https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83272
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|normal |enhancement
Last reconfirmed| |2021-08-14
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Mason from comment #2)
> Anyway, I may have oversimplified the testcase. Consider this one:
>
> char foo(unsigned char *p)
> {
> static const char map[16] = "wxyz";
> return map[*p / 16];
> }
For this case we start in before combine:
(insn 8 7 9 2 (parallel [
(set (reg:QI 89)
(lshiftrt:QI (reg:QI 90 [ *p_5(D) ])
(const_int 4 [0x4])))
(clobber (reg:CC 17 flags))
]) "/app/example.cpp":5:16 729 {*lshrqi3_1}
(expr_list:REG_DEAD (reg:QI 90 [ *p_5(D) ])
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil))))
(insn 9 8 10 2 (set (reg:SI 91)
(zero_extend:SI (reg:QI 89))) "/app/example.cpp":5:16 141
{*zero_extendqisi2}
(expr_list:REG_DEAD (reg:QI 89)
(nil)))
(insn 10 9 11 2 (set (reg:DI 92)
(sign_extend:DI (reg:SI 91))) "/app/example.cpp":5:20 145
{*extendsidi2_rex64}
(expr_list:REG_DEAD (reg:SI 91)
(nil)))
after:
(insn 8 7 9 2 (parallel [
(set (reg:QI 89)
(lshiftrt:QI (reg:QI 90 [ *p_5(D) ])
(const_int 4 [0x4])))
(clobber (reg:CC 17 flags))
]) "/app/example.cpp":5:16 729 {*lshrqi3_1}
(expr_list:REG_DEAD (reg:QI 90 [ *p_5(D) ])
(expr_list:REG_UNUSED (reg:CC 17 flags)
(nil))))
(note 9 8 10 2 NOTE_INSN_DELETED)
(insn 10 9 11 2 (parallel [
(set (reg:DI 92)
(and:DI (subreg:DI (reg:QI 89) 0)
(const_int 15 [0xf])))
(clobber (reg:CC 17 flags))
]) "/app/example.cpp":5:20 498 {*anddi_1}
(expr_list:REG_UNUSED (reg:CC 17 flags)
(expr_list:REG_DEAD (reg:QI 89)
(nil))))
And that is due to knowing the nonzero bits of reg 89 (after the shift):
Trying 9 -> 10:
9: r91:SI=zero_extend(r89:QI)
REG_DEAD r89:QI
10: r92:DI=sign_extend(r91:SI)
REG_DEAD r91:SI
Successfully matched this instruction:
(set (reg:DI 92)
(and:DI (subreg:DI (reg:QI 89) 0)
(const_int 15 [0xf])))
allowing combination of insns 9 and 10
original costs 4 + 4 = 8
replacement cost 4
deferring deletion of insn with uid = 9.
modifying insn i3 10: {r92:DI=r89:QI#0&0xf;clobber flags:CC;}
REG_UNUSED flags:CC
REG_DEAD r89:QI
deferring rescan insn with uid = 10.
------- CUT -----
So combining is over using the nonzero bits sometimes.
