https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78604
--- Comment #3 from amker at gcc dot gnu.org ---
For function sign_lt and uns_lt, the change causes worse code generation
unfortunately. Take uns_lt as example, the difference in optimized dump is as
like:
529,530c529,530
< vect_cst__32 = {val1_8(D), val1_8(D)};
< vect_cst__33 = {val2_7(D), val2_7(D)};
---
> vect_cst__32 = {val2_7(D), val2_7(D)};
> vect_cst__33 = {val1_8(D), val1_8(D)};
536c536
< vect_iftmp.412_34 = VEC_COND_EXPR <vect__1.408_28 < vect__2.411_31,
vect_cst__32, vect_cst__33>;
---
> vect_iftmp.412_34 = VEC_COND_EXPR <vect__1.408_28 >= vect__2.411_31,
> vect_cst__32, vect_cst__33>;
Condition "<" is transformed into ">=" with the change, which results in more
insns after expanding:
***************
*** 1014,1024 ****
REG_EQUAL `uc'
49: r199:V2DI=[r195:SI+r167:SI]
50: r200:V2DI=[r197:SI+r167:SI]
! 51: r202:V2DI=gtu(r200:V2DI,r199:V2DI)
! 52: r201:V2DI={(r202:V2DI!=const_vector)?r170:V2DI:r171:V2DI}
! 53: [r193:SI+r167:SI]=r201:V2DI
! 54: r167:SI=r167:SI+0x10
! 56: r203:CCUNS=cmp(r167:SI,0x2000)
! 57: pc={(r203:CCUNS!=0)?L55:pc}
REG_BR_PROB 9899
! 58: NOTE_INSN_BASIC_BLOCK 5
--- 1016,1028 ----
REG_EQUAL `uc'
49: r199:V2DI=[r195:SI+r167:SI]
50: r200:V2DI=[r197:SI+r167:SI]
! 51: r202:V2DI=gtu(r199:V2DI,r200:V2DI)
! 52: r203:V2DI=r199:V2DI==r200:V2DI
! 53: r204:V2DI=r202:V2DI|r203:V2DI
! 54: r201:V2DI={(r204:V2DI!=const_vector)?r170:V2DI:r171:V2DI}
! 55: [r193:SI+r167:SI]=r201:V2DI
! 56: r167:SI=r167:SI+0x10
! 58: r205:CCUNS=cmp(r167:SI,0x2000)
! 59: pc={(r205:CCUNS!=0)?L57:pc}
REG_BR_PROB 9899
! 60: NOTE_INSN_BASIC_BLOCK 5
So "<" is expanded into instruction (51), but ">=" is expanded into
instructions (51~53).
So powerpc has to compute ">=" with "> || =="? Apart from this, the change
itself only introduces additional canonicalization during tree-ifcvt,
transforming "<" to "<=". If the former form is more efficient on powerpc,
shall we take this into consideration when canonicalization?
Thanks.
