Re: [PATCH] RISC-V: optimize stack manipulation in save-restore

["Fei Gao" <gaofei@eswincomputing.com>]

On 2022-12-01 06:50  Palmer Dabbelt <palmer@dabbelt.com> wrote:
>
>On Wed, 30 Nov 2022 00:37:17 PST (-0800), gaofei@eswincomputing.com wrote:
>> The stack that save-restore reserves is not well accumulated in stack allocation and deallocation.
>> This patch allows less instructions to be used in stack allocation and deallocation if save-restore enabled,
>> and also a much clear logic for save-restore stack manipulation.
>>
>> before patch:
>> bar:
>> call	t0,__riscv_save_4
>> addi	sp,sp,-64
>> ...
>> li	t0,-12288
>> addi	t0,t0,-1968 # optimized out after patch
>> add	sp,sp,t0 # prologue
>> ...
>> li	t0,12288 # epilogue
>> addi	t0,t0,2000 # optimized out after patch
>> add	sp,sp,t0
>> ...
>> addi	sp,sp,32
>> tail	__riscv_restore_4
>>
>> after patch:
>> bar:
>> call	t0,__riscv_save_4
>> addi	sp,sp,-2032
>> ...
>> li	t0,-12288
>> add	sp,sp,t0 # prologue
>> ...
>> li	t0,12288 # epilogue
>> add	sp,sp,t0
>> ...
>> addi	sp,sp,2032
>> tail	__riscv_restore_4
>>
>> gcc/ChangeLog:
>>
>>         * config/riscv/riscv.cc (riscv_first_stack_step): add a new function parameter remaining_size.
>>         (riscv_compute_frame_info): adapt new riscv_first_stack_step interface.
>>         (riscv_expand_prologue): consider save-restore in stack allocation.
>>         (riscv_expand_epilogue): consider save-restore in stack deallocation.
>>
>> gcc/testsuite/ChangeLog:
>>
>>         * gcc.target/riscv/stack_save_restore.c: New test.
>> ---
>>  gcc/config/riscv/riscv.cc                     | 58 ++++++++++---------
>>  .../gcc.target/riscv/stack_save_restore.c     | 40 +++++++++++++
>>  2 files changed, 70 insertions(+), 28 deletions(-)
>>  create mode 100644 gcc/testsuite/gcc.target/riscv/stack_save_restore.c
>
>I guess with the RISC-V backend still being open for things as big as
>the V port we should probably be taking code like this as well?  I
>wouldn't be opposed to making an exception for the V code and holding
>everything else back, though.
>
>> diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
>> index 05bdba5ab4d..9e92e729a5f 100644
>> --- a/gcc/config/riscv/riscv.cc
>> +++ b/gcc/config/riscv/riscv.cc
>> @@ -4634,7 +4634,7 @@ riscv_save_libcall_count (unsigned mask)
>>     They decrease stack_pointer_rtx but leave frame_pointer_rtx and
>>     hard_frame_pointer_rtx unchanged.  */
>>
>> -static HOST_WIDE_INT riscv_first_stack_step (struct riscv_frame_info *frame);
>> +static HOST_WIDE_INT riscv_first_stack_step (struct riscv_frame_info *frame, poly_int64 remaining_size);
>>
>>  /* Handle stack align for poly_int.  */
>>  static poly_int64
>> @@ -4663,7 +4663,7 @@ riscv_compute_frame_info (void)
>>       save/restore t0.  We check for this before clearing the frame struct.  */
>>    if (cfun->machine->interrupt_handler_p)
>>      {
>> -      HOST_WIDE_INT step1 = riscv_first_stack_step (frame);
>> +      HOST_WIDE_INT step1 = riscv_first_stack_step (frame, frame->total_size);
>>        if (! POLY_SMALL_OPERAND_P ((frame->total_size - step1)))
>>  interrupt_save_prologue_temp = true;
>>      }
>> @@ -4913,31 +4913,31 @@ riscv_restore_reg (rtx reg, rtx mem)
>>     without adding extra instructions.  */
>>
>>  static HOST_WIDE_INT
>> -riscv_first_stack_step (struct riscv_frame_info *frame)
>> +riscv_first_stack_step (struct riscv_frame_info *frame, poly_int64 remaining_size)
>>  {
>> -  HOST_WIDE_INT frame_total_constant_size;
>> -  if (!frame->total_size.is_constant ())
>> -    frame_total_constant_size
>> -      = riscv_stack_align (frame->total_size.coeffs[0])
>> -	- riscv_stack_align (frame->total_size.coeffs[1]);
>> +  HOST_WIDE_INT remaining_const_size;
>> +  if (!remaining_size.is_constant ())
>> +    remaining_const_size
>> +      = riscv_stack_align (remaining_size.coeffs[0])
>> +	- riscv_stack_align (remaining_size.coeffs[1]);
>
>The alignment looks off here, at least in the email.  Worth fixing it up
>if you're touching the lines anyway. 

