1; RUN: llc -verify-machineinstrs -O3 -mcpu=pwr7 < %s | FileCheck %s 2; RUN: llc -verify-machineinstrs -O3 -print-after=codegenprepare -mcpu=pwr7 < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-UseAA <%t %s 3target datalayout = "E-m:e-i64:64-n32:64" 4target triple = "powerpc64-unknown-linux-gnu" 5 6; Following test cases test enabling SeparateConstOffsetFromGEP pass in the PPC 7; backend. If useAA() returns true, it will lower a GEP with multiple indices 8; into GEPs with a single index, otherwise it will lower it into a 9; "ptrtoint+arithmetics+inttoptr" form. 10 11%struct = type { i32, i32, i32, i32, [20 x i32] } 12 13; Check that when two complex GEPs are used in two basic blocks, LLVM can 14; elimilate the common subexpression for the second use. 15define void @test_GEP_CSE(ptr %string, ptr %adj, i32 %lib, i64 %idxprom) { 16 %liberties = getelementptr [240 x %struct], ptr %string, i64 1, i64 %idxprom, i32 3 17 %1 = load i32, ptr %liberties, align 4 18 %cmp = icmp eq i32 %1, %lib 19 br i1 %cmp, label %if.then, label %if.end 20 21if.then: ; preds = %entry 22 %origin = getelementptr [240 x %struct], ptr %string, i64 1, i64 %idxprom, i32 2 23 %2 = load i32, ptr %origin, align 4 24 store i32 %2, ptr %adj, align 4 25 br label %if.end 26 27if.end: ; preds = %if.then, %entry 28 ret void 29} 30 31; CHECK-UseAA-LABEL: @test_GEP_CSE( 32; CHECK-UseAA: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96 33; CHECK-UseAA: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8, ptr %string, i64 [[IDX]] 34; CHECK-UseAA: getelementptr i8, ptr [[PTR1]], i64 23052 35; CHECK-UseAA: if.then: 36; CHECK-UseAA: getelementptr i8, ptr [[PTR1]], i64 23048 37 38%class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]} 39%struct.pt = type { ptr, i32, i32 } 40%struct.point = type { i32, i32 } 41 42; Check when a GEP is used across two basic block, LLVM can sink the address 43; calculation and code gen can generate a better addressing mode for the second 44; use. 45define void @test_GEP_across_BB(ptr %this, i64 %idx) { 46 %1 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 1 47 %2 = load i32, ptr %1, align 4 48 %3 = getelementptr %class.my, ptr %this, i64 0, i32 3, i64 %idx, i32 2 49 %4 = load i32, ptr %3, align 4 50 %5 = icmp eq i32 %2, %4 51 br i1 %5, label %if.true, label %exit 52 53if.true: 54 %6 = shl i32 %4, 1 55 store i32 %6, ptr %3, align 4 56 br label %exit 57 58exit: 59 %7 = add nsw i32 %4, 1 60 store i32 %7, ptr %1, align 4 61 ret void 62} 63; CHECK-LABEL: test_GEP_across_BB: 64; CHECK-NOT: lwzu 65; CHECK: blr 66 67; CHECK-UseAA-LABEL: test_GEP_across_BB( 68; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr 69; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 528 70; CHECK-UseAA: getelementptr i8, ptr [[PTR0]], i64 532 71; CHECK-UseAA: if.true: 72; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 532 73; CHECK-UseAA: exit: 74; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, ptr [[PTR0]], i64 528 75 76%struct.S = type { float, double } 77@struct_array = global [1024 x %struct.S] zeroinitializer, align 16 78 79; The following two test cases check we can extract constant from indices of 80; struct type. 81; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the 82; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field 83; offset is 8, the total constant offset is (5 * 16 + 8) = 88. 84define ptr @test-struct_1(i32 %i) { 85entry: 86 %add = add nsw i32 %i, 5 87 %idxprom = sext i32 %add to i64 88 %p = getelementptr [1024 x %struct.S], ptr @struct_array, i64 0, i64 %idxprom, i32 1 89 ret ptr %p 90} 91; CHECK-UseAA-LABEL: @test-struct_1( 92; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 88 93 94%struct3 = type { i64, i32 } 95%struct2 = type { %struct3, i32 } 96%struct1 = type { i64, %struct2 } 97%struct0 = type { i32, i32, ptr, [100 x %struct1] } 98 99; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1". 100; "i32 3" is the 4th element whose field offset is 16. The alloca size of 101; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the 102; total constant offset is 16 + (-2 * 32) + 8 = -40 103define ptr @test-struct_2(ptr %ptr, i64 %idx) { 104entry: 105 %arrayidx = add nsw i64 %idx, -2 106 %ptr2 = getelementptr %struct0, ptr %ptr, i64 0, i32 3, i64 %arrayidx, i32 1 107 ret ptr %ptr2 108} 109; CHECK-UseAA-LABEL: @test-struct_2( 110; CHECK-UseAA: getelementptr i8, ptr %{{[a-zA-Z0-9]+}}, i64 -40 111 112; Test that when a index is added from two constant, SeparateConstOffsetFromGEP 113; pass does not generate incorrect result. 114define void @test_const_add(ptr %in) { 115 %inc = add nsw i32 2, 1 116 %idxprom = sext i32 %inc to i64 117 %arrayidx = getelementptr [3 x i32], ptr %in, i64 %idxprom, i64 2 118 store i32 0, ptr %arrayidx, align 4 119 ret void 120} 121; CHECK-LABEL: test_const_add: 122; CHECK: li [[REG:[0-9]+]], 0 123; CHECK: stw [[REG]], 44(3) 124; CHECK: blr 125 126