201 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --include-generated-funcs2; RUN: opt -S -passes=verify,iroutliner -ir-outlining-no-cost < %s | FileCheck %s3 4; Here we have multiple exits, but the different sources, same outputs are5; needed, this checks that they are compressed, and moved into the appropriate6; output blocks.7 8define void @outline_outputs1() #0 {9entry:10 %output = alloca i32, align 411 %result = alloca i32, align 412 %output2 = alloca i32, align 413 %result2 = alloca i32, align 414 %a = alloca i32, align 415 %b = alloca i32, align 416 br label %block_217block_1:18 %a2 = alloca i32, align 419 %b2 = alloca i32, align 420 br label %block_221block_2:22 %a2val = load i32, ptr %a23 %b2val = load i32, ptr %b24 %add2 = add i32 2, %a2val25 %mul2 = mul i32 2, %b2val26 br label %block_527block_3:28 %aval = load i32, ptr %a29 %bval = load i32, ptr %b30 %add = add i32 2, %aval31 %mul = mul i32 2, %bval32 br label %block_433block_4:34 store i32 %add, ptr %output, align 435 store i32 %mul, ptr %result, align 436 br label %block_637block_5:38 store i32 %add2, ptr %output, align 439 store i32 %mul2, ptr %result, align 440 br label %block_741block_6:42 %div = udiv i32 %aval, %bval43 ret void44block_7:45 %sub = sub i32 %a2val, %b2val46 ret void47}48 49define void @outline_outputs2() #0 {50entry:51 %output = alloca i32, align 452 %result = alloca i32, align 453 %output2 = alloca i32, align 454 %result2 = alloca i32, align 455 %a = alloca i32, align 456 %b = alloca i32, align 457 br label %block_258block_1:59 %a2 = alloca i32, align 460 %b2 = alloca i32, align 461 br label %block_262block_2:63 %a2val = load i32, ptr %a64 %b2val = load i32, ptr %b65 %add2 = add i32 2, %a2val66 %mul2 = mul i32 2, %b2val67 br label %block_568block_3:69 %aval = load i32, ptr %a70 %bval = load i32, ptr %b71 %add = add i32 2, %aval72 %mul = mul i32 2, %bval73 br label %block_474block_4:75 store i32 %add, ptr %output, align 476 store i32 %mul, ptr %result, align 477 br label %block_778block_5:79 store i32 %add2, ptr %output, align 480 store i32 %mul2, ptr %result, align 481 br label %block_682block_6:83 %diff = sub i32 %a2val, %b2val84 ret void85block_7:86 %quot = udiv i32 %aval, %bval87 ret void88}89; CHECK-LABEL: @outline_outputs1(90; CHECK-NEXT: entry:91; CHECK-NEXT: [[BVAL_LOC:%.*]] = alloca i32, align 492; CHECK-NEXT: [[AVAL_LOC:%.*]] = alloca i32, align 493; CHECK-NEXT: [[B2VAL_LOC:%.*]] = alloca i32, align 494; CHECK-NEXT: [[A2VAL_LOC:%.*]] = alloca i32, align 495; CHECK-NEXT: [[OUTPUT:%.*]] = alloca i32, align 496; CHECK-NEXT: [[RESULT:%.*]] = alloca i32, align 497; CHECK-NEXT: [[OUTPUT2:%.*]] = alloca i32, align 498; CHECK-NEXT: [[RESULT2:%.*]] = alloca i32, align 499; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4100; CHECK-NEXT: [[B:%.*]] = alloca i32, align 4101; CHECK-NEXT: br label [[BLOCK_2:%.*]]102; CHECK: block_1:103; CHECK-NEXT: [[A2:%.*]] = alloca i32, align 4104; CHECK-NEXT: [[B2:%.*]] = alloca i32, align 4105; CHECK-NEXT: br label [[BLOCK_2]]106; CHECK: block_2:107; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[A2VAL_LOC]])108; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[B2VAL_LOC]])109; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[AVAL_LOC]])110; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[BVAL_LOC]])111; CHECK-NEXT: [[TARGETBLOCK:%.*]] = call i1 @outlined_ir_func_0(ptr [[A]], ptr [[B]], ptr [[OUTPUT]], ptr [[RESULT]], ptr [[A2VAL_LOC]], ptr [[B2VAL_LOC]], ptr [[AVAL_LOC]], ptr [[BVAL_LOC]])112; CHECK-NEXT: [[A2VAL_RELOAD:%.*]] = load i32, ptr [[A2VAL_LOC]], align 4113; CHECK-NEXT: [[B2VAL_RELOAD:%.*]] = load i32, ptr [[B2VAL_LOC]], align 4114; CHECK-NEXT: [[AVAL_RELOAD:%.*]] = load i32, ptr [[AVAL_LOC]], align 4115; CHECK-NEXT: [[BVAL_RELOAD:%.*]] = load i32, ptr [[BVAL_LOC]], align 4116; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[A2VAL_LOC]])117; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[B2VAL_LOC]])118; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[AVAL_LOC]])119; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[BVAL_LOC]])120; CHECK-NEXT: br i1 [[TARGETBLOCK]], label [[BLOCK_6:%.