, offsetable operand ...

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #10 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 15:01:10 UTC ---
(In reply to comment #9)
> (In reply to comment #8)
> 
> > > Created attachment 24899 [details]
> > > Proposed patch that exploits addr32.
> > > 
> > > H.J., can you please test this patch on mx32.
> > > 
> > > The patch bootstraps and regression tests OK on x86_64-pc-linux-gnu {,-m32}.
> > 
> > It failed the testcase for PR 47744, which I just checked in:
> > 
> > [hjl@gnu-33 ilp32-24]$
> > /export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/xgcc
> > -B/export/build/gnu/gcc-x32-test/build-x86_64-linux/gcc/ -S -o x.s -mx32 -O3
> > -std=gnu99
> > /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c
> > /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c: In
> > function \u2018matmul_i16\u2019:
> > /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
> > error: insn does not satisfy its constraints:
> > (insn 146 66 67 4 (set (reg:TI 0 ax)
> >         (mem:TI (zero_extend:DI (plus:SI (reg:SI 4 si [orig:119 ivtmp.30 ]
> > [119])
> >                     (reg:SI 5 di [orig:102 dest_y ] [102]))) [6 MEM[base:
> > dest_y_18, index: ivtmp.30_63, offset: 0B]+0 S16 A128]))
> > /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:34 60
> > {*movti_internal_rex64}
> >      (nil))
> > /export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/torture/pr47744-2.c:40:1:
> > internal compiler error: in reload_cse_simplify_operands, at postreload.c:403
> > Please submit a full bug report,
> > with preprocessed source if appropriate.
> > See <http://gcc.gnu.org/bugs.html> for instructions.
> > [hjl@gnu-33 ilp32-24]$
> 
> Hm, offsetable operand ...

This additional patch prevents zero_extend when we deal with
wider-than-word-size moves.  These moves need offsetable_operand, which
zero_extend (...) isn't.

Index: i386.c
===================================================================
--- i386.c    (revision 177281)
+++ i386.c    (working copy)
@@ -11681,6 +11689,10 @@ ix86_legitimate_address_p (enum machine_
   rtx base, index, disp;
   HOST_WIDE_INT scale;

+  if (GET_CODE (addr) == ZERO_EXTEND
+      && GET_MODE_SIZE (mode) > UNITS_PER_WORD)
+      return false;
+
   if (ix86_decompose_address (addr, &parts) <= 0)
     /* Decomposition failed.  */
     return false;

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #11 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 15:44:59 UTC ---
(In reply to comment #10)
> This additional patch prevents zero_extend when we deal with
> wider-than-word-size moves.  These moves need offsetable_operand, which
> zero_extend (...) isn't.
> 
> Index: i386.c
> ===================================================================
> --- i386.c    (revision 177281)
> +++ i386.c    (working copy)
> @@ -11681,6 +11689,10 @@ ix86_legitimate_address_p (enum machine_
>    rtx base, index, disp;
>    HOST_WIDE_INT scale;
> 
> +  if (GET_CODE (addr) == ZERO_EXTEND
> +      && GET_MODE_SIZE (mode) > UNITS_PER_WORD)
> +      return false;
> +
>    if (ix86_decompose_address (addr, &parts) <= 0)
>      /* Decomposition failed.  */
>      return false;

gcc.dg/torture/pr47744-2.c compiled with

-mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops

generates codes  like

    leal    (%rax,%r9), %r12d
    leal    (%rax,%rdi), %r10d
    mov    %r12d, %edx
    movq    (%r12d), %rbp
    movq    8(%rdx), %rdx
    movq    (%r12d), %rax

Many leal aren't necessary.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #12 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-03 16:08:47 UTC ---
(In reply to comment #11)

> gcc.dg/torture/pr47744-2.c compiled with
> 
> -mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops
> 
> generates codes  like
> 
>     leal    (%rax,%r9), %r12d
>     leal    (%rax,%rdi), %r10d
>     mov    %r12d, %edx
>     movq    (%r12d), %rbp
>     movq    8(%rdx), %rdx
>     movq    (%r12d), %rax
> 
> Many leal aren't necessary.

