1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
2 ; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve < %s | FileCheck %s
4 ; Verify that DAG combine rules for LD1 + sext/zext don't apply when the
5 ; result of LD1 has multiple uses
7 define <vscale x 2 x i64> @no_dag_combine_zext_sext(<vscale x 2 x i1> %pg,
8 ; CHECK-LABEL: no_dag_combine_zext_sext:
10 ; CHECK-NEXT: ld1b { z0.d }, p0/z, [z0.d, #16]
11 ; CHECK-NEXT: st1b { z0.d }, p1, [x0]
12 ; CHECK-NEXT: and z0.d, z0.d, #0xff
14 <vscale x 2 x i64> %base,
16 <vscale x 2 x i1> %pred) {
17 %load = call <vscale x 2 x i8> @llvm.aarch64.sve.ld1.gather.scalar.offset.nxv2i8.nxv2i64(<vscale x 2 x i1> %pg,
18 <vscale x 2 x i64> %base,
20 %res1 = zext <vscale x 2 x i8> %load to <vscale x 2 x i64>
21 %res2 = sext <vscale x 2 x i8> %load to <vscale x 2 x i64>
22 call void @llvm.masked.store.nxv2i8(<vscale x 2 x i8> %load,
25 <vscale x 2 x i1> %pred)
27 ret <vscale x 2 x i64> %res1
30 define <vscale x 2 x i64> @no_dag_combine_sext(<vscale x 2 x i1> %pg,
31 ; CHECK-LABEL: no_dag_combine_sext:
33 ; CHECK-NEXT: ld1b { z1.d }, p0/z, [z0.d, #16]
34 ; CHECK-NEXT: ptrue p0.d
35 ; CHECK-NEXT: movprfx z0, z1
36 ; CHECK-NEXT: sxtb z0.d, p0/m, z1.d
37 ; CHECK-NEXT: st1b { z1.d }, p1, [x0]
39 <vscale x 2 x i64> %base,
41 <vscale x 2 x i1> %pred) {
42 %load = call <vscale x 2 x i8> @llvm.aarch64.sve.ld1.gather.scalar.offset.nxv2i8.nxv2i64(<vscale x 2 x i1> %pg,
43 <vscale x 2 x i64> %base,
45 %res = sext <vscale x 2 x i8> %load to <vscale x 2 x i64>
46 call void @llvm.masked.store.nxv2i8(<vscale x 2 x i8> %load,
49 <vscale x 2 x i1> %pred)
51 ret <vscale x 2 x i64> %res
54 define <vscale x 2 x i64> @no_dag_combine_zext(<vscale x 2 x i1> %pg,
55 ; CHECK-LABEL: no_dag_combine_zext:
57 ; CHECK-NEXT: ld1b { z0.d }, p0/z, [z0.d, #16]
58 ; CHECK-NEXT: st1b { z0.d }, p1, [x0]
59 ; CHECK-NEXT: and z0.d, z0.d, #0xff
61 <vscale x 2 x i64> %base,
63 <vscale x 2 x i1> %pred) {
64 %load = call <vscale x 2 x i8> @llvm.aarch64.sve.ld1.gather.scalar.offset.nxv2i8.nxv2i64(<vscale x 2 x i1> %pg,
65 <vscale x 2 x i64> %base,
67 %res = zext <vscale x 2 x i8> %load to <vscale x 2 x i64>
68 call void @llvm.masked.store.nxv2i8(<vscale x 2 x i8> %load,
71 <vscale x 2 x i1> %pred)
73 ret <vscale x 2 x i64> %res
76 define <vscale x 16 x i8> @narrow_i64_gather_index_i8_zext(ptr %out, ptr %in, <vscale x 16 x i8> %d, i64 %ptr){
77 ; CHECK-LABEL: narrow_i64_gather_index_i8_zext:
79 ; CHECK-NEXT: ptrue p0.s
80 ; CHECK-NEXT: add x8, x1, x2
81 ; CHECK-NEXT: ld1b { z0.s }, p0/z, [x8, #3, mul vl]
82 ; CHECK-NEXT: ld1b { z1.s }, p0/z, [x8, #2, mul vl]
83 ; CHECK-NEXT: ld1b { z2.s }, p0/z, [x1, x2]
84 ; CHECK-NEXT: ld1b { z3.s }, p0/z, [x8, #1, mul vl]
85 ; CHECK-NEXT: ld1b { z0.s }, p0/z, [x1, z0.s, uxtw]
86 ; CHECK-NEXT: ld1b { z1.s }, p0/z, [x1, z1.s, uxtw]
87 ; CHECK-NEXT: ld1b { z2.s }, p0/z, [x1, z2.s, uxtw]
88 ; CHECK-NEXT: ld1b { z3.s }, p0/z, [x1, z3.s, uxtw]
89 ; CHECK-NEXT: uzp1 z0.h, z1.h, z0.h
90 ; CHECK-NEXT: uzp1 z1.h, z2.h, z3.h
91 ; CHECK-NEXT: uzp1 z0.b, z1.b, z0.b
