From: Uros Bizjak <ubizjak@gmail.com>
To: Ulrich Weigand <uweigand@de.ibm.com>
Cc: gcc-patches@gcc.gnu.org, GCC Development <gcc@gcc.gnu.org>,
"H.J. Lu" <hjl.tools@gmail.com>
Subject: Re: [RFC PATCH, i386]: Allow zero_extended addresses (+ problems with reload and offsetable address, "o" constraint)
Date: Mon, 08 Aug 2011 17:12:00 -0000 [thread overview]
Message-ID: <CAFULd4Y5mie9zBP5TbvFpuDxSwbKpp=n3sh8iyr3UpSfiqiYxQ@mail.gmail.com> (raw)
In-Reply-To: <201108081530.p78FUAgM029764@d06av02.portsmouth.uk.ibm.com>
[-- Attachment #1: Type: text/plain, Size: 2145 bytes --]
On Mon, Aug 8, 2011 at 5:30 PM, Ulrich Weigand <uweigand@de.ibm.com> wrote:
> Uros Bizjak wrote:
>
>> Although, it would be nice for reload to subsequently fix CSE'd
>> non-offsetable memory by copying address to temporary reg (*as said in
>> the documentation*), we could simply require an XMM temporary for
>> TImode reloads to/from integer registers, and this fixes ICE for x32.
>
> Moves are special as far as reload is concerned. If there is already
> a move instruction present *before* reload, it will get fixed up
> according to its constraints as any other instruction.
>
> However, reload will *introduce* new moves as part of its operation,
> and those will *not* themselves get reloaded. Instead, reload simply
> assumes that every plain move will just succeed without requiring
> any reload; if this is not true, the target *must* provide a
> secondary reload for this move.
>
> (Note that the secondary reload could also work by reloading the
> target address into a temporary; that's up to the target to
> implement.)
Whoa, indeed.
Using attached patch that reloads memory address instead of going
through XMM register, the code for the testcase improves from:
test:
.LFB0:
.cfi_startproc
pushq %rbx
.cfi_def_cfa_offset 16
.cfi_offset 3, -16
sall $4, %esi
addl %edi, %esi
movdqa (%esi), %xmm0
movdqa %xmm0, -16(%rsp)
movq -16(%rsp), %rcx
movq -8(%rsp), %rbx
addq $1, %rcx
adcq $0, %rbx
movq %rcx, -16(%rsp)
sall $4, %edx
movq %rbx, -8(%rsp)
movdqa -16(%rsp), %xmm0
movdqa %xmm0, (%esi)
pxor %xmm0, %xmm0
movdqa %xmm0, (%edx,%esi)
popq %rbx
.cfi_def_cfa_offset 8
ret
to:
test:
.LFB0:
.cfi_startproc
sall $4, %esi
pushq %rbx
.cfi_def_cfa_offset 16
.cfi_offset 3, -16
addl %edi, %esi
pxor %xmm0, %xmm0
mov %esi, %eax
movq (%rax), %rcx
movq 8(%rax), %rbx
addq $1, %rcx
adcq $0, %rbx
sall $4, %edx
movq %rcx, (%rax)
movq %rbx, 8(%rax)
movdqa %xmm0, (%edx,%esi)
popq %rbx
.cfi_def_cfa_offset 8
ret
H.J., can you please test attached patch? This optimization won't
trigger on x86_64 anymore.
Thanks,
Uros.
[-- Attachment #2: z.diff.txt --]
[-- Type: text/plain, Size: 2657 bytes --]
Index: config/i386/i386.md
===================================================================
--- config/i386/i386.md (revision 177565)
+++ config/i386/i386.md (working copy)
@@ -2073,6 +2073,40 @@
(const_string "orig")))
(set_attr "mode" "SI,DI,DI,DI,SI,DI,DI,DI,DI,DI,TI,DI,TI,DI,DI,DI,DI,DI")])
+;; Reload patterns to support multi-word load/store
+;; with non-offsetable address.
+(define_expand "reload_noff_store"
+ [(parallel [(match_operand 0 "memory_operand" "=m")
+ (match_operand 1 "register_operand" "r")
+ (match_operand:DI 2 "register_operand" "=&r")])]
+ "TARGET_64BIT"
+{
+ rtx mem = operands[0];
+ rtx addr = XEXP (mem, 0);
+
+ emit_move_insn (operands[2], addr);
+ mem = replace_equiv_address_nv (mem, operands[2]);
+
+ emit_insn (gen_rtx_SET (VOIDmode, mem, operands[1]));
+ DONE;
+})
+
+(define_expand "reload_noff_load"
+ [(parallel [(match_operand 0 "register_operand" "=r")
+ (match_operand 1 "memory_operand" "m")
+ (match_operand:DI 2 "register_operand" "=r")])]
+ "TARGET_64BIT"
+{
+ rtx mem = operands[1];
+ rtx addr = XEXP (mem, 0);
+
+ emit_move_insn (operands[2], addr);
+ mem = replace_equiv_address_nv (mem, operands[2]);
+
+ emit_insn (gen_rtx_SET (VOIDmode, operands[0], mem));
+ DONE;
+})
+
;; Convert impossible stores of immediate to existing instructions.
;; First try to get scratch register and go through it. In case this
;; fails, move by 32bit parts.
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 177566)
+++ config/i386/i386.c (working copy)
@@ -28245,18 +28245,25 @@
static reg_class_t
ix86_secondary_reload (bool in_p, rtx x, reg_class_t rclass,
- enum machine_mode mode,
- secondary_reload_info *sri ATTRIBUTE_UNUSED)
+ enum machine_mode mode, secondary_reload_info *sri)
{
/* Double-word spills from general registers to non-offsettable memory
- references (zero-extended addresses) go through XMM register. */
+ references (zero-extended addresses) require special handling. */
if (TARGET_64BIT
&& MEM_P (x)
&& GET_MODE_SIZE (mode) > UNITS_PER_WORD
&& rclass == GENERAL_REGS
&& !offsettable_memref_p (x))
- return SSE_REGS;
+ {
+ sri->icode = (in_p
+ ? CODE_FOR_reload_noff_load
+ : CODE_FOR_reload_noff_store);
+ /* Add the cost of move to a temporary. */
+ sri->extra_cost = 1;
+ return NO_REGS;
+ }
+
/* QImode spills from non-QI registers require
intermediate register on 32bit targets. */
if (!TARGET_64BIT
next prev parent reply other threads:[~2011-08-08 17:12 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2011-08-05 18:51 Uros Bizjak
2011-08-07 12:39 ` Uros Bizjak
2011-08-08 15:30 ` Ulrich Weigand
2011-08-08 17:12 ` Uros Bizjak [this message]
2011-08-08 17:14 ` H.J. Lu
2011-08-09 7:41 ` Uros Bizjak
2011-08-09 15:40 ` H.J. Lu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAFULd4Y5mie9zBP5TbvFpuDxSwbKpp=n3sh8iyr3UpSfiqiYxQ@mail.gmail.com' \
--to=ubizjak@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=gcc@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--cc=uweigand@de.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).