Skip to content
This repository was archived by the owner on Apr 23, 2020. It is now read-only.

Commit d45c0f1

Browse files
dylanmckaydylanmckay
committed
[AVR] Remove the earlyclobber flag from LDDWRdYQ
Before I started maintaining the AVR backend, this instruction never originally used to have an earlyclobber flag. Some time afterwards (years ago), I must've added it back in, not realising that it was left out for a reason. This pseudo instrction exists solely to work around a long standing bug in the register allocator. Before this commit, the LDDWRdYQ pseudo was not actually working around any bug. With the earlyclobber flag removed again, the LDDWRdYQ pseudo now correctly works around PR13375 again. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@326774 91177308-0d34-0410-b5e6-96231b3b80d8
1 parent 783006e commit d45c0f1

File tree

2 files changed

+155
-2
lines changed

2 files changed

+155
-2
lines changed

‎lib/Target/AVR/AVRInstrInfo.td

+16-2
Original file line numberDiff line numberDiff line change
@@ -1228,9 +1228,23 @@ isReMaterializable = 1 in
12281228
[(set i16:$dst, (load addr:$memri))]>,
12291229
Requires<[HasSRAM]>;
12301230

1231+
// An identical pseudo instruction to LDDWRdPtrQ, expect restricted to the Y
1232+
// register and without the @earlyclobber flag.
1233+
//
1234+
// Used to work around a bug caused by the register allocator not
1235+
// being able to handle the expansion of a COPY into an machine instruction
1236+
// that has an earlyclobber flag. This is because the register allocator will
1237+
// try expand a copy from a register slot into an earlyclobber instruction.
1238+
// Instructions that are earlyclobber need to be in a dedicated earlyclobber slot.
1239+
//
1240+
// This pseudo instruction can be used pre-AVR pseudo expansion in order to
1241+
// get a frame index load without directly using earlyclobber instructions.
1242+
//
1243+
// The pseudo expansion pass trivially expands this into LDDWRdPtrQ.
1244+
//
1245+
// This instruction may be removed once PR13375 is fixed.
12311246
let mayLoad = 1,
1232-
hasSideEffects = 0,
1233-
Constraints = "@earlyclobber $dst" in
1247+
hasSideEffects = 0 in
12341248
def LDDWRdYQ : Pseudo<(outs DREGS:$dst),
12351249
(ins memri:$memri),
12361250
"lddw\t$dst, $memri",

