https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95382
Bug ID: 95382
Summary: [haifa-sched][DO_PREDICATION] execution test: wrong
scheduling result.
Product: gcc
Version: 7.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: xuemaosheng at huawei dot com
Target Milestone: ---
Created attachment 48626
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48626&action=edit
the source file
After we enable the schedule DO_PREDICATION, we get wrong scheduling result in
sched2 pass.
The key dump is shown as following：
.......(Unimportant things)
;; | 93 | 15 | a20=sxn([a19])
;; | 94 | 10 | t2=a20==0
;; | 95 | 9 | pc={(t2!=0)?L107:pc} jump_insn1
;; | 159 | 10 | a21=0xffffffffffff8000
;; | 98 | 9 | t0=a23!=a21
;; | 99 | 8 | pc={(t0==0)?L173:pc} jump_insn2
;; | 105 | 7 | a22=a23/a20 div_insn
;; | 106 | 1 | a23=sxn(a22)
.......(Unimportant things)
;; --- EBB Dependences --- from bb4 to bb6
;; insn code bb dep prio cost reservation
.......(Unimportant things)
;; 93 94 4 4 15 1 : 105(t,1) 99(a,0) 94(t,5)
;; 94 1645 4 1 10 1 : 106(t,1) 95(t,1)
;; 95 139 4 13 9 1 : 106(c,1) 99(a,1)nm
;; 159 129 5 0 10 1 : 98(t,1)
;; 98 1650 5 2 9 1 : 105(t,1) 106(t,1)m 99(t,1)
;; 99 139 5 15 8 1 : 105(c,1) 106(c,1)
;; 105 37 6 4 7 1 : 106(t,6)m
;; 106 94 6 6 1 1 :
.......(Unimportant things)
When we enable the schedule DO_PREDICATION, we can produce conditional
execution insn.
Since insn 98(t0=a23!=a21) doesn't have dependence with jump_insn1 95
(pc={(t2!=0)?L107:pc}), so insn 98 can issue before jump_insn1 95.
And insn 93 (a20=sxn([a19])) can also issue before jump_insn1 95.
After that, insn 105(a22=a23/a20) can become conditional execution insn shown
as following dump:
;; 51--> 93 a20=sxn([a19]) :
rescanning insn with uid = 105.
;; dependencies resolved: insn 105 predicated
;; Ready-->Q: insn 105: queued for 1 cycles (change queue
index).
;; tick updated: insn 105 into queue with cost=1
After predicating, insn 105 have become : [t0]a22=a23/a20
;; 52--> 105 (t0) a22=a23/a20 :
;; Ready list (t = 52):
.......(Unimportant things)
;; Ready list (t = 56): 94:50:prio=10
[52;56]:94
;; 56--> 94 t2=a20==0 :
.......(Unimportant things)
;; Ready list (t = 57): 95:51:prio=9
[56;57]:95
;; 57--> 95 pc={(t2!=0)?L107:pc} :
finally, insn 105([t0]a22=a23/a20) is issued before insn 95
(pc={(t2!=0)?L107:pc}).
However, insn 95 has decided reg a20 is valid not.
In this case, a20 is 0, so we get division by 0 error.
The scheduling result is shown as the following:
tstneq@ags t0, a23, a21 # [28] # 98
....
l16si@agl a20, a19, 0 # [51] # 93
....
[t0] quos@sau a22, a23, a20 # [52] # 105
tsteqi@sau t2, a20, 0 # [53] # 94
[t2] b@pcu .L5, 1 # [54] # 95
[!t0] b@pcu .L28, 0 # [55] # 99
......
Insn 105 (div_insn)should not be issued before insn 95(jump_insn1).
Note： This ebb is 3 basic-blocks.
Since supporting DO_PREDICATION target is little, it's difficult to recurrent
in other target.
How can we solve the problem?