This is the tradeof for using offsetable address for DWI operands.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #13 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 16:18:45 UTC ---
(In reply to comment #10)
> 
> This additional patch prevents zero_extend when we deal with
> wider-than-word-size moves.  These moves need offsetable_operand, which
> zero_extend (...) isn't.
> 
> Index: i386.c
> ===================================================================
> --- i386.c    (revision 177281)
> +++ i386.c    (working copy)
> @@ -11681,6 +11689,10 @@ ix86_legitimate_address_p (enum machine_
>    rtx base, index, disp;
>    HOST_WIDE_INT scale;
> 
> +  if (GET_CODE (addr) == ZERO_EXTEND
> +      && GET_MODE_SIZE (mode) > UNITS_PER_WORD)
> +      return false;
> +
>    if (ix86_decompose_address (addr, &parts) <= 0)
>      /* Decomposition failed.  */
>      return false;


Doesn't work. I got

FAIL: gcc.dg/graphite/pr35356-2.c (internal compiler error)
FAIL: gcc.dg/graphite/pr35356-2.c (test for excess errors)
FAIL: libgomp.fortran/omp_parse4.f90  -Os  (internal compiler error)
FAIL: libgomp.fortran/omp_parse4.f90  -Os  (test for excess errors)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (internal compiler error)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (test for excess errors)

so far:

spawn -ignore SIGHUP /export/build/gnu/gcc-x32/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc-x32/build-x86_64-linux/gcc/
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/graphite/pr35356-2.c -O2
-fgraphite-identity -fdump-tree-graphite-all -S -mx32 -o pr35356-2.s^M
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/graphite/pr35356-2.c: In
function 'foo':^M
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.dg/graphite/pr35356-2.c:17:1:
internal compiler error: Segmentation fault^M
Please submit a full bug report,^M
with preprocessed source if appropriate.^M
See <http://gcc.gnu.org/bugs.html> for instructions.^M
compiler exited with status 1

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #14 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 16:47:09 UTC ---
Created attachment 24907
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24907
A testcase

[hjl@gnu-33 delta]$
/export/build/gnu/gcc-x32-test/release/usr/gcc-4.7.0-x32/bin/gcc -mx32  
-std=gnu99 -fgnu89-inline -O2    -S testcase.c
....
testcase.c:634:2: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #15 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-03 16:49:12 UTC ---
(In reply to comment #13)
> (In reply to comment #10)
> > 
> > This additional patch prevents zero_extend when we deal with
> > wider-than-word-size moves.  These moves need offsetable_operand, which
> > zero_extend (...) isn't.
> > 
> > Index: i386.c
> > ===================================================================
> > --- i386.c    (revision 177281)
> > +++ i386.c    (working copy)
> > @@ -11681,6 +11689,10 @@ ix86_legitimate_address_p (enum machine_
> >    rtx base, index, disp;
> >    HOST_WIDE_INT scale;
> > 
> > +  if (GET_CODE (addr) == ZERO_EXTEND
> > +      && GET_MODE_SIZE (mode) > UNITS_PER_WORD)
> > +      return false;
> > +
> >    if (ix86_decompose_address (addr, &parts) <= 0)
> >      /* Decomposition failed.  */
> >      return false;
> 
> 
> Doesn't work. I got
> 
> FAIL: gcc.dg/graphite/pr35356-2.c (internal compiler error)
> FAIL: gcc.dg/graphite/pr35356-2.c (test for excess errors)
> FAIL: libgomp.fortran/omp_parse4.f90  -Os  (internal compiler error)
> FAIL: libgomp.fortran/omp_parse4.f90  -Os  (test for excess errors)
> FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (internal compiler error)
> FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (test for excess errors)

Total regressions:

FAIL: gcc.c-torture/compile/20000717-1.c  -Os  (internal compiler error)
FAIL: gcc.c-torture/compile/20000717-1.c  -Os  (test for excess errors)
FAIL: gcc.c-torture/execute/builtins/strcat.c compilation,  -O1  (internal
compiler error)
FAIL: gcc.c-torture/unsorted/xdi.c,  -O1   (internal compiler error)
FAIL: gcc.dg/graphite/pr35356-2.c (internal compiler error)
FAIL: gcc.dg/graphite/pr35356-2.c (test for excess errors)
FAIL: gcc.dg/torture/pr45636.c  -O1  (internal compiler error)
FAIL: gcc.dg/torture/pr45636.c  -O1  (test for excess errors)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -Os  (internal compiler error)
FAIL: gfortran.dg/alloc_comp_assign_4.f90  -Os  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_1.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_1.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_6.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_6.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_8.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_8.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/character_array_constructor_1.f90  -O1  (internal compiler
error)
FAIL: gfortran.dg/character_array_constructor_1.f90  -O1  (test for excess
errors)
FAIL: gfortran.dg/char_expr_3.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/char_expr_3.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/char_length_3.f90  -O  (internal compiler error)
FAIL: gfortran.dg/char_length_3.f90  -O  (test for excess errors)
FAIL: gfortran.dg/extends_1.f03  -O1  (internal compiler error)
FAIL: gfortran.dg/extends_1.f03  -O1  (test for excess errors)
FAIL: gfortran.dg/g77/7388.f  -O1  (internal compiler error)
FAIL: gfortran.dg/g77/7388.f  -O1  (test for excess errors)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (internal compiler error)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (test for excess errors)
FAIL: gfortran.dg/namelist_66.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/namelist_66.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -fomit-frame-pointer  (internal compiler
error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -fomit-frame-pointer  (test for excess
errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -g  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -g  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -fomit-frame-pointer  (internal
compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -fomit-frame-pointer  (test for
excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -g  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -g  (test for excess errors)
FAIL: gfortran.fortran-torture/execute/where_11.f90,  -Os  (internal compiler
error)
FAIL: gfortran.fortran-torture/execute/where_2.f90,  -Os  (internal compiler
error)
FAIL: libgomp.fortran/omp_parse4.f90  -Os  (internal compiler error)
FAIL: libgomp.fortran/omp_parse4.f90  -Os  (test for excess errors)

