1 ; RUN: llc -O3 -aarch64-enable-gep-opt=true -verify-machineinstrs %s -o - | FileCheck %s 2 ; RUN: llc -O3 -aarch64-enable-gep-opt=true -mattr=-use-aa -print-after=codegenprepare < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-NoAA <%t %s 3 ; RUN: llc -O3 -aarch64-enable-gep-opt=true -mattr=+use-aa -print-after=codegenprepare < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-UseAA <%t %s 4 ; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cyclone < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-NoAA <%t %s 5 ; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cortex-a53 < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-UseAA <%t %s 6 7 target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128" 8 target triple = "aarch64-linux-gnueabi" 9 10 ; Following test cases test enabling SeparateConstOffsetFromGEP pass in AArch64 11 ; backend. If useAA() returns true, it will lower a GEP with multiple indices 12 ; into GEPs with a single index, otherwise it will lower it into a 13 ; "ptrtoint+arithmetics+inttoptr" form. 14 15 %struct = type { i32, i32, i32, i32, [20 x i32] } 16 17 ; Check that when two complex GEPs are used in two basic blocks, LLVM can 18 ; elimilate the common subexpression for the second use. 19 define void @test_GEP_CSE([240 x %struct]* %string, i32* %adj, i32 %lib, i64 %idxprom) { 20 %liberties = getelementptr [240 x %struct], [240 x %struct]* %string, i64 1, i64 %idxprom, i32 3 21 %1 = load i32, i32* %liberties, align 4 22 %cmp = icmp eq i32 %1, %lib 23 br i1 %cmp, label %if.then, label %if.end 24 25 if.then: ; preds = %entry 26 %origin = getelementptr [240 x %struct], [240 x %struct]* %string, i64 1, i64 %idxprom, i32 2 27 %2 = load i32, i32* %origin, align 4 28 store i32 %2, i32* %adj, align 4 29 br label %if.end 30 31 if.end: ; preds = %if.then, %entry 32 ret void 33 } 34 35 ; CHECK-LABEL: test_GEP_CSE: 36 ; CHECK: madd 37 ; CHECK: ldr 38 ; CHECK-NOT: madd 39 ; CHECK:ldr 40 41 ; CHECK-NoAA-LABEL: @test_GEP_CSE( 42 ; CHECK-NoAA: [[PTR0:%[a-zA-Z0-9]+]] = ptrtoint [240 x %struct]* %string to i64 43 ; CHECK-NoAA: [[PTR1:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96 44 ; CHECK-NoAA: [[PTR2:%[a-zA-Z0-9]+]] = add i64 [[PTR0]], [[PTR1]] 45 ; CHECK-NoAA: add i64 [[PTR2]], 23052 46 ; CHECK-NoAA: inttoptr 47 ; CHECK-NoAA: if.then: 48 ; CHECK-NoAA-NOT: ptrtoint 49 ; CHECK-NoAA-NOT: mul 50 ; CHECK-NoAA: add i64 [[PTR2]], 23048 51 ; CHECK-NoAA: inttoptr 52 53 ; CHECK-UseAA-LABEL: @test_GEP_CSE( 54 ; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = bitcast [240 x %struct]* %string to i8* 55 ; CHECK-UseAA: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96 56 ; CHECK-UseAA: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8, i8* [[PTR0]], i64 [[IDX]] 57 ; CHECK-UseAA: getelementptr i8, i8* [[PTR1]], i64 23052 58 ; CHECK-UseAA: bitcast 59 ; CHECK-UseAA: if.then: 60 ; CHECK-UseAA: getelementptr i8, i8* [[PTR1]], i64 23048 61 ; CHECK-UseAA: bitcast 62 63 %class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]} 64 %struct.pt = type { %struct.point*, i32, i32 } 65 %struct.point = type { i32, i32 } 66 67 ; Check when a GEP is used across two basic block, LLVM can sink the address 68 ; calculation and code gen can generate a better addressing mode for the second 69 ; use. 70 define void @test_GEP_across_BB(%class.my* %this, i64 %idx) { 71 %1 = getelementptr %class.my, %class.my* %this, i64 0, i32 3, i64 %idx, i32 1 72 %2 = load i32, i32* %1, align 4 73 %3 = getelementptr %class.my, %class.my* %this, i64 0, i32 3, i64 %idx, i32 2 74 %4 = load i32, i32* %3, align 4 75 %5 = icmp eq i32 %2, %4 76 