154 lines · plain
1; RUN: llc -O3 -aarch64-enable-gep-opt=true -verify-machineinstrs %s -o - | FileCheck %s2; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare < %s 2>&1 | FileCheck --check-prefix=CHECK-IR %s3; RUN: llc -O3 -aarch64-enable-gep-opt=true -aarch64-use-aa=false -print-after=codegenprepare < %s 2>&1 | FileCheck --check-prefix=CHECK-IR %s4; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cyclone < %s 2>&1 | FileCheck --check-prefix=CHECK-IR %s5; RUN: llc -O3 -aarch64-enable-gep-opt=true -print-after=codegenprepare -mcpu=cortex-a53 < %s 2>&1 | FileCheck --check-prefix=CHECK-IR %s6 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 AArch6411; backend. If useAA() returns true, it will lower a GEP with multiple indices12; into GEPs with a single index, otherwise it will lower it into a13; "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 can18; 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 321 %1 = load i32, ptr %liberties, align 422 %cmp = icmp eq i32 %1, %lib23 br i1 %cmp, label %if.then, label %if.end24 25if.then: ; preds = %entry26 %origin = getelementptr [240 x %struct], ptr %string, i64 1, i64 %idxprom, i32 227 %2 = load i32, ptr %origin, align 428 store i32 %2, ptr %adj, align 429 br label %if.end30 31if.end: ; preds = %if.then, %entry32 ret void33}34 35; CHECK-LABEL: test_GEP_CSE:36; CHECK: madd37; CHECK: ldr38; CHECK-NOT: madd39; CHECK:ldr40 41; CHECK-IR-LABEL: @test_GEP_CSE(42; CHECK-IR: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 9643; CHECK-IR: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8, ptr %string, i64 [[IDX]]44; CHECK-IR: getelementptr i8, ptr [[PTR1]], i64 2305245; CHECK-IR: if.then:46; CHECK-IR: getelementptr i8, ptr [[PTR1]], i64 2304847 48%class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]}49%struct.pt = type { ptr, i32, i32 }50%struct.point = type { i32, i32 }51 52; Check when a GEP is used across two basic block, LLVM can sink the address53; calculation and code gen can generate a better addressing mode for the second54; use.55define void @test_GEP_across_BB(ptr %this, i64 %idx) {56 %1 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 157 %2 = load i32, ptr %1, align 458 %3 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 259 %4 = load i32, ptr %3, align 460 %5 = icmp eq i32 %2, %461 br i1 %5, label %if.true, label %exit62 63if.true:64 %6 = shl i32 %4, 165 store i32 %6, ptr %3, align 466 br label %exit67 68exit:69 %7 = add nsw i32 %4, 170 store i32 %7, ptr %1, align 471 ret void72}73; CHECK-LABEL: test_GEP_across_BB:74; CHECK: ldr {{w[0-9]+}}, [{{x[0-9]+}}, #528]75; CHECK: ldr {{w[0-9]+}}, [{{x[0-9]+}}, #532]76; CHECK-NOT: add77; CHECK: str {{w[0-9]+}}, [{{x[0-9]+}}, #532]78; CHECK: str {{w[0-9]+}}, [{{x[0-9]+}}, #528]79 80; CHECK-NoAA-LABEL: test_GEP_across_BB(81; CHECK-NoAA: add i64 [[TMP:%[a-zA-Z0-9]+]], 52882; CHECK-NoAA: add i64 [[TMP]], 53283; CHECK-NoAA: if.true:84; CHECK-NoAA: inttoptr85; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 53286; CHECK-NoAA: exit:87; CHECK-NoAA: inttoptr88; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, {{.*}}, i64 52889 90; CHECK-UseAA-LABEL: test_GEP_across_BB(91; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr92; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 52893; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 53294; CHECK-UseAA: if.true:95; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 53296; CHECK-UseAA: exit:97; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 52898 99%struct.S = type { float, double }100@struct_array = global [1024 x %struct.S] zeroinitializer, align 16101 102; The following two test cases check we can extract constant from indices of103; struct type.104; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the105; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field106; offset is 8, the total constant offset is (5 * 16 + 8) = 88.107define ptr @test-struct_1(i32 %i) {108entry:109 %add = add nsw i32 %i, 5110 %idxprom = sext i32 %add to i64111 %p = getelementptr [1024 x %struct.S], ptr @struct_array, i64 0, i64 %idxprom, i32 1112 ret ptr %p113}114; CHECK-NoAA-LABEL: @test-struct_1(115; CHECK-NoAA-NOT: getelementptr116; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, 88117 118; CHECK-UseAA-LABEL: @test-struct_1(119; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 88120 121%struct3 = type { i64, i32 }122%struct2 = type { %struct3, i32 }123%struct1 = type { i64, %struct2 }124%struct0 = type { i32, i32, ptr, [100 x %struct1] }125 126; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1".127; "i32 3" is the 4th element whose field offset is 16. The alloca size of128; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the129; total constant offset is 16 + (-2 * 32) + 8 = -40130define ptr @test-struct_2(ptr %ptr, i64 %idx) {131entry:132 %arrayidx = add nsw i64 %idx, -2133 %ptr2 = getelementptr %struct0, ptr %ptr, i64 0, i32 3, i64 %arrayidx, i32 1134 ret ptr %ptr2135}136; CHECK-NoAA-LABEL: @test-struct_2(137; CHECK-NoAA-NOT: = getelementptr138; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, -40139 140; CHECK-UseAA-LABEL: @test-struct_2(141; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 -40142 143; Test that when a index is added from two constant, SeparateConstOffsetFromGEP144; pass does not generate incorrect result.145define void @test_const_add(ptr %in) {146 %inc = add nsw i32 2, 1147 %idxprom = sext i32 %inc to i64148 %arrayidx = getelementptr [3 x i32], ptr %in, i64 %idxprom, i64 2149 store i32 0, ptr %arrayidx, align 4150 ret void151}152; CHECK-LABEL: test_const_add:153; CHECK: str wzr, [x0, #44]154