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #16 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-04 18:49:32 UTC ---
(In reply to comment #14)

> /export/build/gnu/gcc-x32-test/release/usr/gcc-4.7.0-x32/bin/gcc -mx32  
> -std=gnu99 -fgnu89-inline -O2    -S testcase.c
> ....
> testcase.c:634:2: internal compiler error: Segmentation fault
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <http://gcc.gnu.org/bugs.html> for instructions.

Sometimes, a subreg is returned as base or index register.  This patch strips
subregs and fixes ICE (I wonder, why the ICE didn't show earlier):

Index: i386.c
===================================================================
--- i386.c    (revision 177411)
+++ i386.c    (working copy)
@@ -14088,6 +14101,20 @@

   gcc_assert (ok);

+  if (parts.base && GET_CODE (parts.base) == SUBREG)
+    {
+      rtx tmp = SUBREG_REG (parts.base);
+      parts.base = simplify_subreg (GET_MODE (parts.base),
+                    tmp, GET_MODE (tmp), 0);
+    }
+
+  if (parts.index && GET_CODE (parts.index) == SUBREG)
+    {
+      rtx tmp = SUBREG_REG (parts.index);
+      parts.index = simplify_subreg (GET_MODE (parts.index),
+                     tmp, GET_MODE (tmp), 0);
+    }
+
   base = parts.base;
   index = parts.index;
   disp = parts.disp;

However, I think we should hanlde plus (zero_extend (...addr...)(const_int))
offsetable operands, too. Otherwise there will be no benefit for SSE
operands...

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #24899|0                           |1
        is obsolete|                            |

--- Comment #17 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-04 18:53:06 UTC ---
Created attachment 24918
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24918
WIP patch that exploits addr32.

The patch ATM limits addr32 optimization to MODE_SIZE (op) <= WORD_SIZE due to
inability to handle offsetable operands.

But please test it on x32 anyway.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #18 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-04 19:40:38 UTC ---
(In reply to comment #17)
> Created attachment 24918 [details]
> WIP patch that exploits addr32.
> 
> The patch ATM limits addr32 optimization to MODE_SIZE (op) <= WORD_SIZE due to
> inability to handle offsetable operands.
> 
> But please test it on x32 anyway.

I am testing it now.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #19 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-04 19:59:59 UTC ---
(In reply to comment #18)
> (In reply to comment #17)
> > Created attachment 24918 [details]
> > WIP patch that exploits addr32.
> > 
> > The patch ATM limits addr32 optimization to MODE_SIZE (op) <= WORD_SIZE due to
> > inability to handle offsetable operands.
> > 
> > But please test it on x32 anyway.
> 
> I am testing it now.

Uh, I forgot to add this code to predicates.md, aligned_operand (and perhaps
cmpxchg8b_pic_memory_operand)...

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #20 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-04 20:09:48 UTC ---
This is the missing part of predicates.md:

Index: predicates.md
===================================================================
--- predicates.md    (revision 177411)
+++ predicates.md    (working copy)
@@ -840,6 +844,12 @@
   ok = ix86_decompose_address (op, &parts);
   gcc_assert (ok);

