197 lines · plain
1;; This test ensures that the logic which assigns calls to stack nodes2;; correctly handles an inlined callsite with stack ids that partially3;; overlap with a trimmed context. In particular when it also partially4;; overlaps with a longer non-trimmed context that doesn't match all of5;; the inlined callsite stack ids.6 7;; The profile data and call stacks were all manually added, but the code8;; would be structured something like the following (fairly contrived to9;; result in the type of control flow needed to test):10 11;; void A(bool b) {12;; if (b)13;; // cold: stack ids 6, 2, 8 (trimmed ids 10)14;; // not cold: stack ids 6, 7 (trimmed ids 9, 11)15;; new char[10]; // stack id 616;; else17;; // not cold: stack ids 1, 2, 8, 3, 418;; // cold: stack ids 1, 2, 8, 3, 519;; new char[10]; // stack id 120;; }21;;22;; void XZ() {23;; A(false); // stack ids 2, 8 (e.g. X inlined into Z)24;; }25;;26;; void XZN() {27;; // This is the tricky one to get right. We want to ensure it gets28;; // correctly correlated with a stack node for the trimmed 6, 2, 829;; // context shown in A. It should *not* be correlated with the longer30;; // untrimmed 1, 2, 8, 3, 4|5 contexts.31;; A(true); // stack ids 2, 8, 9 (e.g. X inlined into Z inlined into N)32;; }33;;34;; void Y() {35;; A(true); // stack id 736;; }37;;38;; void M() {39;; XZ(); // stack id 340;; }41;;42;; int main() {43;; M(); // stack id 4 (leads to not cold allocation)44;; M(); // stack id 5 (leads to cold allocation)45;; XZN(); // stack id 11 (leads to cold allocation)46;; Y(); // stack id 10 (leads to not cold allocation)47;; }48 49;; -stats requires asserts50; REQUIRES: asserts51 52; RUN: opt -passes=memprof-context-disambiguation -supports-hot-cold-new \53; RUN: -memprof-verify-ccg -memprof-verify-nodes \54; RUN: -stats -pass-remarks=memprof-context-disambiguation \55; RUN: %s -S 2>&1 | FileCheck %s --check-prefix=IR \56; RUN: --check-prefix=STATS --check-prefix=REMARKS57 58; REMARKS: created clone _Z1Ab.memprof.159; REMARKS: created clone _Z2XZv.memprof.160; REMARKS: created clone _Z1Mv.memprof.161;; Make sure the inlined context in _Z3XZNv, which partially overlaps62;; trimmed cold context, and also partially overlaps completely63;; unrelated contexts, correctly calls a cloned version of Z1Ab,64;; which will call the cold annotated allocation.65; REMARKS: call in clone _Z3XZNv assigned to call function clone _Z1Ab.memprof.166; REMARKS: call in clone main assigned to call function clone _Z1Mv.memprof.167; REMARKS: call in clone _Z1Mv.memprof.1 assigned to call function clone _Z2XZv.memprof.168; REMARKS: call in clone _Z2XZv.memprof.1 assigned to call function clone _Z1Ab69; REMARKS: call in clone main assigned to call function clone _Z1Mv70; REMARKS: call in clone _Z1Mv assigned to call function clone _Z2XZv71; REMARKS: call in clone _Z2XZv assigned to call function clone _Z1Ab.memprof.172; REMARKS: call in clone _Z1Ab.memprof.1 marked with memprof allocation attribute cold73; REMARKS: call in clone _Z1Yv assigned to call function clone _Z1Ab74; REMARKS: call in clone _Z1Ab marked with memprof allocation attribute notcold75; REMARKS: call in clone _Z1Ab marked with memprof allocation attribute cold76; REMARKS: call in clone _Z1Ab.memprof.1 marked with memprof allocation attribute notcold77 78 79target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"80target triple = "x86_64-unknown-linux-gnu"81 82define dso_local void @_Z1Ab(i1 noundef zeroext %b) {83entry:84 br i1 %b, label %if.then, label %if.else85 86if.then:87 %call = call noalias noundef nonnull ptr @_Znam(i64 noundef 10) #7, !memprof !5, !callsite !1188 br label %if.end89 