108 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=infer-address-spaces %s | FileCheck %s3 4; Test that pure GetElementPtr instructions not directly connected to5; a memory operation are inferred.6 7@lds = internal unnamed_addr addrspace(3) global [648 x double] poison, align 88 9define void @simplified_constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {10; CHECK-LABEL: @simplified_constexpr_gep_addrspacecast(11; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds double, ptr addrspace(3) getelementptr inbounds ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384), i64 [[IDX0:%.*]]12; CHECK-NEXT: store double 1.000000e+00, ptr addrspace(3) [[GEP0]], align 813; CHECK-NEXT: ret void14;15 %gep0 = getelementptr inbounds double, ptr addrspacecast (ptr addrspace(3) getelementptr inbounds ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384) to ptr), i64 %idx016 %asc = addrspacecast ptr %gep0 to ptr addrspace(3)17 store double 1.000000e+00, ptr addrspace(3) %asc, align 818 ret void19}20 21define void @constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {22; CHECK-LABEL: @constexpr_gep_addrspacecast(23; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds double, ptr addrspace(3) getelementptr ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384), i64 [[IDX0:%.*]]24; CHECK-NEXT: store double 1.000000e+00, ptr addrspace(3) [[GEP0]], align 825; CHECK-NEXT: ret void26;27 %gep0 = getelementptr inbounds double, ptr getelementptr ([648 x double], ptr addrspacecast (ptr addrspace(3) @lds to ptr), i64 0, i64 384), i64 %idx028 %asc = addrspacecast ptr %gep0 to ptr addrspace(3)29 store double 1.0, ptr addrspace(3) %asc, align 830 ret void31}32 33define void @constexpr_gep_gep_addrspacecast(i64 %idx0, i64 %idx1) {34; CHECK-LABEL: @constexpr_gep_gep_addrspacecast(35; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds double, ptr addrspace(3) getelementptr ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384), i64 [[IDX0:%.*]]36; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds double, ptr addrspace(3) [[GEP0]], i64 [[IDX1:%.*]]37; CHECK-NEXT: store double 1.000000e+00, ptr addrspace(3) [[GEP1]], align 838; CHECK-NEXT: ret void39;40 %gep0 = getelementptr inbounds double, ptr getelementptr ([648 x double], ptr addrspacecast (ptr addrspace(3) @lds to ptr), i64 0, i64 384), i64 %idx041 %gep1 = getelementptr inbounds double, ptr %gep0, i64 %idx142 %asc = addrspacecast ptr %gep1 to ptr addrspace(3)43 store double 1.0, ptr addrspace(3) %asc, align 844 ret void45}46 47; Don't crash48define amdgpu_kernel void @vector_gep(<4 x ptr addrspace(3)> %array) nounwind {49; CHECK-LABEL: @vector_gep(50; CHECK-NEXT: [[CAST:%.*]] = addrspacecast <4 x ptr addrspace(3)> [[ARRAY:%.*]] to <4 x ptr>51; CHECK-NEXT: [[P:%.*]] = getelementptr [1024 x i32], <4 x ptr> [[CAST]], <4 x i16> zeroinitializer, <4 x i16> splat (i16 16)52; CHECK-NEXT: [[P0:%.*]] = extractelement <4 x ptr> [[P]], i32 053; CHECK-NEXT: [[P1:%.*]] = extractelement <4 x ptr> [[P]], i32 154; CHECK-NEXT: [[P2:%.*]] = extractelement <4 x ptr> [[P]], i32 255; CHECK-NEXT: [[P3:%.*]] = extractelement <4 x ptr> [[P]], i32 356; CHECK-NEXT: store i32 99, ptr [[P0]], align 457; CHECK-NEXT: store i32 99, ptr [[P1]], align 458; CHECK-NEXT: store i32 99, ptr [[P2]], align 459; CHECK-NEXT: store i32 99, ptr [[P3]], align 460; CHECK-NEXT: ret void61;62 %cast = addrspacecast <4 x ptr addrspace(3)> %array to <4 x ptr>63 %p = getelementptr [1024 x i32], <4 x ptr> %cast, <4 x i16> zeroinitializer, <4 x i16> <i16 16, i16 16, i16 16, i16 16>64 %p0 = extractelement <4 x ptr> %p, i32 065 %p1 = extractelement <4 x ptr> %p, i32 166 %p2 = extractelement <4 x ptr> %p, i32 267 %p3 = extractelement <4 x ptr> %p, i32 368 store i32 99, ptr %p069 store i32 99, ptr %p170 store i32 99, ptr %p271 store i32 99, ptr %p372 ret void73}74 75define void @repeated_constexpr_gep_addrspacecast(i64 %idx0, i64 %idx1) {76; CHECK-LABEL: @repeated_constexpr_gep_addrspacecast(77; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds double, ptr addrspace(3) getelementptr ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384), i64 [[IDX0:%.*]]78; CHECK-NEXT: store double 1.000000e+00, ptr addrspace(3) [[GEP0]], align 879; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds double, ptr addrspace(3) getelementptr ([648 x double], ptr addrspace(3) @lds, i64 0, i64 384), i64 [[IDX1:%.*]]80; CHECK-NEXT: store double 1.000000e+00, ptr addrspace(3) [[GEP1]], align 881; CHECK-NEXT: ret void82;83 %gep0 = getelementptr inbounds double, ptr getelementptr ([648 x double], ptr addrspacecast (ptr addrspace(3) @lds to ptr), i64 0, i64 384), i64 %idx084 %asc0 = addrspacecast ptr %gep0 to ptr addrspace(3)85 store double 1.0, ptr addrspace(3) %asc0, align 886 87 %gep1 = getelementptr inbounds double, ptr getelementptr ([648 x double], ptr addrspacecast (ptr addrspace(3) @lds to ptr), i64 0, i64 384), i64 %idx188 %asc1 = addrspacecast ptr %gep1 to ptr addrspace(3)89 store double 1.0, ptr addrspace(3) %asc1, align 890 91 ret void92}93 94define void @unorder_constexpr_gep_bitcast() {95; CHECK-LABEL: @unorder_constexpr_gep_bitcast(96; CHECK-NEXT: [[X0:%.*]] = load i32, ptr addrspace(3) @lds, align 497; CHECK-NEXT: [[X1:%.*]] = load i32, ptr addrspace(3) getelementptr (i32, ptr addrspace(3) @lds, i32 1), align 498; CHECK-NEXT: call void @use(i32 [[X0]], i32 [[X1]])99; CHECK-NEXT: ret void100;101 %x0 = load i32, ptr addrspacecast (ptr addrspace(3) @lds to ptr), align 4102 %x1 = load i32, ptr getelementptr (i32, ptr addrspacecast (ptr addrspace(3) @lds to ptr), i32 1), align 4103 call void @use(i32 %x0, i32 %x1)104 ret void105}106 107declare void @use(i32, i32)108