+  if (GET_CODE (parts.base) == SUBREG)
+    parts.base = SUBREG_REG (parts.base);
+
+  if (GET_CODE (parts.index) == SUBREG)
+    parts.index = SUBREG_REG (parts.index);
+
   /* Look for some component that isn't known to be aligned.  */
   if (parts.index)
     {
@@ -903,6 +913,10 @@

   ok = ix86_decompose_address (XEXP (op, 0), &parts);
   gcc_assert (ok);
+
+  if (GET_CODE (parts.base) == SUBREG)
+    parts.base = SUBREG_REG (parts.base);
+
   if (parts.base == NULL_RTX
       || parts.base == arg_pointer_rtx
       || parts.base == frame_pointer_rtx
@@ -910,6 +924,9 @@
       || parts.base == stack_pointer_rtx)
     return true;

+  if (GET_CODE (parts.index) == SUBREG)
+    parts.index = SUBREG_REG (parts.index);
+
   if (parts.index == NULL_RTX
       || parts.index == arg_pointer_rtx
       || parts.index == frame_pointer_rtx

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #21 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-04 21:00:11 UTC ---
(In reply to comment #20)
> This is the missing part of predicates.md:
> 

It failed on 64bit bootstrap:

[hjl@gnu-33 libgcc]$ cat /tmp/doit 
/export/build/gnu/gcc-x32-test/build-x86_64-linux/./gcc/xgcc
-B/export/build/gnu/gcc-x32-test/build-x86_64-linux/./gcc/
-B/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/bin/
-B/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/lib/ -isystem
/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/include -isystem
/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/sys-include    -O2 -g -O2  -O2 -g
-DIN_GCC   -W -Wall -Wwrite-strings -Wcast-qual -Wstrict-prototypes
-Wmissing-prototypes -Wold-style-definition  -isystem ./include  -fPIC -g
-DHAVE_GTHR_DEFAULT -DIN_LIBGCC2 -fbuilding-libgcc -fno-stack-protector   -I.
-I. -I../.././gcc -I/export/gnu/import/git/gcc-x32/libgcc
-I/export/gnu/import/git/gcc-x32/libgcc/.
-I/export/gnu/import/git/gcc-x32/libgcc/../gcc
-I/export/gnu/import/git/gcc-x32/libgcc/../include
-I/export/gnu/import/git/gcc-x32/libgcc/config/libbid
-DENABLE_DECIMAL_BID_FORMAT -DHAVE_CC_TLS  -DUSE_TLS -o bid64_div.o -MT
bid64_div.o -MD -MP -MF bid64_div.dep -c
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c
[hjl@gnu-33 libgcc]$
/export/build/gnu/gcc-x32-test/build-x86_64-linux/./gcc/xgcc
-B/export/build/gnu/gcc-x32-test/build-x86_64-linux/./gcc/
-B/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/bin/
-B/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/lib/ -isystem
/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/include -isystem
/usr/gcc-4.7.0-x32/x86_64-unknown-linux-gnu/sys-include    -O2 -g -O2  -O2 -g
-DIN_GCC   -W -Wall -Wwrite-strings -Wcast-qual -Wstrict-prototypes
-Wmissing-prototypes -Wold-style-definition  -isystem ./include  -fPIC -g
-DHAVE_GTHR_DEFAULT -DIN_LIBGCC2 -fbuilding-libgcc -fno-stack-protector   -I.
-I. -I../.././gcc -I/export/gnu/import/git/gcc-x32/libgcc
-I/export/gnu/import/git/gcc-x32/libgcc/.
-I/export/gnu/import/git/gcc-x32/libgcc/../gcc
-I/export/gnu/import/git/gcc-x32/libgcc/../include
-I/export/gnu/import/git/gcc-x32/libgcc/config/libbid
-DENABLE_DECIMAL_BID_FORMAT -DHAVE_CC_TLS  -DUSE_TLS -o bid64_div.o -MT
bid64_div.o -MD -MP -MF bid64_div.dep -c
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c: In function
‘__bid64dq_div’:
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c:523:51:
warning: variable ‘Ql’ set but not used [-Wunused-but-set-variable]
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c: In function
‘__bid64qd_div’:
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c:937:51:
warning: variable ‘Ql’ set but not used [-Wunused-but-set-variable]
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c: In function
‘__bid64qq_div’:
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c:1374:51:
warning: variable ‘Ql’ set but not used [-Wunused-but-set-variable]
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c: In function
‘__bid64_div’:
/export/gnu/import/git/gcc-x32/libgcc/config/libbid/bid64_div.c:516:1: internal
compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #22 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-04 21:03:07 UTC ---
I am testing:

diff --git a/gcc/config/i386/predicates.md b/gcc/config/i386/predicates.md
index ae1fa74..373e74b 100644
--- a/gcc/config/i386/predicates.md
+++ b/gcc/config/i386/predicates.md
@@ -844,10 +844,10 @@
   ok = ix86_decompose_address (op, &parts);
   gcc_assert (ok);

