Skip to content

Commit 63f3a5b

Browse files
RKSimonRKSimon
committed
[PhaseOrdering][X86] Add test coverage for #48223
The X86 backend shuffle combining is saving us from some poor vectorised IR
1 parent 4b89339 commit 63f3a5b

File tree

1 file changed

+163
-0
lines changed

1 file changed

+163
-0
lines changed

‎llvm/test/Transforms/PhaseOrdering/X86/pr48223.ll

Lines changed: 163 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,163 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt -O3 -S -mtriple=x86_64-- -mcpu=x86-64 < %s | FileCheck %s --check-prefixes=CHECK,SSE,SSE2
3+
; RUN: opt -O3 -S -mtriple=x86_64-- -mcpu=x86-64-v2 < %s | FileCheck %s --check-prefixes=CHECK,SSE,SSE4
4+
; RUN: opt -O3 -S -mtriple=x86_64-- -mcpu=x86-64-v3 < %s | FileCheck %s --check-prefixes=CHECK,AVX
5+
; RUN: opt -O3 -S -mtriple=x86_64-- -mcpu=x86-64-v4 < %s | FileCheck %s --check-prefixes=CHECK,AVX
6+
7+
%"struct.std::array" = type { [8 x i16] }
8+
9+
define { i64, i64 } @compute_min(ptr noundef nonnull align 2 dereferenceable(16) %x, ptr noundef nonnull align 2 dereferenceable(16) %y) {
10+
; SSE2-LABEL: @compute_min(
11+
; SSE2-NEXT: entry:
12+
; SSE2-NEXT: [[LD0:%.*]] = load i16, ptr [[Y:%.*]], align 2
13+
; SSE2-NEXT: [[LD1:%.*]] = load i16, ptr [[X:%.*]], align 2
14+
; SSE2-NEXT: [[LD2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0]], i16 [[LD1]])
15+
; SSE2-NEXT: [[PT1_1:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 2
16+
; SSE2-NEXT: [[PT0_1:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 2
17+
; SSE2-NEXT: [[LD0_1:%.*]] = load i16, ptr [[PT0_1]], align 2
18+
; SSE2-NEXT: [[LD1_1:%.*]] = load i16, ptr [[PT1_1]], align 2
19+
; SSE2-NEXT: [[LD2_1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_1]], i16 [[LD1_1]])
20+
; SSE2-NEXT: [[PT1_2:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 4
21+
; SSE2-NEXT: [[PT0_2:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 4
22+
; SSE2-NEXT: [[LD0_2:%.*]] = load i16, ptr [[PT0_2]], align 2
23+
; SSE2-NEXT: [[LD1_2:%.*]] = load i16, ptr [[PT1_2]], align 2
24+
; SSE2-NEXT: [[LD2_2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_2]], i16 [[LD1_2]])
25+
; SSE2-NEXT: [[PT1_3:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 6
26+
; SSE2-NEXT: [[PT0_3:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 6
27+
; SSE2-NEXT: [[LD0_3:%.*]] = load i16, ptr [[PT0_3]], align 2
28+
; SSE2-NEXT: [[LD1_3:%.*]] = load i16, ptr [[PT1_3]], align 2
29+
; SSE2-NEXT: [[LD2_3:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_3]], i16 [[LD1_3]])
30+
; SSE2-NEXT: [[PT1_4:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 8
31+
; SSE2-NEXT: [[PT0_4:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 8
32+
; SSE2-NEXT: [[LD0_4:%.*]] = load i16, ptr [[PT0_4]], align 2
33+
; SSE2-NEXT: [[LD1_4:%.*]] = load i16, ptr [[PT1_4]], align 2
34+
; SSE2-NEXT: [[LD2_4:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_4]], i16 [[LD1_4]])
35+
; SSE2-NEXT: [[PT1_5:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 10
36+
