377 lines · plain
1; RUN: opt < %s -passes=inline -inline-threshold=20 -S | FileCheck %s2; RUN: opt < %s -passes='cgscc(inline)' -inline-threshold=20 -S | FileCheck %s3 4define internal i32 @callee1(i32 %A, i32 %B) {5 %C = sdiv i32 %A, %B6 ret i32 %C7}8 9define i32 @caller1() {10; CHECK-LABEL: define i32 @caller1(11; CHECK-NEXT: ret i32 312 13 %X = call i32 @callee1( i32 10, i32 3 )14 ret i32 %X15}16 17define i32 @caller2() {18; Check that we can constant-prop through instructions after inlining callee2119; to get constants in the inlined callsite to callee22.20; FIXME: Currently, the threshold is fixed at 20 because we don't perform21; *recursive* cost analysis to realize that the nested call site will definitely22; inline and be cheap. We should eventually do that and lower the threshold here23; to 1.24;25; CHECK-LABEL: @caller2(26; CHECK-NOT: call void @callee227; CHECK: ret28 29 %x = call i32 @callee21(i32 42, i32 48)30 ret i32 %x31}32 33define i32 @callee21(i32 %x, i32 %y) {34 %sub = sub i32 %y, %x35 %result = call i32 @callee22(i32 %sub)36 ret i32 %result37}38 39declare ptr @getptr()40 41define i32 @callee22(i32 %x) {42 %icmp = icmp ugt i32 %x, 4243 br i1 %icmp, label %bb.true, label %bb.false44bb.true:45 ; This block musn't be counted in the inline cost.46 %x1 = add i32 %x, 147 %x2 = add i32 %x1, 148 %x3 = add i32 %x2, 149 %x4 = add i32 %x3, 150 %x5 = add i32 %x4, 151 %x6 = add i32 %x5, 152 %x7 = add i32 %x6, 153 %x8 = add i32 %x7, 154 55 ret i32 %x856bb.false:57 ret i32 %x58}59 60define i32 @caller3() {61; Check that even if the expensive path is hidden behind several basic blocks,62; it doesn't count toward the inline cost when constant-prop proves those paths63; dead.64;65; CHECK-LABEL: @caller3(66; CHECK-NOT: call67; CHECK: ret i32 668 69entry:70 %x = call i32 @callee3(i32 42, i32 48)71 ret i32 %x72}73 74define i32 @callee3(i32 %x, i32 %y) {75 %sub = sub i32 %y, %x76 %icmp = icmp ugt i32 %sub, 4277 br i1 %icmp, label %bb.true, label %bb.false78 79bb.true:80 %icmp2 = icmp ult i32 %sub, 6481 br i1 %icmp2, label %bb.true.true, label %bb.true.false82 83bb.true.true:84 ; This block musn't be counted in the inline cost.85 %x1 = add i32 %x, 186 %x2 = add i32 %x1, 187 %x3 = add i32 %x2, 188 %x4 = add i32 %x3, 189 %x5 = add i32 %x4, 190 %x6 = add i32 %x5, 191 %x7 = add i32 %x6, 192 %x8 = add i32 %x7, 193 br label %bb.merge94 95bb.true.false:96 ; This block musn't be counted in the inline cost.97 %y1 = add i32 %y, 198 %y2 = add i32 %y1, 199 %y3 = add i32 %y2, 1100 %y4 = add i32 %y3, 1101 %y5 = add i32 %y4, 1102 %y6 = add i32 %y5, 1103 %y7 = add i32 %y6, 1104 %y8 = add i32 %y7, 1105 br label %bb.merge106 107bb.merge:108 %result = phi i32 [ %x8, %bb.true.true ], [ %y8, %bb.true.false ]109 ret i32 %result110 111bb.false:112 ret i32 %sub113}114 115declare {i8, i1} @llvm.uadd.with.overflow.i8(i8 %a, i8 %b)116 117define i8 @caller4(i8 %z) {118; Check that we can constant fold through intrinsics such as the119; overflow-detecting arithmetic intrinsics. These are particularly important120; as they are used heavily in standard library code and