308 lines · plain
1; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' < %s | FileCheck %s2 3target datalayout =4"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"5target triple = "x86_64-unknown-linux-gnu"6 7@x = external global [1 x [2 x <4 x float>]]8 9; Can we sink single addressing mode computation to use?10define void @test1(i1 %cond, ptr %base) {11; CHECK-LABEL: @test112; CHECK: getelementptr inbounds i8, {{.+}} 4013entry:14 %addr = getelementptr inbounds i64, ptr %base, i64 515 br i1 %cond, label %if.then, label %fallthrough16 17if.then:18 %v = load i32, ptr %addr, align 419 br label %fallthrough20 21fallthrough:22 ret void23}24 25declare void @foo(i32)26 27; Make sure sinking two copies of addressing mode into different blocks works28define void @test2(i1 %cond, ptr %base) {29; CHECK-LABEL: @test230entry:31 %addr = getelementptr inbounds i64, ptr %base, i64 532 br i1 %cond, label %if.then, label %fallthrough33 34if.then:35; CHECK-LABEL: if.then:36; CHECK: getelementptr inbounds i8, {{.+}} 4037 %v1 = load i32, ptr %addr, align 438 call void @foo(i32 %v1)39 %cmp = icmp eq i32 %v1, 040 br i1 %cmp, label %next, label %fallthrough41 42next:43; CHECK-LABEL: next:44; CHECK: getelementptr inbounds i8, {{.+}} 4045 %v2 = load i32, ptr %addr, align 446 call void @foo(i32 %v2)47 br label %fallthrough48 49fallthrough:50 ret void51}52 53; If we have two loads in the same block, only need one copy of addressing mode54; - instruction selection will duplicate if needed55define void @test3(i1 %cond, ptr %base) {56; CHECK-LABEL: @test357entry:58 %addr = getelementptr inbounds i64, ptr %base, i64 559 br i1 %cond, label %if.then, label %fallthrough60 61if.then:62; CHECK-LABEL: if.then:63; CHECK: getelementptr inbounds i8, {{.+}} 4064 %v1 = load i32, ptr %addr, align 465 call void @foo(i32 %v1)66; CHECK-NOT: getelementptr inbounds i8, {{.+}} 4067 %v2 = load i32, ptr %addr, align 468 call void @foo(i32 %v2)69 br label %fallthrough70 71fallthrough:72 ret void73}74 75; Can we still sink addressing mode if there's a cold use of the76; address itself? 77define void @test4(i1 %cond, ptr %base) {78; CHECK-LABEL: @test479entry:80 %addr = getelementptr inbounds i64, ptr %base, i64 581 br i1 %cond, label %if.then, label %fallthrough82 83if.then:84; CHECK-LABEL: if.then:85; CHECK: getelementptr inbounds i8, {{.+}} 4086 %v1 = load i32, ptr %addr, align 487 call void @foo(i32 %v1)88 %cmp = icmp eq i32 %v1, 089 br i1 %cmp, label %rare.1, label %fallthrough90 91fallthrough:92 ret void93 94rare.1:95; CHECK-LABEL: rare.1:96; CHECK: getelementptr inbounds i8, {{.+}} 4097 call void @slowpath(i32 %v1, ptr %addr) cold98 br label %fallthrough99}100 101; Negative test - don't want to duplicate addressing into hot path102define void @test5(i1 %cond, ptr %base) {103; CHECK-LABEL: @test5104entry:105; CHECK: %addr = getelementptr inbounds106 %addr = getelementptr inbounds i64, ptr %base, i64 5107 br i1 %cond, label %if.then, label %fallthrough108 109if.then:110; CHECK-LABEL: if.then:111; CHECK-NOT: getelementptr inbounds i8, {{.+}} 40112 %v1 = load i32, ptr %addr, align 4113 call void @foo(i32 %v1)114 %cmp = icmp eq i32 %v1, 0115 br i1 %cmp, label %rare.1, label %fallthrough116 117fallthrough:118 ret void119 120rare.1:121 call void @slowpath(i32 %v1, ptr %addr) ;; NOT COLD122 br label %fallthrough123}124 125; Negative test - opt for size126define void @test6(i1 %cond, ptr %base) minsize {127; CHECK-LABEL: @test6128entry:129; CHECK: %addr = getelementptr130 %addr = getelementptr inbounds i64, ptr %base, i64 5131 br i1 %cond, label %if.then, label %fallthrough132 133if.then:134; CHECK-LABEL: if.then:135; CHECK-NOT: getelementptr inbounds i8, {{.+}} 40136 %v1 = load i32, ptr %addr, align 4137 call void @foo(i32 %v1)138 %cmp = icmp eq i32 %v1, 0139 br i1 %cmp, label %rare.1, label %fallthrough140 141fallthrough:142 ret void143 144rare.1:145 call void @slowpath(i32 %v1, ptr %addr) cold146 br label %fallthrough147}148 149; Negative test - opt for size150define void @test6_pgso(i1 %cond, ptr %base) !prof !14 {151; CHECK-LABEL: @test6152entry:153; CHECK: %addr = getelementptr154 %addr = getelementptr inbounds i64, ptr %base, i64 5155 br i1 %cond, label %if.then, label %fallthrough156 157if.then:158; CHECK-LABEL: if.then:159; CHECK-NOT: getelementptr inbounds i8, {{.