Re: Base subreg rules on REGMODE_NATURAL_SIZE rather than UNITS_PER_WORD

[Jeff Law <law at redhat dot com>]

On 09/18/2017 05:26 AM, Richard Sandiford wrote:
> Originally subregs operated at the word level and subreg offsets
> were measured in words.  The offset units were later changed from
> words to bytes (SUBREG_WORD became SUBREG_BYTE), but the fundamental
> assumption that subregs should operate at the word level remained.
> Whether (subreg:M1 (reg:M2 R2) N) is well-formed depended on the
> way that M1 and M2 partitioned into words and whether the subword
> part of N represented a lowpart.  However, some questions depended
> instead on the macro REGMODE_NATURAL_SIZE, which was introduced
> as part of the patch that moved from SUBREG_WORD to SUBREG_BYTE.
> It is used to decide whether setting (subreg:M1 (reg:M2 R2) N)
> clobbers all of R2 or just part of it (df_read_modify_subreg).
> 
> Using words doesn't really make sense for modern vector
> architectures.  Vector registers are usually bigger than
> a word and:
> 
> (a) setting the scalar lowpart of them usually clobbers the
>     rest of the register (contrary to the subreg rules,
>     where only the containing words should be clobbered).
> 
> (b) high words of vector registers are often not independently
>     addressable, even though that's what the subreg rules expect.
> 
> This patch therefore uses REGMODE_NATURAL_SIZE instead of
> UNITS_PER_WORD to determine the size of the independently
> addressable blocks in an inner register.
> 
> This is needed for SVE because the number of words in a vector
> mode isn't known at compile time, so isn't a sensible basis
> for calculating the number of registers.
> 
> The only existing port to define REGMODE_NATURAL_SIZE is
> 64-bit SPARC, where FP registers are 32 bits.  (This is the
> opposite of the use case for SVE, since the natural division
> is smaller than a word.)  I compiled the testsuite before and
> after the patch for sparc64-linux-gnu and the only test whose
> assembly changed was g++.dg/debug/pr65678.C, where the order
> of two independent stores was reversed and where a different
> register was picked for one pseudo.  The new code was
> otherwise equivalent to the old code.
> 
> Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64le-linux-gnu.
> Also tested by comparing the testsuite assembly output on at least one
> target per CPU directory, with only the SPARC differences just mentioned.
> OK to install?
> 
> Richard
> 
> 
> 2017-09-18  Richard Sandiford  <richard.sandiford@linaro.org>
> 	    Alan Hayward  <alan.hayward@arm.com>
> 	    David Sherwood  <david.sherwood@arm.com>
> 
> gcc/
> 	* doc/rtl.texi: Rewrite the subreg rules so that they partition
> 	the inner register into REGMODE_NATURAL_SIZE bytes rather than
> 	UNITS_PER_WORD bytes.
> 	* emit-rtl.c (validate_subreg): Divide subregs into blocks
> 	based on REGMODE_NATURAL_SIZE of the inner mode.
> 	(gen_lowpart_common): Split the SCALAR_FLOAT_MODE_P and
> 	!SCALAR_FLOAT_MODE_P cases.  Use REGMODE_NATURAL_SIZE for the latter.
> 	* expr.c (store_constructor): Use REGMODE_NATURAL_SIZE to test
> 	whether something is likely to occupy more than one register.
OK.  Though I must admit, I expected a larger change after reading the
intro.

jeff