132 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt -mtriple=amdgcn-amd-amdhsa -passes=instcombine -S < %s | FileCheck %s3 4%struct.type = type { [256 x <2 x i64>] }5@g1 = external hidden addrspace(3) global %struct.type, align 166 7; This test requires the PtrReplacer to replace users in an RPO traversal.8; Furthermore, %ptr.else need not to be replaced so it must be retained in9; %ptr.sink.10define <2 x i64> @func(ptr addrspace(4) byref(%struct.type) align 16 %0, i1 %cmp.0) {11; CHECK-LABEL: define <2 x i64> @func(12; CHECK-SAME: ptr addrspace(4) byref([[STRUCT_TYPE:%.*]]) align 16 [[TMP0:%.*]], i1 [[CMP_0:%.*]]) {13; CHECK-NEXT: [[ENTRY:.*:]]14; CHECK-NEXT: br i1 [[CMP_0]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]15; CHECK: [[IF_THEN]]:16; CHECK-NEXT: [[VAL_THEN:%.*]] = addrspacecast ptr addrspace(4) [[TMP0]] to ptr17; CHECK-NEXT: br label %[[SINK:.*]]18; CHECK: [[IF_ELSE]]:19; CHECK-NEXT: [[PTR_ELSE:%.*]] = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 1620; CHECK-NEXT: br label %[[SINK]]21; CHECK: [[SINK]]:22; CHECK-NEXT: [[PTR_SINK:%.*]] = phi ptr [ [[PTR_ELSE]], %[[IF_ELSE]] ], [ [[VAL_THEN]], %[[IF_THEN]] ]23; CHECK-NEXT: [[VAL_SINK:%.*]] = load <2 x i64>, ptr [[PTR_SINK]], align 1624; CHECK-NEXT: ret <2 x i64> [[VAL_SINK]]25;26entry:27 %coerce = alloca %struct.type, align 16, addrspace(5)28 call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) align 16 %coerce, ptr addrspace(4) align 16 %0, i64 4096, i1 false)29 br i1 %cmp.0, label %if.then, label %if.else30 31if.then: ; preds = %entry32 %ptr.then = getelementptr inbounds i8, ptr addrspace(5) %coerce, i64 033 %val.then = addrspacecast ptr addrspace(5) %ptr.then to ptr34 br label %sink35 36if.else: ; preds = %entry37 %ptr.else = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 1638 %val.else = getelementptr inbounds nuw i8, ptr %ptr.else, i64 039 br label %sink40 41sink:42 %ptr.sink = phi ptr [ %val.else, %if.else ], [ %val.then, %if.then ]43 %val.sink = load <2 x i64>, ptr %ptr.sink, align 1644 ret <2 x i64> %val.sink45}46 47define <2 x i64> @func_phi_loop(ptr addrspace(4) byref(%struct.type) align 16 %0, i1 %cmp.0) {48; CHECK-LABEL: define <2 x i64> @func_phi_loop(49; CHECK-SAME: ptr addrspace(4) byref([[STRUCT_TYPE:%.*]]) align 16 [[TMP0:%.*]], i1 [[CMP_0:%.*]]) {50; CHECK-NEXT: [[ENTRY:.*]]:51; CHECK-NEXT: [[VAL_0:%.*]] = addrspacecast ptr addrspace(4) [[TMP0]] to ptr52; CHECK-NEXT: br label %[[LOOP:.*]]53; CHECK: [[LOOP]]:54; CHECK-NEXT: [[PTR_PHI_R:%.*]] = phi ptr [ [[PTR_1:%.*]], %[[LOOP]] ], [ [[VAL_0]], %[[ENTRY]] ]55; CHECK-NEXT: [[PTR_1]] = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 1656; CHECK-NEXT: br i1 [[CMP_0]], label %[[LOOP]], label %[[SINK:.*]]57; CHECK: [[SINK]]:58; CHECK-NEXT: [[VAL_SINK:%.