generic C++ code where121; the arguments are oftent constant but complete generality is required.122;123; CHECK-LABEL: @caller4(124; CHECK-NOT: call125; CHECK: ret i8 -1126 127entry:128 %x = call i8 @callee4(i8 254, i8 14, i8 %z)129 ret i8 %x130}131 132define i8 @callee4(i8 %x, i8 %y, i8 %z) {133 %uadd = call {i8, i1} @llvm.uadd.with.overflow.i8(i8 %x, i8 %y)134 %o = extractvalue {i8, i1} %uadd, 1135 br i1 %o, label %bb.true, label %bb.false136 137bb.true:138 ret i8 -1139 140bb.false:141 ; This block musn't be counted in the inline cost.142 %z1 = add i8 %z, 1143 %z2 = add i8 %z1, 1144 %z3 = add i8 %z2, 1145 %z4 = add i8 %z3, 1146 %z5 = add i8 %z4, 1147 %z6 = add i8 %z5, 1148 %z7 = add i8 %z6, 1149 %z8 = add i8 %z7, 1150 ret i8 %z8151}152 153define i64 @caller5(i64 %y) {154; Check that we can round trip constants through various kinds of casts etc w/o155; losing track of the constant prop in the inline cost analysis.156;157; CHECK-LABEL: @caller5(158; CHECK-NOT: call159; CHECK: ret i64 -1160 161entry:162 %x = call i64 @callee5(i64 42, i64 %y)163 ret i64 %x164}165 166define i64 @callee5(i64 %x, i64 %y) {167 %inttoptr = inttoptr i64 %x to ptr168 %ptrtoint = ptrtoint ptr %inttoptr to i64169 %trunc = trunc i64 %ptrtoint to i32170 %zext = zext i32 %trunc to i64171 %cmp = icmp eq i64 %zext, 42172 br i1 %cmp, label %bb.true, label %bb.false173 174bb.true:175 ret i64 -1176 177bb.false:178 ; This block musn't be counted in the inline cost.179 %y1 = add i64 %y, 1180 %y2 = add i64 %y1, 1181 %y3 = add i64 %y2, 1182 %y4 = add i64 %y3, 1183 %y5 = add i64 %y4, 1184 %y6 = add i64 %y5, 1185 %y7 = add i64 %y6, 1186 %y8 = add i64 %y7, 1187 ret i64 %y8188}189 190define float @caller6() {191; Check that we can constant-prop through fcmp instructions192;193; CHECK-LABEL: @caller6(194; CHECK-NOT: call195; CHECK: ret196 %x = call float @callee6(float 42.0)197 ret float %x198}199 200define float @callee6(float %x) {201 %icmp = fcmp ugt float %x, 42.0202 br i1 %icmp, label %bb.true, label %bb.false203 204bb.true:205 ; This block musn't be counted in the inline cost.206 %x1 = fadd float %x, 1.0207 %x2 = fadd float %x1, 1.0208 %x3 = fadd float %x2, 1.0209 %x4 = fadd float %x3, 1.0210 %x5 = fadd float %x4, 1.0211 %x6 = fadd float %x5, 1.0212 %x7 = fadd float %x6, 1.0213 %x8 = fadd float %x7, 1.0214 ret float %x8215 216bb.false:217 ret float %x218}219 220 221 222define i32 @PR13412.main() {223; This is a somewhat complicated three layer subprogram that was reported to224; compute the wrong value for a branch due to assuming that an argument225; mid-inline couldn't be equal to another pointer.226;227; After inlining, the branch should point directly to the exit block, not to228; the intermediate block.229; CHECK: @PR13412.main230; CHECK: br i1 true, label %[[TRUE_DEST:.*]], label %[[FALSE_DEST:.