Sure, i will change RISCV_STACK_ALIGN into riscv_stack_align.

>
>>    else
>> -    frame_total_constant_size = frame->total_size.to_constant ();
>> +    remaining_const_size = remaining_size.to_constant ();
>>
>> -  if (SMALL_OPERAND (frame_total_constant_size))
>> -    return frame_total_constant_size;
>> +  if (SMALL_OPERAND (remaining_const_size))
>> +    return remaining_const_size;
>>
>>    HOST_WIDE_INT min_first_step =
>> -    RISCV_STACK_ALIGN ((frame->total_size - frame->frame_pointer_offset).to_constant());
>> +    RISCV_STACK_ALIGN ((remaining_size - frame->frame_pointer_offset).to_constant());
>>    HOST_WIDE_INT max_first_step = IMM_REACH / 2 - PREFERRED_STACK_BOUNDARY / 8;
>> -  HOST_WIDE_INT min_second_step = frame_total_constant_size - max_first_step;
>> +  HOST_WIDE_INT min_second_step = remaining_const_size - max_first_step;
>>    gcc_assert (min_first_step <= max_first_step);
>>
>>    /* As an optimization, use the least-significant bits of the total frame
>>       size, so that the second adjustment step is just LUI + ADD.  */
>>    if (!SMALL_OPERAND (min_second_step)
>> -      && frame_total_constant_size % IMM_REACH < IMM_REACH / 2
>> -      && frame_total_constant_size % IMM_REACH >= min_first_step)
>> -    return frame_total_constant_size % IMM_REACH;
>> +      && remaining_const_size % IMM_REACH < IMM_REACH / 2
>> +      && remaining_const_size % IMM_REACH >= min_first_step)
>> +    return remaining_const_size % IMM_REACH;
>
>Looks like this entire frame->total_size -> remaining_size conversion
>could be done as an independent patch that would change no
>functionality, that's always a nice way to do things as it makes the
>code easier to read. 

Sure, i will split this patch.

>
>I spent a bit poking around here and nothing wrong is jumping out, but
>trying to keep all these offset differences in my head is a bit tricky. 
>If you have the time to refactor this to be easier to read that'd be
>great, otherwise hopefully I (or someone else) will have the time to
>take a look -- probably not today on my end, though, as I've got some
>Linux backlog to look at. 

I'll try it. Also i propose to add step0 for pre-allocated stack cases like save-restore lib call, 
and the future Zcmp in Zc* extension if needed. 
So we have a clear logic of manipulate stack step by step.