*]] = load <2 x i64>, ptr [[PTR_PHI_R]], align 1659; CHECK-NEXT: ret <2 x i64> [[VAL_SINK]]60;61entry:62 %coerce = alloca %struct.type, align 16, addrspace(5)63 call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) align 16 %coerce, ptr addrspace(4) align 16 %0, i64 4096, i1 false)64 %ptr.0 = getelementptr inbounds i8, ptr addrspace(5) %coerce, i64 065 %val.0 = addrspacecast ptr addrspace(5) %ptr.0 to ptr66 br label %loop67 68loop:69 %ptr.phi = phi ptr [ %val.1, %loop ], [ %val.0, %entry ]70 %ptr.1 = load ptr, ptr addrspace(3) getelementptr inbounds nuw (i8, ptr addrspace(3) @g1, i32 32), align 1671 %val.1 = getelementptr inbounds nuw i8, ptr %ptr.1, i64 072 br i1 %cmp.0, label %loop, label %sink73 74sink:75 %val.sink = load <2 x i64>, ptr %ptr.phi, align 1676 ret <2 x i64> %val.sink77}78 79; Crashed in IC PtrReplacer because an invalid select was generated with addrspace(4) and addrspace(5)80; operands.81define amdgpu_kernel void @select_addr4_addr5(ptr addrspace(4) byref([12 x i8]) align 16 %arg) {82; CHECK-LABEL: define amdgpu_kernel void @select_addr4_addr5(83; CHECK-SAME: ptr addrspace(4) byref([12 x i8]) align 16 [[ARG:%.*]]) {84; CHECK-NEXT: [[BB:.*:]]85; CHECK-NEXT: ret void86;87bb:88 %alloca = alloca i32, i32 0, align 8, addrspace(5)89 %alloca1 = alloca [12 x i8], align 16, addrspace(5)90 call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) %alloca1, ptr addrspace(4) %arg, i64 0, i1 false)91 %select = select i1 false, ptr addrspace(5) %alloca1, ptr addrspace(5) %alloca92 call void @llvm.memcpy.p0.p5.i64(ptr null, ptr addrspace(5) %select, i64 0, i1 false)93 ret void94}95 96; Same as above but with swapped operands on the select.97define amdgpu_kernel void @select_addr4_addr5_swapped(ptr addrspace(4) byref([12 x i8]) align 16 %arg) {98; CHECK-LABEL: define amdgpu_kernel void @select_addr4_addr5_swapped(99; CHECK-SAME: ptr addrspace(4) byref([12 x i8]) align 16 [[ARG:%.*]]) {100; CHECK-NEXT: [[BB:.*:]]101; CHECK-NEXT: ret void102;103bb:104 %alloca = alloca i32, i32 0, align 8, addrspace(5)105 %alloca1 = alloca [12 x i8], align 16, addrspace(5)106 call void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) %alloca1, ptr addrspace(4) %arg, i64 0, i1 false)107 %select = select i1 false, ptr addrspace(5) %alloca, ptr addrspace(5) %alloca1108 call void @llvm.memcpy.p0.p5.i64(ptr null, ptr addrspace(5) %select, i64 0, i1 false)109 ret void110}111 112@global = external addrspace(1) constant [16 x float], align 64113 114define float @issue160302(i1 %cond, ptr addrspace(5) %arg) {115; CHECK-LABEL: define float @issue160302(116; CHECK-SAME: i1 [[COND:%.*]], ptr addrspace(5) [[ARG:%.*]]) {117; CHECK-NEXT: [[AGG_TMP2_I4:%.*]] = alloca [16 x float], align 64, addrspace(5)118; CHECK-NEXT: [[SELECT_PTR:%.*]] = select i1 [[COND]], ptr addrspace(5) [[AGG_TMP2_I4]], ptr addrspace(5) [[ARG]]119; CHECK-NEXT: [[COND_I:%.*]] = load float, ptr addrspace(5) [[SELECT_PTR]], align 4120; CHECK-NEXT: ret float [[COND_I]]121;122 %agg.tmp2.i4 = alloca [16 x float], align 64, addrspace(5)123 call void @llvm.memcpy.p5.p1.i64(ptr addrspace(5) %agg.tmp2.i4, ptr addrspace(1) @global, i64 0, i1 false)124 %m_Data.i14.i = getelementptr [16 x float], ptr addrspace(5) %agg.tmp2.i4, i32 0, i32 0125 %gep = getelementptr [16 x float], ptr addrspace(5) %arg, i32 0, i32 0126 %select.ptr = select i1 %cond, ptr addrspace(5) %m_Data.i14.i, ptr addrspace(5) %gep127 %cond.i = load float, ptr addrspace(5) %select.ptr, align 4128 ret float %cond.i129}130 131declare void @llvm.memcpy.p5.p4.i64(ptr addrspace(5) noalias writeonly captures(none), ptr addrspace(4) noalias readonly captures(none), i64, i1 immarg) #0132