br i1 %5, label %if.true, label %exit 77 78 if.true: 79 %6 = shl i32 %4, 1 80 store i32 %6, i32* %3, align 4 81 br label %exit 82 83 exit: 84 %7 = add nsw i32 %4, 1 85 store i32 %7, i32* %1, align 4 86 ret void 87 } 88 ; CHECK-LABEL: test_GEP_across_BB: 89 ; CHECK: ldr {{w[0-9]+}}, [{{x[0-9]+}}, #528] 90 ; CHECK: ldr {{w[0-9]+}}, [{{x[0-9]+}}, #532] 91 ; CHECK-NOT: add 92 ; CHECK: str {{w[0-9]+}}, [{{x[0-9]+}}, #532] 93 ; CHECK: str {{w[0-9]+}}, [{{x[0-9]+}}, #528] 94 95 ; CHECK-NoAA-LABEL: test_GEP_across_BB( 96 ; CHECK-NoAA: add i64 [[TMP:%[a-zA-Z0-9]+]], 528 97 ; CHECK-NoAA: add i64 [[TMP]], 532 98 ; CHECK-NoAA: if.true: 99 ; CHECK-NoAA: inttoptr 100 ; CHECK-NoAA: bitcast 101 ; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 532 102 ; CHECK-NoAA: exit: 103 ; CHECK-NoAA: inttoptr 104 ; CHECK-NoAA: bitcast 105 ; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 528 106 107 ; CHECK-UseAA-LABEL: test_GEP_across_BB( 108 ; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr 109 ; CHECK-UseAA: getelementptr i8, i8* [[PTR0]], i64 528 110 ; CHECK-UseAA: getelementptr i8, i8* [[PTR0]], i64 532 111 ; CHECK-UseAA: if.true: 112 ; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, i8* [[PTR0]], i64 532 113 ; CHECK-UseAA: exit: 114 ; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, i8* [[PTR0]], i64 528 115 116 %struct.S = type { float, double } 117 @struct_array = global [1024 x %struct.S] zeroinitializer, align 16 118 119 ; The following two test cases check we can extract constant from indices of 120 ; struct type. 121 ; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the 122 ; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field 123 ; offset is 8, the total constant offset is (5 * 16 + 8) = 88. 124 define double* @test-struct_1(i32 %i) { 125 entry: 126 %add = add nsw i32 %i, 5 127 %idxprom = sext i32 %add to i64 128 %p = getelementptr [1024 x %struct.S], [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1 129 ret double* %p 130 } 131 ; CHECK-NoAA-LABEL: @test-struct_1( 132 ; CHECK-NoAA-NOT: getelementptr 133 ; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, 88 134 135 ; CHECK-UseAA-LABEL: @test-struct_1( 136 ; CHECK-UseAA: getelementptr i8, i8* %{{[a-zA-Z0-9]+}}, i64 88 137 138 %struct3 = type { i64, i32 } 139 %struct2 = type { %struct3, i32 } 140 %struct1 = type { i64, %struct2 } 141 %struct0 = type { i32, i32, i64*, [100 x %struct1] } 142 143 ; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1". 144 ; "i32 3" is the 4th element whose field offset is 16. The alloca size of 145 ; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the 146 ; total constant offset is 16 + (-2 * 32) + 8 = -40 147 define %struct2* @test-struct_2(%struct0* %ptr, i64 %idx) { 148 entry: 149 %arrayidx = add nsw i64 %idx, -2 150 %ptr2 = getelementptr %struct0, %struct0* %ptr, i64 0, i32 3, i64 %arrayidx, i32 1 151 ret %struct2* %ptr2 152 } 153 ; CHECK-NoAA-LABEL: @test-struct_2( 154 ; CHECK-NoAA-NOT: = getelementptr 155 ; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, -40 156 157 ; CHECK-UseAA-LABEL: @test-struct_2( 158 ; CHECK-UseAA: getelementptr i8, i8* %{{[a-zA-Z0-9]+}}, i64 -40 159 160 ; Test that when a index is added from two constant, SeparateConstOffsetFromGEP 161 ; pass does not generate incorrect result. 162 define void @test_const_add([3 x i32]* %in) { 163 %inc = add nsw i32 2, 1 164 %idxprom = sext i32 %inc to i64 165 %arrayidx = getelementptr [3 x i32], [3 x i32]* %in, i64 %idxprom, i64 2 166 store i32 0, i32* %arrayidx, align 4 167 ret void 168 } 169 ; CHECK-LABEL: test_const_add: 170 ; CHECK: str wzr, [x0, #44] 171