; SSE2-NEXT: [[PT0_5:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 10
37+
; SSE2-NEXT: [[LD0_5:%.*]] = load i16, ptr [[PT0_5]], align 2
38+
; SSE2-NEXT: [[LD1_5:%.*]] = load i16, ptr [[PT1_5]], align 2
39+
; SSE2-NEXT: [[LD2_5:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_5]], i16 [[LD1_5]])
40+
; SSE2-NEXT: [[PT1_6:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 12
41+
; SSE2-NEXT: [[PT0_6:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 12
42+
; SSE2-NEXT: [[LD0_6:%.*]] = load i16, ptr [[PT0_6]], align 2
43+
; SSE2-NEXT: [[LD1_6:%.*]] = load i16, ptr [[PT1_6]], align 2
44+
; SSE2-NEXT: [[LD2_6:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_6]], i16 [[LD1_6]])
45+
; SSE2-NEXT: [[PT1_7:%.*]] = getelementptr inbounds nuw i8, ptr [[X]], i64 14
46+
; SSE2-NEXT: [[PT0_7:%.*]] = getelementptr inbounds nuw i8, ptr [[Y]], i64 14
47+
; SSE2-NEXT: [[LD0_7:%.*]] = load i16, ptr [[PT0_7]], align 2
48+
; SSE2-NEXT: [[LD1_7:%.*]] = load i16, ptr [[PT1_7]], align 2
49+
; SSE2-NEXT: [[LD2_7:%.*]] = tail call i16 @llvm.smin.i16(i16 [[LD0_7]], i16 [[LD1_7]])
50+
; SSE2-NEXT: [[RETVAL_SROA_4_0_INSERT_EXT:%.*]] = zext i16 [[LD2_3]] to i64
51+
; SSE2-NEXT: [[RETVAL_SROA_4_0_INSERT_SHIFT:%.*]] = shl nuw i64 [[RETVAL_SROA_4_0_INSERT_EXT]], 48
52+
; SSE2-NEXT: [[RETVAL_SROA_3_0_INSERT_EXT:%.*]] = zext i16 [[LD2_2]] to i64
53+
; SSE2-NEXT: [[RETVAL_SROA_3_0_INSERT_SHIFT:%.*]] = shl nuw nsw i64 [[RETVAL_SROA_3_0_INSERT_EXT]], 32
54+
; SSE2-NEXT: [[RETVAL_SROA_3_0_INSERT_INSERT:%.*]] = or disjoint i64 [[RETVAL_SROA_4_0_INSERT_SHIFT]], [[RETVAL_SROA_3_0_INSERT_SHIFT]]
55+
; SSE2-NEXT: [[RETVAL_SROA_2_0_INSERT_EXT:%.*]] = zext i16 [[LD2_1]] to i64
56+
; SSE2-NEXT: [[RETVAL_SROA_2_0_INSERT_SHIFT:%.*]] = shl nuw nsw i64 [[RETVAL_SROA_2_0_INSERT_EXT]], 16
57+
; SSE2-NEXT: [[RETVAL_SROA_2_0_INSERT_INSERT:%.*]] = or disjoint i64 [[RETVAL_SROA_3_0_INSERT_INSERT]], [[RETVAL_SROA_2_0_INSERT_SHIFT]]
58+
; SSE2-NEXT: [[RETVAL_SROA_0_0_INSERT_EXT:%.*]] = zext i16 [[LD2]] to i64
59+
; SSE2-NEXT: [[TMP20:%.*]] = or disjoint i64 [[RETVAL_SROA_2_0_INSERT_INSERT]], [[RETVAL_SROA_0_0_INSERT_EXT]]
60+
; SSE2-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue { i64, i64 } poison, i64 [[TMP20]], 0
61+
; SSE2-NEXT: [[RETVAL_SROA_9_8_INSERT_EXT:%.*]] = zext i16 [[LD2_7]] to i64
62+
; SSE2-NEXT: [[RETVAL_SROA_9_8_INSERT_SHIFT:%.*]] = shl nuw i64 [[RETVAL_SROA_9_8_INSERT_EXT]], 48
63+
; SSE2-NEXT: [[RETVAL_SROA_8_8_INSERT_EXT:%.*]] = zext i16 [[LD2_6]] to i64
64+
; SSE2-NEXT: [[RETVAL_SROA_8_8_INSERT_SHIFT:%.*]] = shl nuw nsw i64 [[RETVAL_SROA_8_8_INSERT_EXT]], 32
65+
; SSE2-NEXT: [[RETVAL_SROA_8_8_INSERT_INSERT:%.*]] = or disjoint i64 [[RETVAL_SROA_9_8_INSERT_SHIFT]], [[RETVAL_SROA_8_8_INSERT_SHIFT]]
66+
; SSE2-NEXT: [[RETVAL_SROA_7_8_INSERT_EXT:%.*]] = zext i16 [[LD2_5]] to i64
67+
; SSE2-NEXT: [[RETVAL_SROA_7_8_INSERT_SHIFT:%.*]] = shl nuw nsw i64 [[RETVAL_SROA_7_8_INSERT_EXT]], 16
68+
; SSE2-NEXT: [[RETVAL_SROA_7_8_INSERT_INSERT:%.*]] = or disjoint i64 [[RETVAL_SROA_8_8_INSERT_INSERT]], [[RETVAL_SROA_7_8_INSERT_SHIFT]]
69+
