237 lines · plain
1; REQUIRES: asserts2; RUN: opt < %s -passes='sroa<preserve-cfg>' -disable-output -debug-only=sroa 2>&1 | FileCheck %s3; RUN: opt < %s -passes='sroa<modify-cfg>' -disable-output -debug-only=sroa 2>&1 | FileCheck %s4 5target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-n8:16:32:64"6 7; CHECK-NOT: Tree structured merge rewrite8define i32 @test_alloca_not_fixed_vector() {9entry:10 %alloca = alloca [4 x float]11 12 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 013 store float 1.0, ptr %ptr014 15 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 116 store float 2.0, ptr %ptr117 18 %result = load i32, ptr %alloca19 ret i32 %result20}21 22define <4 x float> @test_more_than_one_load(<2 x float> %a, <2 x float> %b) {23entry:24 %alloca = alloca [4 x float]25 26 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 027 store <2 x float> %a, ptr %ptr028 29 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 230 store <2 x float> %b, ptr %ptr131 32 %result1 = load <4 x float>, ptr %alloca33 %result2 = load <4 x float>, ptr %alloca34 35 %final = fadd <4 x float> %result1, %result236 ret <4 x float> %final37}38 39define void @test_no_load(<4 x float> %a) {40entry:41 %alloca = alloca [4 x float]42 store <4 x float> %a, ptr %alloca43 ret void44}45 46define i32 @test_load_not_fixed_vector(<2 x float> %a, <2 x float> %b) {47entry:48 %alloca = alloca [4 x float]49 50 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 051 store <2 x float> %a, ptr %ptr052 53 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 254 store <2 x float> %b, ptr %ptr155 56 %result = load i32, ptr %alloca57 ret i32 %result58}59 60define <3 x float> @test_load_not_covering_alloca(<2 x float> %a, <2 x float> %b) {61entry:62 %alloca = alloca [4 x float]63 64 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 065 store <2 x float> %a, ptr %ptr066 67 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 268 store <2 x float> %b, ptr %ptr169 70 %result = load <3 x float>, ptr %ptr071 ret <3 x float> %result72}73 74define <4 x float> @test_store_not_fixed_vector(<vscale x 2 x float> %a) {75entry:76 %alloca = alloca [4 x float]77 78 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 079 %fixed = extractelement <vscale x 2 x float> %a, i32 080 store float %fixed, ptr %ptr081 82 %result = load <4 x float>, ptr %alloca83 ret <4 x float> %result84}85 86 87define <4 x float> @test_no_stores() {88entry:89 %alloca = alloca [4 x float]90 91 %result = load <4 x float>, ptr %alloca92 ret <4 x float> %result93}94 95define <4 x float> @test_stores_overlapping(<2 x float> %a, <2 x float> %b, <2 x float> %c) {96entry:97 %alloca = alloca [4 x float]98 99 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0100 store <2 x float> %a, ptr %ptr0101 102 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 1103 store <2 x float> %b, ptr %ptr1104 105 %ptr2 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2106 store <2 x float> %c, ptr %ptr2107 108 %result = load <4 x float>, ptr %alloca109 ret <4 x float> %result110}111 112define <4 x float> @test_stores_not_covering_alloca(<2 x float> %a) {113entry:114 %alloca = alloca [4 x float]115 116 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0117 store <2 x float> %a, ptr %ptr0118 119 %result = load <4 x float>, ptr %alloca120 ret <4 x float> %result121}122 123define <4 x float> @test_stores_not_same_basic_block(<2 x float> %a, <2 x float> %b, i1 %cond) {124entry:125 %alloca = alloca [4 x float]126 127 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0128 store <2 x float> %a, ptr %ptr0129 130 br i1 %cond, label %then, label %else131 132then:133 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2134 store <2 x float> %b, ptr %ptr1135 br label %merge136 137else:138 br label %merge139 140merge:141 %result = load <4 x float>, ptr %alloca142 ret <4 x float> %result143}144 145define <4 x float> @test_load_before_stores(<2 x float> %a, <2 x float> %b) {146entry:147 %alloca = alloca [4 x float]148 149 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0150 store <2 x float> %a, ptr %ptr0151 152 %intermediate = load <4 x float>, ptr %alloca153 154 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2155 store <2 x float> %b, ptr %ptr1156 157 ret <4 x float> %intermediate158}159 160define <4 x float> @test_other_instructions(<2 x float> %a, <2 x float> %b) {161entry:162 %alloca = alloca [4 x float]163 164 ; Store first vector165 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0166 store <2 x float> %a, ptr %ptr0167 168 ; Other instruction (memset) that's not a simple load/store169 call void @llvm.memset.p0.i64(ptr %alloca, i8 0, i64 8, i1 false)170 171 ; Store second vector172 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2173 store <2 x float> %b, ptr %ptr1174 175 %result = load <4 x float>, ptr %alloca176 ret <4 x float> %result177}178 179define <4 x float> @volatile_stores(<2 x i32> %a, <2 x i32> %b) {180entry:181 %alloca = alloca [4 x float]182 183 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0184 store volatile <2 x i32> %a, ptr %ptr0185 186 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2187 store volatile <2 x i32> %b, ptr %ptr1188 189 %result = load <4 x float>, ptr %alloca190 ret <4 x float> %result191}192 193define <4 x float> @volatile_loads(<2 x i32> %a, <2 x i32> %b) {194entry:195 %alloca = alloca [4 x float]196 197 %ptr0 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 0198 store <2 x i32> %a, ptr %ptr0199 200 %ptr1 = getelementptr inbounds [4 x float], ptr %alloca, i32 0, i32 2201 store <2 x i32> %b, ptr %ptr1202 203 %result = load volatile <4 x float>, ptr %alloca204 ret <4 x float> %result205}206 207define <4 x i15> @non_byte_aligned_alloca(<2 x i15> %a, <2 x i15> %b) {208entry:209 %alloca = alloca [4 x i15]210 211 %ptr0 = getelementptr inbounds [4 x i15], ptr %alloca, i32 0, i32 0212 store <2 x i15> %a, ptr %ptr0213 214 %ptr1 = getelementptr inbounds [4 x i15], ptr %alloca, i32 0, i32 2215 store <2 x i15> %b, ptr %ptr1216 217 %result = load <4 x i15>, ptr %alloca218 ret <4 x i15> %result219 220}221 222define <1 x i32> @test_store_value_size_not_multiple_of_allocated_element_type_size(<1 x i16> %a, <1 x i16> %b) {223entry:224 %alloca = alloca [2 x i16]225 226 %ptr0 = getelementptr inbounds [2 x i16], ptr %alloca, i32 0, i32 0227 store <1 x i16> %a, ptr %ptr0228 229 %ptr1 = getelementptr inbounds [2 x i16], ptr %alloca, i32 0, i32 1230 store <1 x i16> %b, ptr %ptr1231 232 %result = load <1 x i32>, ptr %alloca233 ret <1 x i32> %result234}235 236declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg)237