1; RUN: llc -O3 -aarch64-enable-gep-opt=true -verify-machineinstrs %s -o - | FileCheck %s 2; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare < %s 2>&1 | FileCheck --check-prefix=CHECK-UseAA %s 3; RUN: llc -O3 -aarch64-enable-gep-opt=true -aarch64-use-aa=false -print-after=codegenprepare < %s 2>&1 | FileCheck --check-prefix=CHECK-NoAA %s 4; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cyclone < %s 2>&1 | FileCheck --check-prefix=CHECK-UseAA %s 5; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cortex-a53 < %s 2>&1 | FileCheck --check-prefix=CHECK-UseAA %s 6 7target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128" 8target triple = "aarch64" 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; eliminate the common subexpression for the second use. 19define void @test_GEP_CSE(ptr %string, ptr %adj, i32 %lib, i64 %idxprom) { 20 %liberties = getelementptr [240 x %struct], ptr %string, i64 1, i64 %idxprom, i32 3 21 %1 = load i32, ptr %liberties, align 4 22 %cmp = icmp eq i32 %1, %lib 23 br i1 %cmp, label %if.then, label %if.end 24 25if.then: ; preds = %entry 26 %origin = getelementptr [240 x %struct], ptr %string, i64 1, i64 %idxprom, i32 2 27 %2 = load i32, ptr %origin, align 4 28 store i32 %2, ptr %adj, align 4 29 br label %if.end 30 31if.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 ptr %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: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96 55; CHECK-UseAA: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8, ptr %string, i64 [[IDX]] 56; CHECK-UseAA: getelementptr i8, ptr [[PTR1]], i64 23052 57; CHECK-UseAA: if.then: 58; CHECK-UseAA: getelementptr i8, ptr [[PTR1]], i64 23048 59 60%class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]} 61%struct.pt = type { ptr, 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(ptr %this, i64 %idx) { 68 %1 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 1 69 %2 = load i32, ptr %1, align 4 70 %3 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 2 71 %4 = load i32, ptr %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, ptr %3, align 4 78 br label %exit 79 80exit: 81 %7 = add nsw i32 %4, 1 82 store i32 %7, ptr %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: inttoptr 97; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 532 98; CHECK-NoAA: exit: 99; CHECK-NoAA: inttoptr 100; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 528 101 102; CHECK-UseAA-LABEL: test_GEP_across_BB( 103; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr 104; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 528 105; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 532 106; CHECK-UseAA: if.true: 107; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 532 108; CHECK-UseAA: exit: 109; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 528 110 111%struct.S = type { float, double } 112@struct_array = global [1024 x %struct.S] zeroinitializer, align 16 113 114; The following two test cases check we can extract constant from indices of 115; struct type. 116; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the 117; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field 118; offset is 8, the total constant offset is (5 * 16 + 8) = 88. 119define ptr @test-struct_1(i32 %i) { 120entry: 121 %add = add nsw i32 %i, 5 122 %idxprom = sext i32 %add to i64 123 %p = getelementptr [1024 x %struct.S], ptr @struct_array, i64 0, i64 %idxprom, i32 1 124 ret ptr %p 125} 126; CHECK-NoAA-LABEL: @test-struct_1( 127; CHECK-NoAA-NOT: getelementptr 128; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, 88 129 130; CHECK-UseAA-LABEL: @test-struct_1( 131; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 88 132 133%struct3 = type { i64, i32 } 134%struct2 = type { %struct3, i32 } 135%struct1 = type { i64, %struct2 } 136%struct0 = type { i32, i32, ptr, [100 x %struct1] } 137 138; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1". 139; "i32 3" is the 4th element whose field offset is 16. The alloca size of 140; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the 141; total constant offset is 16 + (-2 * 32) + 8 = -40 142define ptr @test-struct_2(ptr %ptr, i64 %idx) { 143entry: 144 %arrayidx = add nsw i64 %idx, -2 145 %ptr2 = getelementptr %struct0, ptr %ptr, i64 0, i32 3, i64 %arrayidx, i32 1 146 ret ptr %ptr2 147} 148; CHECK-NoAA-LABEL: @test-struct_2( 149; CHECK-NoAA-NOT: = getelementptr 150; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, -40 151 152; CHECK-UseAA-LABEL: @test-struct_2( 153; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 -40 154 155; Test that when a index is added from two constant, SeparateConstOffsetFromGEP 156; pass does not generate incorrect result. 157define void @test_const_add(ptr %in) { 158 %inc = add nsw i32 2, 1 159 %idxprom = sext i32 %inc to i64 160 %arrayidx = getelementptr [3 x i32], ptr %in, i64 %idxprom, i64 2 161 store i32 0, ptr %arrayidx, align 4 162 ret void 163} 164; CHECK-LABEL: test_const_add: 165; CHECK: str wzr, [x0, #44] 166