-  if (GET_CODE (parts.base) == SUBREG)
+  if (parts.base && GET_CODE (parts.base) == SUBREG)
     parts.base = SUBREG_REG (parts.base);

-  if (GET_CODE (parts.index) == SUBREG)
+  if (parts.index && GET_CODE (parts.index) == SUBREG)
     parts.index = SUBREG_REG (parts.index);

   /* Look for some component that isn't known to be aligned.  */
@@ -914,7 +914,7 @@
   ok = ix86_decompose_address (XEXP (op, 0), &parts);
   gcc_assert (ok);

-  if (GET_CODE (parts.base) == SUBREG)
+  if (parts.base && GET_CODE (parts.base) == SUBREG)
     parts.base = SUBREG_REG (parts.base);

   if (parts.base == NULL_RTX
@@ -924,7 +924,7 @@
       || parts.base == stack_pointer_rtx)
     return true;

-  if (GET_CODE (parts.index) == SUBREG)
+  if (parts.index && GET_CODE (parts.index) == SUBREG)
     parts.index = SUBREG_REG (parts.index);

   if (parts.index == NULL_RTX

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #23 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-04 22:24:58 UTC ---
I still got

FAIL: gcc.c-torture/execute/builtins/strcat.c compilation,  -O1  (internal
compiler error)
FAIL: gcc.c-torture/unsorted/xdi.c,  -O1   (internal compiler error)
FAIL: gcc.dg/torture/pr45636.c  -O1  (internal compiler error)
FAIL: gcc.dg/torture/pr45636.c  -O1  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_1.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_1.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_6.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_6.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/bounds_check_array_ctor_8.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/bounds_check_array_ctor_8.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/character_array_constructor_1.f90  -O1  (internal compiler
error)
FAIL: gfortran.dg/character_array_constructor_1.f90  -O1  (test for excess
errors)
FAIL: gfortran.dg/char_expr_3.f90  -O1  (internal compiler error)
FAIL: gfortran.dg/char_expr_3.f90  -O1  (test for excess errors)
FAIL: gfortran.dg/char_length_3.f90  -O  (internal compiler error)
FAIL: gfortran.dg/char_length_3.f90  -O  (test for excess errors)
FAIL: gfortran.dg/extends_1.f03  -O1  (internal compiler error)
FAIL: gfortran.dg/extends_1.f03  -O1  (test for excess errors)
FAIL: gfortran.dg/g77/7388.f  -O1  (internal compiler error)
FAIL: gfortran.dg/g77/7388.f  -O1  (test for excess errors)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (internal compiler error)
FAIL: gfortran.dg/gomp/crayptr5.f90  -O  (test for excess errors)
FAIL: gfortran.dg/namelist_66.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/namelist_66.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -fomit-frame-pointer  (internal compiler
error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -fomit-frame-pointer  (test for excess
errors)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -g  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_22.f90  -O3 -g  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O2  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O2  (test for excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -fomit-frame-pointer  (internal
compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -fomit-frame-pointer  (test for
excess errors)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -g  (internal compiler error)
FAIL: gfortran.dg/proc_ptr_comp_12.f90  -O3 -g  (test for excess errors)

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #24 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-05 06:25:55 UTC ---
(In reply to comment #23)
> I still got
> 
> FAIL: gcc.c-torture/execute/builtins/strcat.c compilation,  -O1  (internal
> compiler error)

movdi_internal_rex64 has wrong constraint for operand 0 for moving 64bit
immediate. This move needs offsetable memory, so "!o" instead of "!m" in
alternative 4.

Nice, how many problems did -mx32 uncovered.  I will prepare a patch for
mainline that fixes all these inconsistencies first.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #25 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-05 19:01:44 UTC ---
New revision of patch at [1].

[1] http://gcc.gnu.org/ml/gcc-patches/2011-08/msg00642.html

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #26 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-05 21:58:21 UTC ---
(In reply to comment #25)
> New revision of patch at [1].
> 
> [1] http://gcc.gnu.org/ml/gcc-patches/2011-08/msg00642.html

It failed gcc.dg/torture/pr47744-2.c the same way as comment 8.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #27 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-05 22:01:31 UTC ---
(In reply to comment #26)

> > [1] http://gcc.gnu.org/ml/gcc-patches/2011-08/msg00642.html
> 
> It failed gcc.dg/torture/pr47744-2.c the same way as comment 8.

