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