https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100922
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> Pre takes:
> <bb 5> [local count: 1073741824]:
> # str_9 = PHI <str_27(4), str_16(8)>
> # out_10 = PHI <str_27(4), out_20(8)>
> str_16 = str_9 + 1;
> c_17 = *str_9;
> _5 = (sizetype) c_17;
> _6 = keep_lut_14(D) + _5;
> inc_18 = *_6;
> *out_10 = c_17;
> _7 = (sizetype) inc_18;
> out_20 = out_10 + _7;
> if (c_17 != 0)
> goto <bb 8>; [89.00%]
> else
> goto <bb 6>; [11.00%]
>
> <bb 8> [local count: 955630224]:
> goto <bb 5>; [100.00%]
>
> Changes it to:
> <bb 5> [local count: 1073741824]:
> # str_9 = PHI <str_27(4), str_16(6)>
> # out_10 = PHI <str_27(4), out_20(6)>
> # prephitmp_60 = PHI <prephitmp_54(4), pretmp_59(6)>
> str_16 = str_9 + 1;
> c_17 = *str_9;
> _5 = (sizetype) c_17;
> _6 = keep_lut_14(D) + _5;
> *out_10 = c_17;
> _7 = (sizetype) prephitmp_60;
> out_20 = out_10 + _7;
> if (c_17 != 0)
> goto <bb 6>; [89.00%]
> else
> goto <bb 7>; [11.00%]
>
> <bb 6> [local count: 955630224]:
> pretmp_55 = MEM[(char *)str_9 + 1B];
> _57 = (sizetype) pretmp_55;
> _58 = keep_lut_14(D) + _57;
> pretmp_59 = *_58;
> goto <bb 5>; [100.00%]
>
> And then ch_vect changes it to (basically undoing the PRE):
> <bb 14> [local count: 850510900]:
>
> <bb 6> [local count: 955630224]:
> # str_42 = PHI <str_41(14), str_27(13)>
> # str_41 = PHI <str_16(14), str_47(13)>
> # out_38 = PHI <out_20(14), out_43(13)>
> pretmp_55 = MEM[(char *)str_42 + 1B];
> _57 = (sizetype) pretmp_55;
> _58 = keep_lut_14(D) + _57;
> pretmp_59 = *_58;
> str_16 = str_41 + 1;
> c_17 = *str_41;
> *out_38 = c_17;
> _7 = (sizetype) pretmp_59;
> out_20 = out_38 + _7;
> if (c_17 != 0)
> goto <bb 14>; [89.00%]
> else
> goto <bb 7>; [11.00%]
>
> And then IVOPTS comes around and does:
> <bb 14> [local count: 850510900]:
>
> <bb 6> [local count: 955630224]:
> # str_41 = PHI <str_16(14), str_47(13)>
> # out_38 = PHI <out_20(14), out_43(13)>
> pretmp_55 = MEM[(char *)str_41];
> _57 = (sizetype) pretmp_55;
> _58 = keep_lut_14(D) + _57;
> pretmp_59 = *_58;
> str_16 = str_41 + 1;
> c_17 = MEM[(char *)str_16 + -1B];
> *out_38 = c_17;
> _7 = (sizetype) pretmp_59;
> out_20 = out_38 + _7;
> if (c_17 != 0)
> goto <bb 14>; [89.00%]
> else
> goto <bb 7>; [11.00%]
>
> Which is fine except FRE (and DOM), don't recognize the MEM[(char *)str_16 +
> -1B] and MEM[(char *)str_41] being the same. For FRE, it almost looks like
> TARGET_MEM_REF is not handled ....
TARGET_MEM_REF is handled but what is not is "forwarding" the str_16 definition
to the str_16 + -1B TARGET_MEM_REF - vn_reference_maybe_forwprop_address only
forwards to MEM_REFs (see the caller in valueize_refs_1 - the TARGET_MEM_REF
representation is different enough to make a trivial patch impossible).
Now, there's code to fold TARGET_MEM_REF back to MEM_REF, but it doesn't
trigger here as
/* If possible use a plain MEM_REF instead of a TARGET_MEM_REF.
??? As IVOPTs does not follow restrictions to where the base
pointer may point to create a MEM_REF only if we know that
base is valid. */
if ((TREE_CODE (base) == ADDR_EXPR || TREE_CODE (base) == INTEGER_CST)
I _think_ that IVOPTs is fixed, but I'm not 100% sure ...