Expected, please see the above link for explanation.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #28 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-07 07:31:06 UTC ---
Reduced testcase:

--cut here--
void test (__int128 *array, int idx, int off)
{
  __int128 *dest = &array [idx];

  dest[0] += 1;
  dest[off] = 0;
}
--cut here--

$ cc1 -O2 -mx32 -quiet t.c
t.c: In function ‘test’:
t.c:7:1: error: insn does not satisfy its constraints:
(insn 20 10 11 2 (set (reg:TI 2 cx)
        (mem:TI (zero_extend:DI (reg:SI 4 si [71])) [2 *dest_5+0 S16 A128]))
t.c:5 60 {*movti_internal_rex64}
     (nil))
t.c:7:1: internal compiler error: in reload_cse_simplify_operands, at
postreload.c:403
Please submit a full bug report,
...

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #24918|0                           |1
        is obsolete|                            |
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011.08.07 12:53:50
         AssignedTo|unassigned at gcc dot       |ubizjak at gmail dot com
                   |gnu.org                     |
     Ever Confirmed|0                           |1

--- Comment #29 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-07 12:53:50 UTC ---
Created attachment 24938
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24938
WIP patch that exploits addr32.

New version of patch for testing. Survives bootstrap + regtest on
x86_64-pc-linux-gnu.

2011-08-07  Uros Bizjak  <ubizjak@gmail.com>

    PR target/49781
    * config/i386/i386.c (ix86_decompose_address): Allow zero-extended
    SImode addresses.
    (ix86_print_operand_address): Handle zero-extended addresses.
    (memory_address_length): Add length of addr32 prefix for
    zero-extended addresses.
    (ix86_secondary_reload): Handle moves to/from double-word general
    registers from/to zero-extended addresses.
    * config/i386/predicates.md (lea_address_operand): Reject
    zero-extended operands.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #30 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-07 13:09:23 UTC ---
(In reply to comment #29)
> Created attachment 24938 [details]
> WIP patch that exploits addr32.

BTW: This patch also fixes following FAIL in [1]:

FAIL: gcc.target/i386/pr39543-3.c (test for excess errors)

and exposes interesting problem in combine in 

FAIL: gcc.target/i386/pr43766.c scan-assembler-not lea[lq]?[ \\t]

where combine converts zero_extended address on-the-fly to an address involving
and.  We can also recognize this as a valid address.

Trying 9 -> 10:
Failed to match this instruction:
(prefetch (and:DI (subreg:DI (plus:SI (ashift:SI (reg/v:SI 63 [ i ])
                    (const_int 2 [0x2]))
                (subreg:SI (reg/v/f:DI 62 [ a ]) 0)) 0)
        (const_int 4294967295 [0xffffffff]))
    (const_int 0 [0])
    (const_int 3 [0x3]))

[1] http://gcc.gnu.org/ml/gcc-testresults/2011-08/msg00601.html

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #31 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-07 13:50:17 UTC ---
(In reply to comment #29)
> Created attachment 24938 [details]
> WIP patch that exploits addr32.
> 
> New version of patch for testing. Survives bootstrap + regtest on
> x86_64-pc-linux-gnu.

Oh, I forgot to add MEM_P ..., so please change the condition to:

  if (TARGET_64BIT
&& MEM_P (x)

      && GET_MODE_SIZE (mode) > UNITS_PER_WORD
      && rclass == GENERAL_REGS
      && !offsettable_memref_p (x))

Patch still works OK.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |UNCONFIRMED
     Ever Confirmed|1                           |0

--- Comment #32 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-07 20:23:13 UTC ---
(In reply to comment #31)
> (In reply to comment #29)
> > Created attachment 24938 [details]
> > WIP patch that exploits addr32.
> > 
> > New version of patch for testing. Survives bootstrap + regtest on
> > x86_64-pc-linux-gnu.
> 
> Oh, I forgot to add MEM_P ..., so please change the condition to:
> 
>   if (TARGET_64BIT
> && MEM_P (x)
> 
>       && GET_MODE_SIZE (mode) > UNITS_PER_WORD
>       && rclass == GENERAL_REGS
>       && !offsettable_memref_p (x))
> 
> Patch still works OK.

I am testing it now.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #33 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-08 13:34:47 UTC ---
(In reply to comment #32)
> (In reply to comment #31)
> > (In reply to comment #29)
> > > Created attachment 24938 [details]
> > > WIP patch that exploits addr32.
> > > 
> > > New version of patch for testing. Survives bootstrap + regtest on
> > > x86_64-pc-linux-gnu.
> > 
> > Oh, I forgot to add MEM_P ..., so please change the condition to:
> > 
> >   if (TARGET_64BIT
> > && MEM_P (x)
> > 
> >       && GET_MODE_SIZE (mode) > UNITS_PER_WORD
> >       && rclass == GENERAL_REGS
> >       && !offsettable_memref_p (x))
> > 
> > Patch still works OK.
> 
> I am testing it now.