; SSE2-NEXT: [[RETVAL_SROA_5_8_INSERT_EXT:%.*]] = zext i16 [[LD2_4]] to i64
70+
; SSE2-NEXT: [[TMP21:%.*]] = or disjoint i64 [[RETVAL_SROA_7_8_INSERT_INSERT]], [[RETVAL_SROA_5_8_INSERT_EXT]]
71+
; SSE2-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue { i64, i64 } [[DOTFCA_0_INSERT]], i64 [[TMP21]], 1
72+
; SSE2-NEXT: ret { i64, i64 } [[DOTFCA_1_INSERT]]
73+
;
74+
; SSE4-LABEL: @compute_min(
75+
; SSE4-NEXT: entry:
76+
; SSE4-NEXT: [[TMP0:%.*]] = load <8 x i16>, ptr [[Y:%.*]], align 2
77+
; SSE4-NEXT: [[TMP1:%.*]] = load <8 x i16>, ptr [[X:%.*]], align 2
78+
; SSE4-NEXT: [[TMP2:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
79+
; SSE4-NEXT: [[TMP3:%.*]] = shufflevector <8 x i16> [[TMP2]], <8 x i16> poison, <2 x i32> <i32 0, i32 4>
80+
; SSE4-NEXT: [[TMP4:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
81+
; SSE4-NEXT: [[TMP5:%.*]] = shufflevector <8 x i16> [[TMP4]], <8 x i16> poison, <2 x i32> <i32 1, i32 5>
82+
; SSE4-NEXT: [[TMP6:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
83+
; SSE4-NEXT: [[TMP7:%.*]] = shufflevector <8 x i16> [[TMP6]], <8 x i16> poison, <2 x i32> <i32 3, i32 6>
84+
; SSE4-NEXT: [[TMP8:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
85+
; SSE4-NEXT: [[TMP9:%.*]] = shufflevector <8 x i16> [[TMP8]], <8 x i16> poison, <2 x i32> <i32 2, i32 7>
86+
; SSE4-NEXT: [[TMP10:%.*]] = zext <2 x i16> [[TMP9]] to <2 x i64>
87+
; SSE4-NEXT: [[TMP11:%.*]] = shl nuw <2 x i64> [[TMP10]], <i64 32, i64 48>
88+
; SSE4-NEXT: [[TMP12:%.*]] = zext <2 x i16> [[TMP7]] to <2 x i64>
89+
; SSE4-NEXT: [[TMP13:%.*]] = shl nuw <2 x i64> [[TMP12]], <i64 48, i64 32>
90+
; SSE4-NEXT: [[TMP14:%.*]] = or disjoint <2 x i64> [[TMP11]], [[TMP13]]
91+
; SSE4-NEXT: [[TMP15:%.*]] = zext <2 x i16> [[TMP5]] to <2 x i64>
92+
; SSE4-NEXT: [[TMP16:%.*]] = shl nuw nsw <2 x i64> [[TMP15]], splat (i64 16)
93+
; SSE4-NEXT: [[TMP17:%.*]] = or disjoint <2 x i64> [[TMP14]], [[TMP16]]
94+
; SSE4-NEXT: [[TMP18:%.*]] = zext <2 x i16> [[TMP3]] to <2 x i64>
95+
; SSE4-NEXT: [[TMP19:%.*]] = or disjoint <2 x i64> [[TMP17]], [[TMP18]]
96+
; SSE4-NEXT: [[TMP20:%.*]] = extractelement <2 x i64> [[TMP19]], i64 0
97+
; SSE4-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue { i64, i64 } poison, i64 [[TMP20]], 0
98+
; SSE4-NEXT: [[TMP21:%.*]] = extractelement <2 x i64> [[TMP19]], i64 1
99+
; SSE4-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue { i64, i64 } [[DOTFCA_0_INSERT]], i64 [[TMP21]], 1
100+
; SSE4-NEXT: ret { i64, i64 } [[DOTFCA_1_INSERT]]
101+
;
102+
; AVX-LABEL: @compute_min(
103+
; AVX-NEXT: entry:
104+
; AVX-NEXT: [[TMP0:%.*]] = load <8 x i16>, ptr [[Y:%.*]], align 2
105+
; AVX-NEXT: [[TMP1:%.*]] = load <8 x i16>, ptr [[X:%.*]], align 2
106+
; AVX-NEXT: [[TMP2:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
107+
; AVX-NEXT: [[TMP3:%.*]] = shufflevector <8 x i16> [[TMP2]], <8 x i16> poison, <2 x i32> <i32 0, i32 4>
108+
; AVX-NEXT: [[TMP4:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
109+
; AVX-NEXT: [[TMP5:%.*]] = shufflevector <8 x i16> [[TMP4]], <8 x i16> poison, <2 x i32> <i32 1, i32 5>
