https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97387
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |uros at gcc dot gnu.org
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I actually think that what we emit for these builtins is right, the problem is
that combiner is not able to optimize
(insn 10 9 11 2 (set (reg:QI 88 [ _31 ])
(ltu:QI (reg:CCC 17 flags)
(const_int 0 [0]))) "include/adxintrin.h":69:10 784 {*setcc_qi}
(expr_list:REG_DEAD (reg:CCC 17 flags)
(nil)))
followed (with instructions that don't clobber flags in between) by:
(insn 17 15 18 2 (parallel [
(set (reg:CCC 17 flags)
(compare:CCC (plus:QI (reg:QI 88 [ _31 ])
(const_int -1 [0xffffffffffffffff]))
(reg:QI 88 [ _31 ])))
(clobber (scratch:QI))
]) "include/adxintrin.h":69:10 349 {*addqi3_cconly_overflow_1}
(expr_list:REG_DEAD (reg:QI 88 [ _31 ])
(nil)))
into nothing. It tries that:
Trying 10 -> 17:
10: r88:QI=ltu(flags:CCC,0)
REG_DEAD flags:CCC
17: {flags:CCC=cmp(r88:QI-0x1,r88:QI);clobber scratch;}
REG_DEAD r88:QI
Failed to match this instruction:
(parallel [
(set (reg:CC 17 flags)
(compare:CC (neg:QI (geu:QI (reg:CCC 17 flags)
(const_int 0 [0])))
(ltu:QI (reg:CCC 17 flags)
(const_int 0 [0]))))
(clobber (scratch:QI))
])
Failed to match this instruction:
(set (reg:CC 17 flags)
(compare:CC (neg:QI (geu:QI (reg:CCC 17 flags)
(const_int 0 [0])))
(ltu:QI (reg:CCC 17 flags)
(const_int 0 [0]))))
Similarly, the
Trying 10, 17 -> 18:
10: r88:QI=ltu(flags:CCC,0)
REG_DEAD flags:CCC
17: {flags:CCC=cmp(r88:QI-0x1,r88:QI);clobber scratch;}
REG_DEAD r88:QI
18:
{flags:CCC=cmp(zero_extend(ltu(flags:CCC,0)+r106:DI+r107:DI),zero_extend(r107:DI)+ltu(flags:CCC,0));r109:DI=ltu(flags:CCC,0)+r106:DI+r107:DI
;}
REG_DEAD r107:DI
REG_DEAD r106:DI
Can't combine i1 into i3
fails because it would want to set flags multiple times and punts because of
that.
The 10 -> 17 combination seems more promissing, though I'm not sure the CCmode
rather than CCCmode in that case is desirable.
