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([240 x %struct]* %string, i32* %adj, i32 %lib, i64 %idxprom) { 16 %liberties = getelementptr [240 x %struct], [240 x %struct]* %string, i64 1, i64 %idxprom, i32 3 17 %1 = load i32, i32* %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], [240 x %struct]* %string, i64 1, i64 %idxprom, i32 2 23 %2 = load i32, i32* %origin, align 4 24 store i32 %2, i32* %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: [[PTR0:%[a-zA-Z0-9]+]] = bitcast [240 x %struct]* %string to i8* 33; CHECK-UseAA: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96 34; CHECK-UseAA: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8, i8* [[PTR0]], i64 [[IDX]] 35; CHECK-UseAA: getelementptr i8, i8* [[PTR1]], i64 23052 36; CHECK-UseAA: bitcast 37; CHECK-UseAA: if.then: 38; CHECK-UseAA: getelementptr i8, i8* [[PTR1]], i64 23048 39; CHECK-UseAA: bitcast 40 41%class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]} 42%struct.pt = type { %struct.point*, i32, i32 } 43%struct.point = type { i32, i32 } 44 45; Check when a GEP is used across two basic block, LLVM can sink the address 46; calculation and code gen can generate a better addressing mode for the second 47; use. 48define void @test_GEP_across_BB(%class.my* %this, i64 %idx) { 49 %1 = getelementptr %class.my, %class.my* %this, i64 0, i32 3, i64 %idx, i32 1 50 %2 = load i32, i32* %1, align 4 51 %3 = getelementptr %class.my, %class.my* %this, i64 0, i32 3, i64 %idx, i32 2 52 %4 = load i32, i32* %3, align 4 53 %5 = icmp eq i32 %2, %4 54 br i1 %5, label %if.true, label %exit 55 56if.true: 57 %6 = shl i32 %4, 1 58 store i32 %6, i32* %3, align 4 59 br label %exit 60 61exit: 62 %7 = add nsw i32 %4, 1 63 store i32 %7, i32* %1, align 4 64 ret void 65} 66; CHECK-LABEL: test_GEP_across_BB: 67; CHECK-NOT: lwzu 68; CHECK: blr 69 70; CHECK-UseAA-LABEL: test_GEP_across_BB( 71; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr 72; CHECK-UseAA: getelementptr i8, i8* [[PTR0]], i64 528 73; CHECK-UseAA: getelementptr i8, i8* [[PTR0]], i64 532 74; CHECK-UseAA: if.true: 75; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, i8* [[PTR0]], i64 532 76; CHECK-UseAA: exit: 77; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8, i8* [[PTR0]], i64 528 78 79%struct.S = type { float, double } 80@struct_array = global [1024 x %struct.S] zeroinitializer, align 16 81 82; The following two test cases check we can extract constant from indices of 83; struct type. 84; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the 85; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field 86; offset is 8, the total constant offset is (5 * 16 + 8) = 88. 87define double* @test-struct_1(i32 %i) { 88entry: 89 %add = add nsw i32 %i, 5 90 %idxprom = sext i32 %add to i64 91 %p = getelementptr [1024 x %struct.S], [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1 92 ret double* %p 93} 94; CHECK-UseAA-LABEL: @test-struct_1( 95; CHECK-UseAA: getelementptr i8, i8* %{{[a-zA-Z0-9]+}}, i64 88 96 97%struct3 = type { i64, i32 } 98%struct2 = type { %struct3, i32 } 99%struct1 = type { i64, %struct2 } 100%struct0 = type { i32, i32, i64*, [100 x %struct1] } 101 102; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1". 103; "i32 3" is the 4th element whose field offset is 16. The alloca size of 104; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the 105; total constant offset is 16 + (-2 * 32) + 8 = -40 106define %struct2* @test-struct_2(%struct0* %ptr, i64 %idx) { 107entry: 108 %arrayidx = add nsw i64 %idx, -2 109 %ptr2 = getelementptr %struct0, %struct0* %ptr, i64 0, i32 3, i64 %arrayidx, i32 1 110 ret %struct2* %ptr2 111} 112; CHECK-UseAA-LABEL: @test-struct_2( 113; CHECK-UseAA: getelementptr i8, i8* %{{[a-zA-Z0-9]+}}, i64 -40 114 115; Test that when a index is added from two constant, SeparateConstOffsetFromGEP 116; pass does not generate incorrect result. 117define void @test_const_add([3 x i32]* %in) { 118 %inc = add nsw i32 2, 1 119 %idxprom = sext i32 %inc to i64 120 %arrayidx = getelementptr [3 x i32], [3 x i32]* %in, i64 %idxprom, i64 2 121 store i32 0, i32* %arrayidx, align 4 122 ret void 123} 124; CHECK-LABEL: test_const_add: 125; CHECK: li [[REG:[0-9]+]], 0 126; CHECK: stw [[REG]], 44(3) 127; CHECK: blr 128 129