93 %1 = getelementptr inbounds i8, ptr %in, i64 %ptr
94 %2 = bitcast ptr %1 to ptr
95 %wide.load = load <vscale x 16 x i8>, ptr %2, align 1
96 %3 = zext <vscale x 16 x i8> %wide.load to <vscale x 16 x i64>
97 %4 = getelementptr inbounds i8, ptr %in, <vscale x 16 x i64> %3
98 %wide.masked.gather = call <vscale x 16 x i8> @llvm.masked.gather.nxv16i8.nxv16p0(<vscale x 16 x ptr> %4, i32 1, <vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> undef)
99 ret <vscale x 16 x i8> %wide.masked.gather
102 define <vscale x 16 x i8> @narrow_i64_gather_index_i8_sext(ptr %out, ptr %in, <vscale x 16 x i8> %d, i64 %ptr){
103 ; CHECK-LABEL: narrow_i64_gather_index_i8_sext:
105 ; CHECK-NEXT: ptrue p0.s
106 ; CHECK-NEXT: add x8, x1, x2
107 ; CHECK-NEXT: ld1sb { z0.s }, p0/z, [x8, #3, mul vl]
108 ; CHECK-NEXT: ld1sb { z1.s }, p0/z, [x8, #2, mul vl]
109 ; CHECK-NEXT: ld1sb { z2.s }, p0/z, [x1, x2]
110 ; CHECK-NEXT: ld1sb { z3.s }, p0/z, [x8, #1, mul vl]
111 ; CHECK-NEXT: ld1b { z0.s }, p0/z, [x1, z0.s, sxtw]
112 ; CHECK-NEXT: ld1b { z1.s }, p0/z, [x1, z1.s, sxtw]
113 ; CHECK-NEXT: ld1b { z2.s }, p0/z, [x1, z2.s, sxtw]
114 ; CHECK-NEXT: ld1b { z3.s }, p0/z, [x1, z3.s, sxtw]
115 ; CHECK-NEXT: uzp1 z0.h, z1.h, z0.h
116 ; CHECK-NEXT: uzp1 z1.h, z2.h, z3.h
117 ; CHECK-NEXT: uzp1 z0.b, z1.b, z0.b
119 %1 = getelementptr inbounds i8, ptr %in, i64 %ptr
120 %2 = bitcast ptr %1 to ptr
121 %wide.load = load <vscale x 16 x i8>, ptr %2, align 1
122 %3 = sext <vscale x 16 x i8> %wide.load to <vscale x 16 x i64>
123 %4 = getelementptr inbounds i8, ptr %in, <vscale x 16 x i64> %3
124 %wide.masked.gather = call <vscale x 16 x i8> @llvm.masked.gather.nxv16i8.nxv16p0(<vscale x 16 x ptr> %4, i32 1, <vscale x 16 x i1> splat (i1 true), <vscale x 16 x i8> undef)
125 ret <vscale x 16 x i8> %wide.masked.gather
128 define <vscale x 8 x i16> @narrow_i64_gather_index_i16_zext(ptr %out, ptr %in, <vscale x 8 x i16> %d, i64 %ptr){
129 ; CHECK-LABEL: narrow_i64_gather_index_i16_zext:
131 ; CHECK-NEXT: ptrue p0.s
132 ; CHECK-NEXT: add x8, x1, x2, lsl #1
133 ; CHECK-NEXT: ld1h { z0.s }, p0/z, [x1, x2, lsl #1]
134 ; CHECK-NEXT: ld1h { z1.s }, p0/z, [x8, #1, mul vl]
135 ; CHECK-NEXT: ld1h { z0.s }, p0/z, [x1, z0.s, uxtw #1]
136 ; CHECK-NEXT: ld1h { z1.s }, p0/z, [x1, z1.s, uxtw #1]
137 ; CHECK-NEXT: uzp1 z0.h, z0.h, z1.h
139 %1 = getelementptr inbounds i16, ptr %in, i64 %ptr
140 %2 = bitcast ptr %1 to ptr
141 %wide.load = load <vscale x 8 x i16>, ptr %2, align 1
142 %3 = zext <vscale x 8 x i16> %wide.load to <vscale x 8 x i64>
143 %4 = getelementptr inbounds i16, ptr %in, <vscale x 8 x i64> %3
144 %wide.masked.gather = call <vscale x 8 x i16> @llvm.masked.gather.nxv8i16.nxv8p0(<vscale x 8 x ptr> %4, i32 1, <vscale x 8 x i1> splat (i1 true), <vscale x 8 x i16> undef)
145 ret <vscale x 8 x i16> %wide.masked.gather
148 define <vscale x 8 x i16> @narrow_i64_gather_index_i16_sext(ptr %out, ptr %in, <vscale x 8 x i16> %d, i64 %ptr){
149 ; CHECK-LABEL: narrow_i64_gather_index_i16_sext:
151 ; CHECK-NEXT: ptrue p0.s
152 ; CHECK-NEXT: add x8, x1, x2, lsl #1
