Patchwork RFA: Simplifying truncation and integer lowpart subregs

login
register
mail settings
Submitter Richard Sandiford
Date Nov. 28, 2012, 9:45 p.m.
Message ID <87boehmmfi.fsf@talisman.default>
Download mbox | patch
Permalink /patch/202558/
State New
Headers show

Comments

Richard Sandiford - Nov. 28, 2012, 9:45 p.m.
Ramana Radhakrishnan <ramana.gcc@googlemail.com> writes:
> On Sun, Oct 7, 2012 at 8:56 AM, Richard Sandiford
> <rdsandiford@googlemail.com> wrote:
>> Eric Botcazou <ebotcazou@adacore.com> writes:
>>>> I think modelling it as a TRUNCATE operation is correct for
>>>> !TRULY_NOOP_TRUNCATION (it's the bug that Andrew pointed out).
>>>> And we shouldn't generate an actual TRUNCATE rtx for
>>>> TRULY_NOOP_TRUNCATION (the thing about making
>>>> simplify_gen_unary (TRUNCATE, ...) no worse than simplify_gen_subreg
>>>> for those targets).  I suppose:
>>>>
>>>>       /* We can't handle truncation to a partial integer mode here
>>>>          because we don't know the real bitsize of the partial
>>>>          integer mode.  */
>>>>       if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT)
>>>>         break;
>>>>
>>>> might be a problem though; we should still allow a subreg to be
>>>> generated.  Is that what you were thinking of, or something else?
>>>
>>> I was thinking of the !TRULY_NOOP_TRUNCATION case, where the two operations
>>> aren't equivalent.  Generating TRUNCATE in simplify_subreg seems
>>> suspicious to
>>> me in this case but, if not doing it is the source of the bug, I guess I need
>>> to do some homework on this TRULY_NOOP_TRUNCATION stuff. :-)
>>>
>>> Maybe add a blurb to the head comment of simplify_truncation, explaining that
>>> it is valid to call the function both for TRUNCATEs and truncations to the
>>> lowpart, and why it is correct to generate new TRUNCATEs in the latter case.
>>
>> Yeah, in hindsight, the patch was definitely lacking commentary.
>> How about the patch below?  It also fixes the partial int case
>> and gets rid of the errant NOT hunk.  Tested in the same way as before.
>>
>> Richard
>>
>>
>> gcc/
>>         * machmode.h (GET_MODE_UNIT_PRECISION): New macro.
>>         * simplify-rtx.c (simplify_truncation): New function,
>>         extracted from simplify_subreg and (in small part) from
>>         simplify_unary_operation_1.
>>         (simplify_unary_operation_1) <TRUNCATE>: Use it.  Remove sign bit
>>         test for !TRULY_NOOP_TRUNCATION_MODES_P.
>>         (simplify_subreg): Use simplify_truncate for lowpart subregs
>>         where both the inner and outer modes are scalar integers.
>>         * config/mips/mips.c (mips_truncated_op_cost): New function.
>>         (mips_rtx_costs): Adjust test for BADDU.
>>         * config/mips/mips.md (*baddu_di<mode>): Push truncates to operands.
>
> This triggers PR55052 on ARM.I've attached the .i file and the dumps
> to the bug report.

Thanks.  I'd managed to drop a SCALAR_INT_MODE_P check when splitting
the ZERO_EXTEND handling into two.

This patch reinstates the check.  Tested on x86_64-linux-gnu and applied
as obvious.

Richard


gcc/
	PR rtl-optimization/55052
	* simplify-rtx.c (simplify_subreg): Restore SCALAR_INT_MODE_P check.

Patch

Index: gcc/simplify-rtx.c
===================================================================
--- gcc/simplify-rtx.c	2012-11-27 18:52:29.000000000 +0000
+++ gcc/simplify-rtx.c	2012-11-28 19:54:30.500525576 +0000
@@ -5875,7 +5875,7 @@  simplify_subreg (enum machine_mode outer
 
   /* A SUBREG resulting from a zero extension may fold to zero if
      it extracts higher bits that the ZERO_EXTEND's source bits.  */
-  if (GET_CODE (op) == ZERO_EXTEND)
+  if (GET_CODE (op) == ZERO_EXTEND && SCALAR_INT_MODE_P (innermode))
     {
       unsigned int bitpos = subreg_lsb_1 (outermode, innermode, byte);
       if (bitpos >= GET_MODE_PRECISION (GET_MODE (XEXP (op, 0)))