*]]231; CHECK: [[FALSE_DEST]]:232; CHECK-NEXT: call void @PR13412.fail()233; CHECK: [[TRUE_DEST]]:234; CHECK-NEXT: ret i32 0235 236entry:237 %i1 = alloca i64238 store i64 0, ptr %i1239 %call = call i1 @PR13412.first(ptr %i1, ptr %i1)240 br i1 %call, label %cond.end, label %cond.false241 242cond.false:243 call void @PR13412.fail()244 br label %cond.end245 246cond.end:247 ret i32 0248}249 250define internal i1 @PR13412.first(ptr %a, ptr %b) {251entry:252 %call = call ptr @PR13412.second(ptr %a, ptr %b)253 %cmp = icmp eq ptr %call, %b254 ret i1 %cmp255}256 257declare void @PR13412.fail()258 259define internal ptr @PR13412.second(ptr %a, ptr %b) {260entry:261 %sub.ptr.lhs.cast = ptrtoint ptr %b to i64262 %sub.ptr.rhs.cast = ptrtoint ptr %a to i64263 %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast264 %sub.ptr.div = ashr exact i64 %sub.ptr.sub, 2265 %cmp = icmp ugt i64 %sub.ptr.div, 1266 br i1 %cmp, label %if.then, label %if.end3267 268if.then:269 %0 = load i32, ptr %a270 %1 = load i32, ptr %b271 %cmp1 = icmp eq i32 %0, %1272 br i1 %cmp1, label %return, label %if.end3273 274if.end3:275 br label %return276 277return:278 %retval.0 = phi ptr [ %b, %if.end3 ], [ %a, %if.then ]279 ret ptr %retval.0280}281 282declare i32 @PR28802.external(i32 returned %p1)283 284define internal i32 @PR28802.callee() {285entry:286 br label %cont287 288cont:289 %0 = phi i32 [ 0, %entry ]290 %call = call i32 @PR28802.external(i32 %0)291 ret i32 %call292}293 294define i32 @PR28802() {295entry:296 %call = call i32 @PR28802.callee()297 ret i32 %call298}299 300; CHECK-LABEL: define i32 @PR28802(301; CHECK: %[[call:.*]] = call i32 @PR28802.external(i32 0)302; CHECK: ret i32 %[[call]]303 304define internal i32 @PR28848.callee(i32 %p2, i1 %c) {305entry:306 br i1 %c, label %cond.end, label %cond.true307 308cond.true:309 br label %cond.end310 311cond.end:312 %cond = phi i32 [ 0, %cond.true ], [ %p2, %entry ]313 %or = or i32 %cond, %p2314 ret i32 %or315}316 317define i32 @PR28848() {318entry:319 %call = call i32 @PR28848.callee(i32 0, i1 false)320 ret i32 %call321}322; CHECK-LABEL: define i32 @PR28848(323; CHECK: ret i32 0324 325define internal void @callee7(i16 %param1, i16 %param2) {326entry:327 br label %bb328 329bb:330 %phi = phi i16 [ %param2, %entry ]331 %add = add i16 %phi, %param1332 ret void333}334 335declare i16 @caller7.external(i16 returned)336 337define void @caller7() {338bb1:339 %call = call i16 @caller7.external(i16 1)340 call void @callee7(i16 0, i16 %call)341 ret void342}343; CHECK-LABEL: define void @caller7(344; CHECK: %call = call i16 @caller7.external(i16 1)345; CHECK-NEXT: ret void346 347define float @caller8(float %y) {348; Check that we can constant-prop through fneg instructions349;350; CHECK-LABEL: @caller8(351; CHECK-NOT: call352; CHECK: ret353 %x = call float @callee8(float -42.0, float %y)354 ret float %x355}356 357define float @callee8(float %x, float %y) {358 %neg = fneg float %x359 %icmp = fcmp ugt float %neg, 42.0360 br i1 %icmp, label %bb.true, label %bb.false361 362bb.true:363 ; This block musn't be counted in the inline cost.364 %y1 = fadd float %y, 1.0365 %y2 = fadd float %y1, 1.0366 %y3 = fadd float %y2, 1.0367 %y4 = fadd float %y3, 1.0368 %y5 = fadd float %y4, 1.0369 %y6 = fadd float %y5, 1.0370 %y7 = fadd float %y6, 1.0371 %y8 = fadd float %y7, 1.0372 ret float %y8373 374bb.false:375 ret float %x376}377