153 ; CHECK-NEXT: ld1sh { z0.s }, p0/z, [x1, x2, lsl #1]
154 ; CHECK-NEXT: ld1sh { z1.s }, p0/z, [x8, #1, mul vl]
155 ; CHECK-NEXT: ld1h { z0.s }, p0/z, [x1, z0.s, sxtw #1]
156 ; CHECK-NEXT: ld1h { z1.s }, p0/z, [x1, z1.s, sxtw #1]
157 ; CHECK-NEXT: uzp1 z0.h, z0.h, z1.h
159 %1 = getelementptr inbounds i16, ptr %in, i64 %ptr
160 %2 = bitcast ptr %1 to ptr
161 %wide.load = load <vscale x 8 x i16>, ptr %2, align 1
162 %3 = sext <vscale x 8 x i16> %wide.load to <vscale x 8 x i64>
163 %4 = getelementptr inbounds i16, ptr %in, <vscale x 8 x i64> %3
164 %wide.masked.gather = call <vscale x 8 x i16> @llvm.masked.gather.nxv8i16.nxv8p0(<vscale x 8 x ptr> %4, i32 1, <vscale x 8 x i1> splat (i1 true), <vscale x 8 x i16> undef)
165 ret <vscale x 8 x i16> %wide.masked.gather
168 define <vscale x 4 x i32> @no_narrow_i64_gather_index_i32(ptr %out, ptr %in, <vscale x 4 x i32> %d, i64 %ptr){
169 ; CHECK-LABEL: no_narrow_i64_gather_index_i32:
171 ; CHECK-NEXT: ptrue p0.s
172 ; CHECK-NEXT: ld1w { z0.s }, p0/z, [x1, x2, lsl #2]
173 ; CHECK-NEXT: ld1w { z0.s }, p0/z, [x1, z0.s, uxtw #2]
175 %1 = getelementptr inbounds i32, ptr %in, i64 %ptr
176 %2 = bitcast ptr %1 to ptr
177 %wide.load = load <vscale x 4 x i32>, ptr %2, align 1
178 %3 = zext <vscale x 4 x i32> %wide.load to <vscale x 4 x i64>
179 %4 = getelementptr inbounds i32, ptr %in, <vscale x 4 x i64> %3
180 %wide.masked.gather = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> %4, i32 1, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> undef)
181 ret <vscale x 4 x i32> %wide.masked.gather
184 define <vscale x 2 x i64> @no_narrow_i64_gather_index_i64(ptr %out, ptr %in, <vscale x 2 x i64> %d, i64 %ptr){
185 ; CHECK-LABEL: no_narrow_i64_gather_index_i64:
187 ; CHECK-NEXT: ptrue p0.d
188 ; CHECK-NEXT: ld1d { z0.d }, p0/z, [x1, x2, lsl #3]
189 ; CHECK-NEXT: ld1d { z0.d }, p0/z, [x1, z0.d, lsl #3]
191 %1 = getelementptr inbounds i64, ptr %in, i64 %ptr
192 %2 = bitcast ptr %1 to ptr
193 %wide.load = load <vscale x 2 x i64>, ptr %2, align 1
194 %3 = getelementptr inbounds i64, ptr %in, <vscale x 2 x i64> %wide.load
195 %wide.masked.gather = call <vscale x 2 x i64> @llvm.masked.gather.nxv2i64.nxv2p0(<vscale x 2 x ptr> %3, i32 1, <vscale x 2 x i1> splat (i1 true), <vscale x 2 x i64> undef)
196 ret <vscale x 2 x i64> %wide.masked.gather
199 declare <vscale x 2 x i8> @llvm.aarch64.sve.ld1.gather.scalar.offset.nxv2i8.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>, i64)
200 declare void @llvm.masked.store.nxv2i8(<vscale x 2 x i8>, ptr, i32, <vscale x 2 x i1>)
201 declare <vscale x 16 x i8> @llvm.masked.gather.nxv16i8.nxv16p0(<vscale x 16 x ptr>, i32, <vscale x 16 x i1>, <vscale x 16 x i8>)
202 declare <vscale x 8 x i16> @llvm.masked.gather.nxv8i16.nxv8p0(<vscale x 8 x ptr>, i32, <vscale x 8 x i1>, <vscale x 8 x i16>)
203 declare <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr>, i32, <vscale x 4 x i1>, <vscale x 4 x i32>)
204 declare <vscale x 2 x i64> @llvm.masked.gather.nxv2i64.nxv2p0(<vscale x 2 x ptr>, i32, <vscale x 2 x i1>, <vscale x 2 x i64>)