https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114729
--- Comment #16 from Vineet Gupta <vineetg at gcc dot gnu.org> ---
After ECC hack, the issue persists.
Toggles (for cc1plus): -O2 -march=rv64gc_zfa -mabi=lp64d
%sfp is the spill because
-fno-schedule-insns | -fschedule-insns
|
li t3,2 | li t5,2
li t1,1 | li t4,1
---xxx--- | sd s10,8(sp) # %sfp
.L2: | .L2:
---xxx--- | ld a5,8(sp) # %sfp
beq s10,zero,.L9 | beq a5,zero,.L9
mv a1,s6 | mv a1,a6
beq s6,zero,.L7 | beq a6,zero,.L7
mulw a2,s5,s6 | mulw a2,a0,a6
mv a0,s5 | mv s10,a0
slli a6,s5,3 | slli s11,a0,3
slli a3,a2,3 | slli a3,a2,3
.L6: | .L6:
fcvt.d.w fa5,a2 # 56 | fcvt.d.w fa5,a2 # 56
add a5,s4,a3 # 63 | add a4,s5,a3 # 63
fsd fa5,%lo(e)(t6) # 57 | add a5,s6,a3 # 58
fld fa4,0(a5) # 64 | fsd fa5,%lo(e)(t2) # 57
fld fa5,%lo(o)(t5) | fld fa3,0(a4) # 64
add a5,s3,a3 # 58 | fld fa4,0(a5) # 60
fmul.d fa4,fa4,fa5 | fld fa5,%lo(o)(t0)
fld fa5,0(a5) # 60 | fcvt.l.d a4,fa3,rtz
schedule_insn () is called as follows:
;; 0--> b 0: i 56 r210=flt(r185#0) :alu: GR_REGS+0(0) FP_REGS+1(1):model 0
;; 1--> b 0: i 63 r215=r141+r183 :alu: GR_REGS+1(1) FP_REGS+0(0)
;; 2--> b 0: i 58 r211=r138+r183 :alu:@GR_REGS+1(1)@FP_REGS+0(0)
The prints between insn 63 and insn 58 show the issue
;; point uid prio delay class : del ECC-gpr del ECC-fpr
ECC-tot
;;| 1 57 | 12 +2 | GR_REGS:[0 base cost 0] FP_REGS:[-1 base cost 0]ECC 0
;;| 10 61 | 4 +1 | GR_REGS:[0 base cost 0] FP_REGS:[1 base cost 0] ECC 0
;;| 8 58 | 5 +1 | GR_REGS:[1 base cost 0] FP_REGS:[0 base cost 0] ECC 0
;; --1-- RFS_PRESSURE_DELAY winner 58 looser 57
;; --2-- RFS_PRESSURE_DELAY winner 58 looser 57
1, insn 57 delta is -1 (reduces pressure), while insn 58 delta is 1 (increases
pressure) yet ECC is 0 for both, meaning for pressure considerations they are
treated the same. This ECC "dilution" happens in model_spill_cost ().
2. insn 57 is a store (so prio 2) while insn 58 is an add (so prio 1).
rank_for_schedule (): RFS_PRESSURE_DELAY chooses lower of (ECC + insn_delay)
and thus picks 58 over 57 causing the issue.