*]], label [[BLOCK_7:%.*]]121; CHECK: block_6:122; CHECK-NEXT: [[DIV:%.*]] = udiv i32 [[AVAL_RELOAD]], [[BVAL_RELOAD]]123; CHECK-NEXT: ret void124; CHECK: block_7:125; CHECK-NEXT: [[SUB:%.*]] = sub i32 [[A2VAL_RELOAD]], [[B2VAL_RELOAD]]126; CHECK-NEXT: ret void127;128;129; CHECK-LABEL: @outline_outputs2(130; CHECK-NEXT: entry:131; CHECK-NEXT: [[BVAL_LOC:%.*]] = alloca i32, align 4132; CHECK-NEXT: [[AVAL_LOC:%.*]] = alloca i32, align 4133; CHECK-NEXT: [[B2VAL_LOC:%.*]] = alloca i32, align 4134; CHECK-NEXT: [[A2VAL_LOC:%.*]] = alloca i32, align 4135; CHECK-NEXT: [[OUTPUT:%.*]] = alloca i32, align 4136; CHECK-NEXT: [[RESULT:%.*]] = alloca i32, align 4137; CHECK-NEXT: [[OUTPUT2:%.*]] = alloca i32, align 4138; CHECK-NEXT: [[RESULT2:%.*]] = alloca i32, align 4139; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4140; CHECK-NEXT: [[B:%.*]] = alloca i32, align 4141; CHECK-NEXT: br label [[BLOCK_2:%.*]]142; CHECK: block_1:143; CHECK-NEXT: [[A2:%.*]] = alloca i32, align 4144; CHECK-NEXT: [[B2:%.*]] = alloca i32, align 4145; CHECK-NEXT: br label [[BLOCK_2]]146; CHECK: block_2:147; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[A2VAL_LOC]])148; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[B2VAL_LOC]])149; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[AVAL_LOC]])150; CHECK-NEXT: call void @llvm.lifetime.start.p0(ptr [[BVAL_LOC]])151; CHECK-NEXT: [[TARGETBLOCK:%.*]] = call i1 @outlined_ir_func_0(ptr [[A]], ptr [[B]], ptr [[OUTPUT]], ptr [[RESULT]], ptr [[A2VAL_LOC]], ptr [[B2VAL_LOC]], ptr [[AVAL_LOC]], ptr [[BVAL_LOC]])152; CHECK-NEXT: [[A2VAL_RELOAD:%.*]] = load i32, ptr [[A2VAL_LOC]], align 4153; CHECK-NEXT: [[B2VAL_RELOAD:%.*]] = load i32, ptr [[B2VAL_LOC]], align 4154; CHECK-NEXT: [[AVAL_RELOAD:%.*]] = load i32, ptr [[AVAL_LOC]], align 4155; CHECK-NEXT: [[BVAL_RELOAD:%.*]] = load i32, ptr [[BVAL_LOC]], align 4156; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[A2VAL_LOC]])157; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[B2VAL_LOC]])158; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[AVAL_LOC]])159; CHECK-NEXT: call void @llvm.lifetime.end.p0(ptr [[BVAL_LOC]])160; CHECK-NEXT: br i1 [[TARGETBLOCK]], label [[BLOCK_7:%.*]], label [[BLOCK_6:%.*]]161; CHECK: block_6:162; CHECK-NEXT: [[DIFF:%.*]] = sub i32 [[A2VAL_RELOAD]], [[B2VAL_RELOAD]]163; CHECK-NEXT: ret void164; CHECK: block_7:165; CHECK-NEXT: [[QUOT:%.*]] = udiv i32 [[AVAL_RELOAD]], [[BVAL_RELOAD]]166; CHECK-NEXT: ret void167;168;169; CHECK: define internal i1 @outlined_ir_func_0(170; CHECK-NEXT: newFuncRoot:171; CHECK-NEXT: br label [[BLOCK_2_TO_OUTLINE:%.*]]172; CHECK: block_2_to_outline:173; CHECK-NEXT: [[A2VAL:%.*]] = load i32, ptr [[TMP0:%.*]], align 4174; CHECK-NEXT: [[B2VAL:%.*]] = load i32, ptr [[TMP1:%.*]], align 4175; CHECK-NEXT: [[ADD2:%.*]] = add i32 2, [[A2VAL]]176; CHECK-NEXT: [[MUL2:%.*]] = mul i32 2, [[B2VAL]]177; CHECK-NEXT: br label [[BLOCK_5:%.*]]178; CHECK: block_3:179; CHECK-NEXT: [[AVAL:%.*]] = load i32, ptr [[TMP0]], align 4180; CHECK-NEXT: [[BVAL:%.*]] = load i32, ptr [[TMP1]], align 4181; CHECK-NEXT: [[ADD:%.*]] = add i32 2, [[AVAL]]182; CHECK-NEXT: [[MUL:%.*]] = mul i32 2, [[BVAL]]183; CHECK-NEXT: br label [[BLOCK_4:%.*]]184; CHECK: block_4:185; CHECK-NEXT: store i32 [[ADD]], ptr [[TMP2:%.*]], align 4186; CHECK-NEXT: store i32 [[MUL]], ptr [[TMP3:%.*]], align 4187; CHECK-NEXT: br label [[BLOCK_6_EXITSTUB:%.*]]188; CHECK: block_5:189; CHECK-NEXT: store i32 [[ADD2]], ptr [[TMP2]], align 4190; CHECK-NEXT: store i32 [[MUL2]], ptr [[TMP3]], align 4191; CHECK-NEXT: br label [[BLOCK_7_EXITSTUB:%.*]]192; CHECK: block_6.exitStub:193; CHECK-NEXT: store i32 [[AVAL]], ptr [[TMP6:%.*]], align 4194; CHECK-NEXT: store i32 [[BVAL]], ptr [[TMP7:%.*]], align 4195; CHECK-NEXT: ret i1 true196; CHECK: block_7.exitStub:197; CHECK-NEXT: store i32 [[A2VAL]], ptr [[TMP4:%.*]], align 4198; CHECK-NEXT: store i32 [[B2VAL]], ptr [[TMP5:%.*]], align 4199; CHECK-NEXT: ret i1 false200;201