It works.  There are no regressions in GCC, glibc and SPEC CPU
2000/2006.  Thanks.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[uros at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #34 from uros at gcc dot gnu.org 2011-08-08 14:59:22 UTC ---
Author: uros
Date: Mon Aug  8 14:59:19 2011
New Revision: 177566

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177566
Log:
    PR target/49781
    * config/i386/i386.c (ix86_decompose_address): Allow zero-extended
    SImode addresses.
    (ix86_print_operand_address): Handle zero-extended addresses.
    (memory_address_length): Add length of addr32 prefix for
    zero-extended addresses.
    (ix86_secondary_reload): Handle moves to/from double-word general
    registers from/to zero-extended addresses.
    * config/i386/predicates.md (lea_address_operand): Reject
    zero-extended operands.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/config/i386/predicates.md

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #35 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-08 16:28:58 UTC ---
It works much better now. But gcc.dg/torture/pr47744-2.c compiled with

-mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops

still generates those leal:

    leal    (%rsi,%r9), %ebp
    leal    (%r10,%r9), %r12d
        leal    (%r10,%r9), %r12d

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #36 from hjl at gcc dot gnu.org <hjl at gcc dot gnu.org> 2011-08-08 16:33:10 UTC ---
Author: hjl
Date: Mon Aug  8 16:33:06 2011
New Revision: 177569

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177569
Log:
Add a testcase for PR target/49781.

2011-08-08  H.J. Lu  <hongjiu.lu@intel.com>

    PR target/49781
    * gcc.target/i386/pr49781-1.c: New.

Added:
    trunk/gcc/testsuite/gcc.target/i386/pr49781-1.c
Modified:
    trunk/gcc/testsuite/ChangeLog

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #37 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-08 17:16:44 UTC ---
(In reply to comment #35)
> It works much better now. But gcc.dg/torture/pr47744-2.c compiled with
> 
> -mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops
> 
> still generates those leal:
> 
>     leal    (%rsi,%r9), %ebp
>     leal    (%r10,%r9), %r12d
>         leal    (%r10,%r9), %r12d

This partial address won't combine due to multiple uses.

IIRC, there was some discussion to distribute partial addresses back to their
use sites, but OTOH it all depends on how hard you want to hammer your address
generation unit.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[uros at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #38 from uros at gcc dot gnu.org 2011-08-09 07:38:07 UTC ---
Author: uros
Date: Tue Aug  9 07:38:02 2011
New Revision: 177583

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177583
Log:
    PR target/49781
    * config/i386/i386.md (reload_noff_load): New.
    (reload_noff_store): Ditto.
    * config/i386/i386.c (ix86_secondary_reload): Use
    CODE_FOR_reload_noff_load and CODE_FOR_reload_noff_store to handle
    double-word moves from/to non-offsetable addresses instead of
    generating XMM temporary.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/config/i386/i386.md

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #39 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-09 18:37:37 UTC ---
Created attachment 24967
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24967
Patch that recognizes addresses, zero-extended with AND

Attached patch adds recognision of addresses, zero-extended with AND.

The patch fixes gcc-target/i386/pr43766.c testcase.

H.J., can you please test it on x32 ?

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #40 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-09 18:50:10 UTC ---
(In reply to comment #39)
> Created attachment 24967 [details]
> Patch that recognizes addresses, zero-extended with AND
> 
> Attached patch adds recognision of addresses, zero-extended with AND.
> 
> The patch fixes gcc-target/i386/pr43766.c testcase.
> 
> H.J., can you please test it on x32 ?

I am testing it now.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #41 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-09 18:59:46 UTC ---
(In reply to comment #39)
> Created attachment 24967 [details]
> Patch that recognizes addresses, zero-extended with AND
> 
> Attached patch adds recognision of addresses, zero-extended with AND.
> 
> The patch fixes gcc-target/i386/pr43766.c testcase.
> 
> H.J., can you please test it on x32 ?

