MIPS: Long branch implementation and trampoline improvement.

MIPS: Long branch implementation and trampoline improvement.

Improve the branch and branch-trampoline mechanism to automatically
use long-jumps when function size grows large. Reduce size of emitted
trampoline pools.

Now passes mozilla regress-80981.js.

BUG=
TEST=

Committed: http://code.google.com/p/v8/source/detail?r=8433

[Søren Thygesen Gjesse, 2011-06-24 13:05:15]

LGTM with comments

I find the additional argument for lui and ori somewhat obscure.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc
File src/mips/assembler-mips.cc (right):

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:633: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
IsLui(instr_lui) is not needed in the ASSERT (already covered by the if).

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:685: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
Ditto.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:1836: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
Ditto.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:2000: ori(at, at, (imm32 & kImm16Mask), false);
Can't the jr go between the lui and the ori and then the nop can be avoided?

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h
File src/mips/assembler-mips.h (right):

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:647: void ori(Register rd, Register rs, int32_t j,
bool check_buffer = true);
This boolean flag seems rather obscure - when should it be set to false?

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:1121: // already a slow case which can possibly break
our code generation for
for -> for the / in the

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:1122: // extreme case, we use this information to
trigger different mode for
for -> of

[Paul Lind, 2011-06-28 06:53:14]

Changed the method of blocking buffer growth from a boolean passed around, to a
class & member variable. I think it is more readable this way. Other minor
cleanup, and review comments responded to.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc
File src/mips/assembler-mips.cc (right):

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:633: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> IsLui(instr_lui) is not needed in the ASSERT (already covered by the if).

Done.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:685: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> Ditto.

Done.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:1836: ASSERT(IsLui(instr_lui) && IsOri(instr_ori));
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> Ditto.

Done.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.cc#newc...
src/mips/assembler-mips.cc:2000: ori(at, at, (imm32 & kImm16Mask), false);
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> Can't the jr go between the lui and the ori and then the nop can be avoided?

Unfortunately, it cannot. This is jr (jump) to 'at' register, and lui/ori are
loading value to 'at' register. That must be valid at the start of the 'jr'
instruction.

You can alter a register (such as 'at') within the delay slot of a jump/branch
utilizing that same register, but the new value would be seen only when you
arrived at the branch destination. It doesn't make sense in code like this,
since you need the branch to go the correct place.

It can make sense in other situations, such as incrementing a loop counter
register in the delay slot of the branch that is testing for loop termination
(if you are careful with the count).

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h
File src/mips/assembler-mips.h (right):

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:647: void ori(Register rd, Register rs, int32_t j,
bool check_buffer = true);
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> This boolean flag seems rather obscure - when should it be set to false?

Agree that it was confusing. I have removed the flag and used a blocking scope
class & new class member boolean for this purpose, which is to signal the low
level emit() routine to NOT check & grow the assembly buffer.

The big picture of this is that we can now use internal references that are
absolute pointers within the code for very long functions. These need to be
relocated by patching the lui/ori instruction-pair, or optionally a j
instruction. But if the relocation is triggered by buffer check/growth before
those instructions are emitted, there is nothing to patch and it all fails. So
we delay buffer growth during these sequences using BlockGrowBufferScope.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:1121: // already a slow case which can possibly break
our code generation for
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> for -> for the / in the

Done.

http://codereview.chromium.org/7239020/diff/1/src/mips/assembler-mips.h#newco...
src/mips/assembler-mips.h:1122: // extreme case, we use this information to
trigger different mode for
On 2011/06/24 13:05:15, Søren Gjesse wrote:
> for -> of

Done.

[Søren Thygesen Gjesse, 2011-06-28 06:58:19]

LGTM

I will commit it.