>
>Thanks!
>
>>    if (TARGET_RVC)
>>      {
>> @@ -5037,9 +5037,7 @@ riscv_expand_prologue (void)
>>    /* Save the registers.  */
>>    if ((frame->mask | frame->fmask) != 0)
>>      {
>> -      HOST_WIDE_INT step1 = riscv_first_stack_step (frame);
>> -      if (size.is_constant ())
>> -	step1 = MIN (size.to_constant(), step1);
>> +      HOST_WIDE_INT step1 = riscv_first_stack_step (frame, size);
>>
>>        insn = gen_add3_insn (stack_pointer_rtx,
>>      stack_pointer_rtx,
>> @@ -5142,6 +5140,8 @@ riscv_expand_epilogue (int style)
>>    HOST_WIDE_INT step2 = 0;
>>    bool use_restore_libcall = ((style == NORMAL_RETURN)
>>        && riscv_use_save_libcall (frame));
>> +  unsigned libcall_size = use_restore_libcall ?
>> +                            frame->save_libcall_adjustment : 0;
>>    rtx ra = gen_rtx_REG (Pmode, RETURN_ADDR_REGNUM);
>>    rtx insn;
>>
>> @@ -5212,13 +5212,18 @@ riscv_expand_epilogue (int style)
>>        REG_NOTES (insn) = dwarf;
>>      }
>>
>> +  if (use_restore_libcall)
>> +    frame->mask = 0; /* Temporarily fib for GPRs.  */
>> +
>>    /* If we need to restore registers, deallocate as much stack as
>>       possible in the second step without going out of range.  */
>>    if ((frame->mask | frame->fmask) != 0)
>> -    {
>> -      step2 = riscv_first_stack_step (frame);
>> -      step1 -= step2;
>> -    }
>> +    step2 = riscv_first_stack_step (frame, frame->total_size - libcall_size);
>> +
>> +  if (use_restore_libcall)
>> +    frame->mask = mask; /* Undo the above fib.  */
>> +
>> +  step1 -= step2 + libcall_size;
>>
>>    /* Set TARGET to BASE + STEP1.  */
>>    if (known_gt (step1, 0))
>> @@ -5272,15 +5277,12 @@ riscv_expand_epilogue (int style)
>>      frame->mask = 0; /* Temporarily fib that we need not save GPRs.  */
>>
>>    /* Restore the registers.  */
>> -  riscv_for_each_saved_reg (frame->total_size - step2, riscv_restore_reg,
>> +  riscv_for_each_saved_reg (frame->total_size - step2 - libcall_size,
>> +                            riscv_restore_reg,
>>      true, style == EXCEPTION_RETURN);
>>
>>    if (use_restore_libcall)
>> -    {
>>        frame->mask = mask; /* Undo the above fib.  */
>> -      gcc_assert (step2 >= frame->save_libcall_adjustment);
>> -      step2 -= frame->save_libcall_adjustment;
>> -    }
>>
>>    if (need_barrier_p)
>>      riscv_emit_stack_tie ();
>> diff --git a/gcc/testsuite/gcc.target/riscv/stack_save_restore.c b/gcc/testsuite/gcc.target/riscv/stack_save_restore.c
>> new file mode 100644
>> index 00000000000..4695ef9469a
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/riscv/stack_save_restore.c
>> @@ -0,0 +1,40 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-march=rv32imafc -mabi=ilp32f -msave-restore -O2 -fno-schedule-insns -fno-schedule-insns2 -fno-unroll-loops -fno-peel-loops" } */
>> +/* { dg-final { check-function-bodies "**" "" } } */
>> +
>> +char my_getchar();
>> +float getf();
>> +
>> +/*
>> +**bar:
>> +**	call	t0,__riscv_save_4
>> +**	addi	sp,sp,-2032
>> +**	...
>> +**	li	t0,-12288
>> +**	add	sp,sp,t0
>> +**	...
>> +**	li	t0,12288
>> +**	add	sp,sp,t0
>> +**	...
>> +**	addi	sp,sp,2032
>> +**	tail	__riscv_restore_4
>> +*/
>
>The test needs to actually check this, it can't just be manual. 

I didn't get your point. 
The { dg-final { check-function-bodies "**" "" } } matches the output with the expected result above automatically. 
Please let me know your idea. 

Thanks & BR, 
Fei

>
>> +int bar()
>> +{
>> +  float volatile farray[3568];
>> +
>> +  float sum = 0;
>> +  float f1 = getf();
>> +  float f2 = getf();
>> +  float f3 = getf();
>> +  float f4 = getf();
>> +
>> +  for (int i = 0; i < 3568; i++)
>> +  {
>> +    farray[i] = my_getchar() * 1.2;
>> +    sum += farray[i];
>> +  }
>> +
>> +  return sum + f1 + f2 + f3 + f4;
>> +}
>> +