It seems to generate more leal for gcc.dg/torture/pr47744-2.c
compiled with

-mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #42 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-09 21:21:35 UTC ---
(In reply to comment #41)
> (In reply to comment #39)
> > Created attachment 24967 [details]
> > Patch that recognizes addresses, zero-extended with AND
> > 
> > Attached patch adds recognision of addresses, zero-extended with AND.
> > 
> > The patch fixes gcc-target/i386/pr43766.c testcase.
> > 
> > H.J., can you please test it on x32 ?
> 
> It seems to generate more leal for gcc.dg/torture/pr47744-2.c
> compiled with
> 
> -mx32 -O3 -std=gnu99 -ftree-vectorize -funroll-loops

It also caused:

spawn -ignore SIGHUP /export/build/gnu/gcc-x32/build-x86_64-linux/gcc/xgcc
-B/export/build/gnu/gcc-x32/build-x86_64-linux/gcc/
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.c-torture/execute/builtin-prefetch-4.c
-w -O2 -lm -mx32 -o
/export/build/gnu/gcc-x32/build-x86_64-linux/gcc/testsuite/gcc2/builtin-prefetch-4.x2^M
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.c-torture/execute/builtin-prefetch-4.c:
In function ‘preinc_glob_ptr’:^M
/export/gnu/import/git/gcc-x32/gcc/testsuite/gcc.c-torture/execute/builtin-prefetch-4.c:67:1:
internal compiler error: output_operand: invalid expression as operand^M
Please submit a full bug report,^M
with preprocessed source if appropriate.^M
See <http://gcc.gnu.org/bugs.html> for instructions.^M
compiler exited with status 1

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #43 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-10 18:19:05 UTC ---
(In reply to comment #41)

> > Patch that recognizes addresses, zero-extended with AND

> It seems to generate more leal for gcc.dg/torture/pr47744-2.c

No, the extra leas are due to reload-nonoffsetable-address patch (the previous
one). These instructions avoids ICE by reloading non-offsetable address of
double-word operand to a temporary.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #24967|0                           |1
        is obsolete|                            |

--- Comment #44 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-10 18:29:34 UTC ---
Created attachment 24973
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24973
Patch that recognizes addresses, zero-extended with AND, v2.

(In reply to comment #42)

> It also caused:
> 
> internal compiler error: output_operand: invalid expression as operand^M
> Please submit a full bug report,^M

Oh, nice - we can just discard paradoxical subregs in this case.  Please find
attached updated revision of the patch that avoids ICE.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #45 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-10 18:44:19 UTC ---
(In reply to comment #44)
> Created attachment 24973 [details]
> Patch that recognizes addresses, zero-extended with AND, v2.
> 
> (In reply to comment #42)
> 
> > It also caused:
> > 
> > internal compiler error: output_operand: invalid expression as operand^M
> > Please submit a full bug report,^M
> 
> Oh, nice - we can just discard paradoxical subregs in this case.  Please find
> attached updated revision of the patch that avoids ICE.

I am testing it now.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #46 from H.J. Lu <hjl.tools at gmail dot com> 2011-08-11 12:55:11 UTC ---
(In reply to comment #45)
> (In reply to comment #44)
> > Created attachment 24973 [details]
> > Patch that recognizes addresses, zero-extended with AND, v2.
> > 
> > (In reply to comment #42)
> > 
> > > It also caused:
> > > 
> > > internal compiler error: output_operand: invalid expression as operand^M
> > > Please submit a full bug report,^M
> > 
> > Oh, nice - we can just discard paradoxical subregs in this case.  Please find
> > attached updated revision of the patch that avoids ICE.
> 
> I am testing it now.

No regressions in GCC, glibc and SPEC CPU 2K/2006.

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[uros at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

--- Comment #47 from uros at gcc dot gnu.org 2011-08-11 20:03:34 UTC ---
Author: uros
Date: Thu Aug 11 20:03:29 2011
New Revision: 177683

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177683
Log:
    PR target/49781
    * config/i386/i386.md (*lea_5_zext): New.
    (*lea_6_zext): Ditto.
    * config/i386/predicates.md (const_32bit_mask): New predicate.
    (lea_address_operand): Reject AND.
    * config/i386/i386.c (ix86_decompose_address): Allow Dimode AND with
    const_32bit_mask immediate.
    (ix86_print_operand_address): Handle AND.
    (memory_address_length): Ditto.


Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/config/i386/i386.md
    trunk/gcc/config/i386/predicates.md

[Bug target/49781] [x32] Unnecessary lea in x32 mode

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49781

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x32
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.7.0

--- Comment #48 from Uros Bizjak <ubizjak at gmail dot com> 2011-08-12 06:31:31 UTC ---
The remaining LEAs are due to:
- CSE'd partial address, these can be distributed into uses, based on some cost
model, etc...
- reloaded non-offsetable address to handle "o" constraint, necessary.

Let's close this bug and eventually open new one on a case-by-case basis.

So, fixed.