Making rtxes use less memory (sometimes!)

[Zack Weinberg]

Richard Sandiford <rsandifo@redhat.com> writes:

> Here's the sort of speed-up I get when compiling mips.i
> using a mips64-linux-gnu native:
>
>      -O0
> 	     before         after
>      real    0m31.478s      0m29.830s
>      user    0m31.270s      0m29.670s
>      sys     0m0.200s       0m0.130s
>
>      -O1
> 	     before         after
>      real    0m16.364s      0m15.465s
>      user    0m16.220s      0m15.270s
>      sys     0m0.140s       0m0.190s
>
>      -O2
> 	     before         after
>      real    0m7.804s       0m7.104s
>      user    0m7.690s       0m7.030s
>      sys     0m0.100s       0m0.070s
>
> (these are reproducable to within ~0.1s).

Do you have the -O numbers backward?  It seems odd that -O0 would be
four times slower than -O2.

Still, the trend is clear.  I'm impressed that a small data layout
change like this can make that much difference.

> 	* rtl.h (rtl_size): Declare.
> 	(rtunion): Remove rtwint.
> 	(rtx_def): Replace 'fld' with a union of an rtunion or a HOST_WIDE_INT.
> 	(RTX_HDR_SIZE, RTX_SIZE): New macros.
> 	(RTL_CHECK1): Adjust for new rtx_def layout.
> 	(RTL_CHECK2, RTL_CHECKC1, RTL_CHECKC2): Likewise.
> 	(XWINT, XCWINT): Likewise.  Access the rtx structure directly.
> 	(X0WINT): Remove.
> 	(X0ANY): New macro.
> 	* rtl.def: Adjust comments for new rtx_def layout.
> 	* ggc.h (ggc_alloc_rtx): Take the rtx code as argument, not the
> 	number of slots.
> 	* rtl.c (rtx_size): New array.
> 	(rtx_alloc): Adjust call to ggc_alloc_rtx.  Use RTX_HDR_SIZE.
> 	(copy_rtx): Use RTX_HDR_SIZE.  Adjust for new rtx_def layout.
> 	(shallow_copy_rtx): Adjust call to ggc_alloc_rtx.  Use RTX_SIZE.
> 	* integrate.c (copy_rtx_and_substitute): Use X0ANY to copy '0' fields.
> 	* emit-rtl.c (copy_most_rtx): Likewise.
> 	(copy_rtx_if_shared): Use RTX_SIZE.
> 	(copy_insn_1): Use RTX_HDR_SIZE.  Adjust for new rtx_def layout.
> 	* gengenrtl.c (gendef): Adjust ggc_alloc_rtx call.  Use RTX_HDR_SIZE.
> 	* gengtype.c (write_rtx_next): Use RTX_HDR_SIZE.
> 	(adjust_field_rtx_def): Expect "rtx_def" to be a union rather than
> 	an array.  Adjust output for new rtx_def layout.
> 	* ggc-page.c (RTL_SIZE): Use RTX_HDR_SIZE.
> 	* reload1.c (eliminate_regs): Use RTX_SIZE.
> 	* rtlanal.c (loc_mentioned_in_p): Adjust for new rtx_def layout.
> 	* gdbinit.in (pi): Likewise.

This is OK.

In general, I agree with Gerald; patches whose sole effect is to speed
up the compiler (as long as they're not too invasive) are acceptable for
stage 3.

zw