| 1 | // This is a regression test for ThinLTO indirect-call-promotion when candidate |
| 2 | // callees need to be imported from another IR module. In the C++ test case, |
| 3 | // `main` calls `global_func` which is defined in another module. `global_func` |
| 4 | // has two indirect callees, one has external linkage and one has local linkage. |
| 5 | // All three functions should be imported into the IR module of main. |
| 6 | |
| 7 | // What the test does: |
| 8 | // - Generate raw profiles from executables and convert it to indexed profiles. |
| 9 | // During the conversion, a profiled callee address in raw profiles will be |
| 10 | // converted to function hash in indexed profiles. |
| 11 | // - Run IRPGO profile use and ThinTLO prelink pipeline and get LLVM bitcodes |
| 12 | // for both cpp files in the C++ test case. |
| 13 | // - Generate ThinLTO summary file with LLVM bitcodes, and run `function-import` pass. |
| 14 | // - Run `pgo-icall-prom` pass for the IR module which needs to import callees. |
| 15 | |
| 16 | // REQUIRES: windows || linux || darwin |
| 17 | |
| 18 | // The test failed on ppc when building the instrumented binary. |
| 19 | // ld.lld: error: /lib/../lib64/Scrt1.o: ABI version 1 is not supported |
| 20 | // UNSUPPORTED: ppc |
| 21 | |
| 22 | // This test and IR test llvm/test/Transforms/PGOProfile/thinlto_indirect_call_promotion.ll |
| 23 | // are complementary to each other; a compiler-rt test has better test coverage |
| 24 | // on different platforms, and the IR test is less restrictive in terms of |
| 25 | // running environment and could be executed more widely. |
| 26 | |
| 27 | // Use lld as linker for more robust test. We need to REQUIRE LLVMgold.so for |
| 28 | // LTO if default linker is GNU ld or gold anyway. |
| 29 | // REQUIRES: lld-available |
| 30 | |
| 31 | // RUN: rm -rf %t && split-file %s %t && cd %t |
| 32 | |
| 33 | // Do setup work for all below tests. |
| 34 | // Generate raw profiles from real programs and convert it into indexed profiles. |
| 35 | // Use clangxx_pgogen for IR level instrumentation for C++. |
| 36 | // RUN: %clangxx_pgogen -fuse-ld=lld -O2 lib.cpp main.cpp -o main |
| 37 | // RUN: env LLVM_PROFILE_FILE=main.profraw %run ./main |
| 38 | // RUN: llvm-profdata merge main.profraw -o main.profdata |
| 39 | |
| 40 | // Use profile on lib and get bitcode. Explicitly skip ICP pass to test ICP happens as |
| 41 | // expected in the IR module that imports functions from lib. |
| 42 | // RUN: %clang -mllvm -disable-icp -fprofile-use=main.profdata -flto=thin -O2 -c lib.cpp -o lib.bc |
| 43 | |
| 44 | // Use profile on main and get bitcode. |
| 45 | // RUN: %clang -fprofile-use=main.profdata -flto=thin -O2 -c main.cpp -o main.bc |
| 46 | |
| 47 | // Run llvm-lto to get summary file. |
| 48 | // RUN: llvm-lto -thinlto -o summary main.bc lib.bc |
| 49 | |
| 50 | // Test the imports of functions. Default import thresholds would work but do |
| 51 | // explicit override to be more futureproof. Note all functions have one basic |
| 52 | // block with a function-entry-count of one, so they are actually hot functions |
| 53 | // per default profile summary hotness cutoff. |
| 54 | // RUN: opt -passes=function-import -import-instr-limit=100 -import-cold-multiplier=1 -summary-file summary.thinlto.bc main.bc -o main.import.bc -print-imports 2>&1 | FileCheck %s --check-prefix=IMPORTS |
| 55 | |
| 56 | // Test that both candidates are ICP'ed and there is no `!VP` in the IR. |
| 57 | // RUN: opt main.import.bc -icp-lto -passes=pgo-icall-prom -S -pass-remarks=pgo-icall-prom 2>&1 | FileCheck %s --check-prefixes=ICP-IR,ICP-REMARK --implicit-check-not="!VP" |
| 58 | |
| 59 | // IMPORTS-DAG: main.cpp: Import {{.*}}callee1{{.*}} |
| 60 | // IMPORTS-DAG: main.cpp: Import {{.*}}callee0{{.*}}llvm.[[#]] |
| 61 | // IMPORTS-DAG: main.cpp: Import {{.*}}global_func{{.*}} |
| 62 | |
| 63 | // PGOName-DAG: define {{.*}}callee1{{.*}} !prof ![[#]] { |
| 64 | // PGOName-DAG: define internal {{.*}}callee0{{.*}} !prof ![[#]] !PGOFuncName ![[#MD:]] { |
| 65 | // PGOName-DAG: ![[#MD]] = !{!"{{.*}}lib.cpp;{{.*}}callee0{{.*}}"} |
| 66 | |
| 67 | // ICP-REMARK: Promote indirect call to {{.*}}callee0{{.*}}llvm.[[#]] with count 1 out of 1 |
| 68 | // ICP-REMARK: Promote indirect call to {{.*}}callee1{{.*}} with count 1 out of 1 |
| 69 | |
| 70 | // ICP-IR: br i1 %[[#]], label %if.true.direct_targ, label %if.false.orig_indirect, !prof ![[#BRANCH_WEIGHT1:]] |
| 71 | // ICP-IR: br i1 %[[#]], label %if.true.direct_targ1, label %if.false.orig_indirect2, !prof ![[#BRANCH_WEIGHT1]] |
| 72 | // ICP-IR: ![[#BRANCH_WEIGHT1]] = !{!"branch_weights", i32 1, i32 0} |
| 73 | |
| 74 | //--- lib.h |
| 75 | void global_func(); |
| 76 | |
| 77 | //--- lib.cpp |
| 78 | #include "lib.h" |
| 79 | static void callee0() {} |
| 80 | void callee1() {} |
| 81 | typedef void (*FPT)(); |
| 82 | FPT calleeAddrs[] = {callee0, callee1}; |
| 83 | // `global_func`` might call one of two indirect callees. callee0 has internal |
| 84 | // linkage and callee1 has external linkage. |
| 85 | void global_func() { |
| 86 | FPT fp = calleeAddrs[0]; |
| 87 | fp(); |
| 88 | fp = calleeAddrs[1]; |
| 89 | fp(); |
| 90 | } |
| 91 | |
| 92 | //--- main.cpp |
| 93 | #include "lib.h" |
| 94 | int main() { global_func(); } |
| 95 | |