110+
; AVX-NEXT: [[TMP6:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
111+
; AVX-NEXT: [[TMP7:%.*]] = shufflevector <8 x i16> [[TMP6]], <8 x i16> poison, <2 x i32> <i32 3, i32 6>
112+
; AVX-NEXT: [[TMP8:%.*]] = tail call <8 x i16> @llvm.smin.v8i16(<8 x i16> [[TMP0]], <8 x i16> [[TMP1]])
113+
; AVX-NEXT: [[TMP9:%.*]] = shufflevector <8 x i16> [[TMP8]], <8 x i16> poison, <2 x i32> <i32 2, i32 7>
114+
; AVX-NEXT: [[TMP10:%.*]] = zext <2 x i16> [[TMP9]] to <2 x i64>
115+
; AVX-NEXT: [[TMP11:%.*]] = shl nuw <2 x i64> [[TMP10]], <i64 32, i64 48>
116+
; AVX-NEXT: [[TMP12:%.*]] = zext <2 x i16> [[TMP7]] to <2 x i64>
117+
; AVX-NEXT: [[TMP13:%.*]] = shl nuw <2 x i64> [[TMP12]], <i64 48, i64 32>
118+
; AVX-NEXT: [[TMP14:%.*]] = or disjoint <2 x i64> [[TMP11]], [[TMP13]]
119+
; AVX-NEXT: [[TMP15:%.*]] = zext <2 x i16> [[TMP5]] to <2 x i64>
120+
; AVX-NEXT: [[TMP16:%.*]] = shl nuw nsw <2 x i64> [[TMP15]], splat (i64 16)
121+
; AVX-NEXT: [[TMP17:%.*]] = or disjoint <2 x i64> [[TMP14]], [[TMP16]]
122+
; AVX-NEXT: [[TMP18:%.*]] = zext <2 x i16> [[TMP3]] to <2 x i64>
123+
; AVX-NEXT: [[TMP19:%.*]] = or disjoint <2 x i64> [[TMP17]], [[TMP18]]
124+
; AVX-NEXT: [[TMP20:%.*]] = extractelement <2 x i64> [[TMP19]], i64 0
125+
; AVX-NEXT: [[DOTFCA_0_INSERT:%.*]] = insertvalue { i64, i64 } poison, i64 [[TMP20]], 0
126+
; AVX-NEXT: [[TMP21:%.*]] = extractelement <2 x i64> [[TMP19]], i64 1
127+
; AVX-NEXT: [[DOTFCA_1_INSERT:%.*]] = insertvalue { i64, i64 } [[DOTFCA_0_INSERT]], i64 [[TMP21]], 1
128+
; AVX-NEXT: ret { i64, i64 } [[DOTFCA_1_INSERT]]
129+
;
130+
entry:
131+
%retval = alloca %"struct.std::array", align 2
132+
br label %for.cond
133+
134+
for.cond: ; preds = %for.body, %entry
135+
%i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
136+
%cmp.not = icmp eq i32 %i.0, 8
137+
br i1 %cmp.not, label %for.cond.cleanup, label %for.body
138+
139+
for.cond.cleanup: ; preds = %for.cond
140+
%.fca.0.load = load i64, ptr %retval, align 2
141+
%.fca.0.insert = insertvalue { i64, i64 } poison, i64 %.fca.0.load, 0
142+
%.fca.1.gep = getelementptr inbounds nuw i8, ptr %retval, i64 8
143+
%.fca.1.load = load i64, ptr %.fca.1.gep, align 2
144+
%.fca.1.insert = insertvalue { i64, i64 } %.fca.0.insert, i64 %.fca.1.load, 1
145+
ret { i64, i64 } %.fca.1.insert
146+
147+
for.body: ; preds = %for.cond
148+
%conv = zext nneg i32 %i.0 to i64
149+
%pt1 = getelementptr inbounds nuw [8 x i16], ptr %x, i64 0, i64 %conv
150+
%pt0 = getelementptr inbounds nuw [8 x i16], ptr %y, i64 0, i64 %conv
151+
%ld0 = load i16, ptr %pt0, align 2
152+
%ld1 = load i16, ptr %pt1, align 2
153+
%cmp.i = icmp slt i16 %ld0, %ld1
154+
%sel = select i1 %cmp.i, ptr %pt0, ptr %pt1
155+
%ld2 = load i16, ptr %sel, align 2
156+
%pt2 = getelementptr inbounds nuw [8 x i16], ptr %retval, i64 0, i64 %conv
157+
store i16 %ld2, ptr %pt2, align 2
158+
%inc = add nuw nsw i32 %i.0, 1
159+
br label %for.cond
160+
}
161+
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
162+
; CHECK: {{.*}}
163+
; SSE: {{.*}}

0 commit comments

Comments
 (0)