+}} 40160 %v1 = load i32, ptr %addr, align 4161 call void @foo(i32 %v1)162 %cmp = icmp eq i32 %v1, 0163 br i1 %cmp, label %rare.1, label %fallthrough164 165fallthrough:166 ret void167 168rare.1:169 call void @slowpath(i32 %v1, ptr %addr) cold170 br label %fallthrough171}172 173; Make sure sinking two copies of addressing mode into different blocks works174; when there are cold paths for each.175define void @test7(i1 %cond, ptr %base) {176; CHECK-LABEL: @test7177entry:178 %addr = getelementptr inbounds i64, ptr %base, i64 5179 br i1 %cond, label %if.then, label %fallthrough180 181if.then:182; CHECK-LABEL: if.then:183; CHECK: getelementptr inbounds i8, {{.+}} 40184 %v1 = load i32, ptr %addr, align 4185 call void @foo(i32 %v1)186 %cmp = icmp eq i32 %v1, 0187 br i1 %cmp, label %rare.1, label %next188 189next:190; CHECK-LABEL: next:191; CHECK: getelementptr inbounds i8, {{.+}} 40192 %v2 = load i32, ptr %addr, align 4193 call void @foo(i32 %v2)194 %cmp2 = icmp eq i32 %v2, 0195 br i1 %cmp2, label %rare.1, label %fallthrough196 197fallthrough:198 ret void199 200rare.1:201; CHECK-LABEL: rare.1:202; CHECK: getelementptr inbounds i8, {{.+}} 40203 call void @slowpath(i32 %v1, ptr %addr) cold204 br label %next205 206rare.2:207; CHECK-LABEL: rare.2:208; CHECK: getelementptr inbounds i8, {{.+}} 40209 call void @slowpath(i32 %v2, ptr %addr) cold210 br label %fallthrough211}212 213declare void @slowpath(i32, ptr)214 215; Make sure we don't end up in an infinite loop after we fail to sink.216; CHECK-LABEL: define void @test8217; CHECK: %ptr = getelementptr i8, ptr %aFOO_load_ptr2int_2void, i32 undef218define void @test8() {219allocas:220 %aFOO_load = load ptr, ptr undef221 %aFOO_load_ptr2int = ptrtoint ptr %aFOO_load to i64222 %aFOO_load_ptr2int_broadcast_init = insertelement <4 x i64> undef, i64 %aFOO_load_ptr2int, i32 0223 %aFOO_load_ptr2int_2void = inttoptr i64 %aFOO_load_ptr2int to ptr224 %ptr = getelementptr i8, ptr %aFOO_load_ptr2int_2void, i32 undef225 br label %load.i145226 227load.i145:228 %valall.i144 = load <4 x float>, ptr %ptr, align 4229 br label %pl_loop.i.i122230 231pl_loop.i.i122:232 br label %pl_loop.i.i122233}234 235; Make sure we can sink address computation even236; if there is a cycle in phi nodes.237define void @test9(i1 %cond, ptr %base) {238; CHECK-LABEL: @test9239entry:240 %addr = getelementptr inbounds i64, ptr %base, i64 5241 br label %header242 243header:244 %iv = phi i32 [0, %entry], [%iv.inc, %backedge]245 %casted.loop = phi ptr [%addr, %entry], [%casted.merged, %backedge]246 br i1 %cond, label %if.then, label %backedge247 248if.then:249 call void @foo(i32 %iv)250 %addr.1 = getelementptr inbounds i64, ptr %base, i64 5251 br label %backedge252 253backedge:254; CHECK-LABEL: backedge:255; CHECK: getelementptr inbounds i8, {{.+}} 40256 %casted.merged = phi ptr [%casted.loop, %header], [%addr.1, %if.then]257 %v = load i32, ptr %casted.merged, align 4258 call void @foo(i32 %v)259 %iv.inc = add i32 %iv, 1260 %cmp = icmp slt i32 %iv.inc, 1000261 br i1 %cmp, label %header, label %exit262 263exit:264 ret void265}266 267; Make sure we can eliminate a select when both arguments perform equivalent268; address computation.269define void @test10(i1 %cond, ptr %base) {270; CHECK-LABEL: @test10271; CHECK: getelementptr inbounds i8, {{.+}} 40272; CHECK-NOT: select273entry:274 %gep1 = getelementptr inbounds i64, ptr %base, i64 5275 %gep2 = getelementptr inbounds i32, ptr %base, i64 10276 %casted.merged = select i1 %cond, ptr %gep1, ptr %gep2277 %v = load i32, ptr %casted.merged, align 4278 call void @foo(i32 %v)279 ret void280}281 282; Found by fuzzer, getSExtValue of > 64 bit constant283define void @i96_mul(ptr %base, i96 %offset) {284BB:285 ;; RHS = 0x7FFFFFFFFFFFFFFFFFFFFFFF286 %B84 = mul i96 %offset, 39614081257132168796771975167287 %G23 = getelementptr i1, ptr %base, i96 %B84288 store i1 false, ptr %G23289 ret void290}291 292!llvm.module.flags = !{!0}293!0 = !{i32 1, !"ProfileSummary", !1}294!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}295!2 = !{!"ProfileFormat", !"InstrProf"}296!3 = !{!"TotalCount", i64 10000}297!4 = !{!"MaxCount", i64 10}298!5 = !{!"MaxInternalCount", i64 1}299!6 = !{!"MaxFunctionCount", i64 1000}300!7 = !{!"NumCounts", i64 3}301!8 = !{!"NumFunctions", i64 3}302!9 = !{!"DetailedSummary", !10}303!10 = !{!11, !12, !13}304!11 = !{i32 10000, i64 100, i32 1}305!12 = !{i32 999000, i64 100, i32 1}306!13 = !{i32 999999, i64 1, i32 2}307!14 = !{!"function_entry_count", i64 0}308