90if.else:91 %call2 = call noalias noundef nonnull ptr @_Znam(i64 noundef 10) #7, !memprof !0, !callsite !1092 br label %if.end93 94if.end:95 ret void96}97 98; Function Attrs: nobuiltin99declare ptr @_Znam(i64) #0100 101define dso_local void @_Z2XZv() local_unnamed_addr #0 {102entry:103 tail call void @_Z1Ab(i1 noundef zeroext false), !callsite !12104 ret void105}106 107define dso_local void @_Z1Mv() local_unnamed_addr #0 {108entry:109 tail call void @_Z2XZv(), !callsite !19110 ret void111}112 113define dso_local void @_Z3XZNv() local_unnamed_addr {114entry:115 tail call void @_Z1Ab(i1 noundef zeroext true), !callsite !15116 ret void117}118 119define dso_local void @_Z1Yv() local_unnamed_addr {120entry:121 tail call void @_Z1Ab(i1 noundef zeroext true), !callsite !17122 ret void123}124 125define dso_local noundef i32 @main() local_unnamed_addr {126entry:127 tail call void @_Z1Mv(), !callsite !13 ;; Not cold context128 tail call void @_Z1Mv(), !callsite !14 ;; Cold context129 tail call void @_Z3XZNv(), !callsite !16 ;; Cold context130 tail call void @_Z1Yv(), !callsite !18 ;; Not cold context131 ret i32 0132}133 134attributes #0 = { nobuiltin }135attributes #7 = { builtin }136 137!0 = !{!1, !3}138;; Not cold context via first call to _Z1Mv in main139!1 = !{!2, !"notcold"}140!2 = !{i64 1, i64 2, i64 8, i64 3, i64 4}141;; Cold context via second call to _Z1Mv in main142!3 = !{!4, !"cold"}143!4 = !{i64 1, i64 2, i64 8, i64 3, i64 5}144!5 = !{!6, !8}145;; Cold (trimmed) context via call to _Z3XZNv in main146!6 = !{!7, !"cold"}147!7 = !{i64 6, i64 2, i64 8}148;; Not cold (trimmed) context via call to _Z1Yv in main149!8 = !{!9, !"notcold"}150!9 = !{i64 6, i64 7}151!10 = !{i64 1}152!11 = !{i64 6}153!12 = !{i64 2, i64 8}154!13 = !{i64 4}155!14 = !{i64 5}156;; Inlined context in _Z3XZNv, which includes part of trimmed cold context157!15 = !{i64 2, i64 8, i64 9}158!16 = !{i64 11}159!17 = !{i64 7}160!18 = !{i64 10}161!19 = !{i64 3}162 163; IR: define {{.*}} @_Z1Ab(i1 noundef zeroext %b)164; IR: call {{.*}} @_Znam(i64 noundef 10) #[[NOTCOLD:[0-9]+]]165; IR: call {{.*}} @_Znam(i64 noundef 10) #[[COLD:[0-9]+]]166; IR: define {{.*}} @_Z2XZv()167; IR: call {{.*}} @_Z1Ab.memprof.1(i1 noundef zeroext false)168; IR: define {{.*}} @_Z1Mv()169; IR: call {{.*}} @_Z2XZv()170;; Make sure the inlined context in _Z3XZNv, which partially overlaps171;; trimmed cold context, and also partially overlaps completely172;; unrelated contexts, correctly calls the cloned version of Z1Ab173;; that will call the cold annotated allocation.174; IR: define {{.*}} @_Z3XZNv()175; IR: call {{.*}} @_Z1Ab.memprof.1(i1 noundef zeroext true)176; IR: define {{.*}} @_Z1Yv()177; IR: call {{.*}} @_Z1Ab(i1 noundef zeroext true)178; IR: define {{.*}} @main()179; IR: call {{.*}} @_Z1Mv()180; IR: call {{.*}} @_Z1Mv.memprof.1()181; IR: call {{.*}} @_Z3XZNv()182; IR: call {{.*}} @_Z1Yv()183; IR: define {{.*}} @_Z1Ab.memprof.1(i1 noundef zeroext %b)184; IR: call {{.*}} @_Znam(i64 noundef 10) #[[COLD]]185; IR: call {{.*}} @_Znam(i64 noundef 10) #[[NOTCOLD]]186; IR: define {{.*}} @_Z2XZv.memprof.1()187; IR: call {{.*}} @_Z1Ab(i1 noundef zeroext false)188; IR: define {{.*}} @_Z1Mv.memprof.1()189; IR: call {{.*}} @_Z2XZv.memprof.1()190 191; IR: attributes #[[NOTCOLD]] = { builtin "memprof"="notcold" }192; IR: attributes #[[COLD]] = { builtin "memprof"="cold" }193 194; STATS: 2 memprof-context-disambiguation - Number of cold static allocations (possibly cloned)195; STATS: 2 memprof-context-disambiguation - Number of not cold static allocations (possibly cloned)196; STATS: 3 memprof-context-disambiguation - Number of function clones created during whole program analysis197