‎test/CodeGen/AVR/issue-regalloc-stackframe-folding-earlyclobber.ll

+139
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,139 @@
1+
; RUN: llc < %s -march=avr -mcpu=atmega328 | FileCheck %s
2+
3+
; This test case is designed to trigger a bug caused by the register
4+
; allocator not handling the case where a target generates a load/store with
5+
; a frame index that happens to be an earlyclobber instruction.
6+
;
7+
; It is taken from the translated LLVM IR of the Rust core library.
8+
;
9+
; The error message looked like
10+
; Assertion failed: (MI && "No instruction defining live value"), function computeDeadValues
11+
;
12+
; See PR13375 for details.
13+
14+
%values = type { i16, [2 x i8], [40 x i32], [0 x i8] }
15+
16+
@POW10TO128 = external constant [14 x i32], align 4
17+
18+
; CHECK: core_num_flt2dec_strategy_dragon
19+
define nonnull dereferenceable(164) %values* @core_num_flt2dec_strategy_dragon(%values* returned dereferenceable(164) %arg, i16 %arg1) unnamed_addr {
20+
start:
21+
%ret.i = alloca [40 x i32], align 4
22+
%tmp = icmp eq i16 undef, 0
23+
br i1 %tmp, label %bb5, label %bb1
24+
25+
bb1:
26+
unreachable
27+
28+
bb5:
29+
br i1 undef, label %bb14, label %bb11
30+
31+
bb11:
32+
unreachable
33+
34+
bb14:
35+
%tmp2 = bitcast [40 x i32]* %ret.i to i8*
36+
call void @llvm.memset.p0i8.i16(i8* align 4 %tmp2, i8 0, i16 160, i1 false)
37+
%tmp3 = getelementptr inbounds %values, %values* %arg, i16 0, i32 0
38+
%tmp4 = load i16, i16* %tmp3, align 2
39+
%tmp5 = icmp ult i16 %tmp4, 14
40+
%tmp6 = getelementptr inbounds %values, %values* %arg, i16 0, i32 2, i16 0
41+
br i1 %tmp5, label %bb2.i38, label %bb3.i39
42+
43+
bb2.i38:
44+
%tmp7 = getelementptr inbounds %values, %values* %arg, i16 0, i32 2, i16 %tmp4
45+
%tmp8 = ptrtoint i32* %tmp6 to i16
46+
br label %bb4.outer.i122
47+
48+
bb4.outer.i122:
49+
%iter.sroa.0.0.ph.i119 = phi i16 [ %tmp12, %bb24.i141 ], [ %tmp8, %bb2.i38 ]
50+
%retsz.0.ph.i121 = phi i16 [ %.retsz.0.i140, %bb24.i141 ], [ 0, %bb2.i38 ]
51+
br label %bb4.i125
52+
53+
bb4.i125:
54+
%iter.sroa.0.0.i123 = phi i16 [ %tmp12, %core.iter.Enumerate.ALPHA ], [ %iter.sroa.0.0.ph.i119, %bb4.outer.i122 ]
55+
%tmp9 = inttoptr i16 %iter.sroa.0.0.i123 to i32*
56+
%tmp10 = icmp eq i32* %tmp9, %tmp7
57+
br i1 %tmp10, label %core.num.bignum.Big32x40.exit44, label %core.iter.Enumerate.ALPHA
58+
59+
core.iter.Enumerate.ALPHA:
60+
%tmp11 = getelementptr inbounds i32, i32* %tmp9, i16 1
61+
%tmp12 = ptrtoint i32* %tmp11 to i16
62+
%tmp13 = load i32, i32* %tmp9, align 4
63+
%tmp14 = icmp eq i32 %tmp13, 0
64+
br i1 %tmp14, label %bb4.i125, label %core..iter..Enumerate.exit17
65+
66+
core..iter..Enumerate.exit17:
67+
%tmp15 = zext i32 %tmp13 to i64
68+
br label %core..iter..Enumerate.exit17.i132
69+
70+
core..iter..Enumerate.exit17.i132:
71+
%carry.085.i129 = phi i32 [ 0, %core..iter..Enumerate.exit17 ], [ %tmp28, %core_slice_IndexMut.exit13 ]
72+
%tmp16 = icmp ult i16 undef, 40
73+
br i1 %tmp16, label %core_slice_IndexMut.exit13, label %panic.i.i14.i134
74+
75+
bb16.i133:
76+
%tmp17 = icmp eq i32 %tmp28, 0
77+
br i1 %tmp17, label %bb24.i141, label %bb21.i136
78+
79+
panic.i.i14.i134:
80+
unreachable
81+
82+
core_slice_IndexMut.exit13:
83+
%tmp18 = load i32, i32* null, align 4
84+
%tmp19 = getelementptr inbounds [40 x i32], [40 x i32]* %ret.i, i16 0, i16 undef
85+
%tmp20 = load i32, i32* %tmp19, align 4
86+
%tmp21 = zext i32 %tmp18 to i64
87+
%tmp22 = mul nuw i64 %tmp21, %tmp15
88+
%tmp23 = zext i32 %tmp20 to i64
89+
%tmp24 = zext i32 %carry.085.i129 to i64
90+
%tmp25 = add nuw nsw i64 %tmp23, %tmp24
91+
%tmp26 = add i64 %tmp25, %tmp22
92+
%tmp27 = lshr i64 %tmp26, 32
93+
%tmp28 = trunc i64 %tmp27 to i32
94+
%tmp29 = icmp eq i32* undef, getelementptr inbounds ([14 x i32], [14 x i32]* @POW10TO128, i16 1, i16 0)
95+
br i1 %tmp29, label %bb16.i133, label %core..iter..Enumerate.exit17.i132
96+
97+
bb21.i136:
98+
%tmp30 = icmp ult i16 undef, 40
99+
br i1 %tmp30, label %"_ZN4core5slice70_$LT$impl$u20$core..ops..IndexMut$LT$I$GT$$u20$for$u20$$u5b$T$u5d$$GT$9index_mut17h8eccc0af1ec6f971E.exit.i138", label %panic.i.i.i137
100+
101+
panic.i.i.i137:
102+
unreachable
103+
104+
"_ZN4core5slice70_$LT$impl$u20$core..ops..IndexMut$LT$I$GT$$u20$for$u20$$u5b$T$u5d$$GT$9index_mut17h8eccc0af1ec6f971E.exit.i138":
105+
store i32 %tmp28, i32* undef, align 4
106+
br label %bb24.i141
107+
108+
bb24.i141:
109+
%sz.0.i139 = phi i16 [ 15, %"_ZN4core5slice70_$LT$impl$u20$core..ops..IndexMut$LT$I$GT$$u20$for$u20$$u5b$T$u5d$$GT$9index_mut17h8eccc0af1ec6f971E.exit.i138" ], [ 14, %bb16.i133 ]
110+
%tmp31 = add i16 %sz.0.i139, 0
111+
%tmp32 = icmp ult i16 %retsz.0.ph.i121, %tmp31
112+
%.retsz.0.i140 = select i1 %tmp32, i16 %tmp31, i16 %retsz.0.ph.i121
113+
br label %bb4.outer.i122
114+
115+
bb3.i39:
116+
%tmp33 = call fastcc i16 @_ZN4core3num6bignum8Big32x4010mul_digits9mul_inner17h5d3461bce04d16ccE([40 x i32]* nonnull dereferenceable(160) %ret.i, i32* noalias nonnull readonly getelementptr inbounds ([14 x i32], [14 x i32]* @POW10TO128, i16 0, i16 0), i16 14, i32* noalias nonnull readonly %tmp6, i16 %tmp4)
117+
br label %core.num.bignum.Big32x40.exit44
118+
119+
core.num.bignum.Big32x40.exit44:
120+
%retsz.0.i40 = phi i16 [ %tmp33, %bb3.i39 ], [ %retsz.0.ph.i121, %bb4.i125 ]
121+
call void @llvm.memcpy.p0i8.p0i8.i16(i8* align 4 undef, i8* align 4 %tmp2, i16 160, i1 false)
122+
store i16 %retsz.0.i40, i16* %tmp3, align 2
123+
%tmp34 = and i16 %arg1, 256
124+
%tmp35 = icmp eq i16 %tmp34, 0
125+
br i1 %tmp35, label %bb30, label %bb27
126+
127+
bb27:
128+
unreachable
129+
130+
bb30:
131+
ret %values* %arg
132+
}
133+
134+
declare fastcc i16 @_ZN4core3num6bignum8Big32x4010mul_digits9mul_inner17h5d3461bce04d16ccE([40 x i32]* nocapture dereferenceable(160), i32* noalias nonnull readonly, i16, i32* noalias nonnull readonly, i16)
135+
136+
declare void @llvm.memset.p0i8.i16(i8* nocapture writeonly, i8, i16, i1)
137+
138+
declare void @llvm.memcpy.p0i8.p0i8.i16(i8* nocapture writeonly, i8* nocapture readonly, i16, i1)
139+

0 commit comments

Comments
 (0)