| 1 |
//===-- PerfTests.cpp -----------------------------------------------------===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
|
| 9 |
#ifdef __x86_64__ |
| 10 |
|
| 11 |
#include "Perf.h" |
| 12 |
|
| 13 |
#include "llvm/Support/Error.h" |
| 14 |
|
| 15 |
#include "gtest/gtest.h" |
| 16 |
|
| 17 |
#include <chrono> |
| 18 |
#include <cstdint> |
| 19 |
|
| 20 |
using namespace lldb_private; |
| 21 |
using namespace process_linux; |
| 22 |
using namespace llvm; |
| 23 |
|
| 24 |
/// Helper function to read current TSC value. |
| 25 |
/// |
| 26 |
/// This code is based on llvm/xray. |
| 27 |
static Expected<uint64_t> readTsc() { |
| 28 |
|
| 29 |
unsigned int eax, ebx, ecx, edx; |
| 30 |
|
| 31 |
// We check whether rdtscp support is enabled. According to the x86_64 manual, |
| 32 |
// level should be set at 0x80000001, and we should have a look at bit 27 in |
| 33 |
// EDX. That's 0x8000000 (or 1u << 27). |
| 34 |
__asm__ __volatile__("cpuid" |
| 35 |
: "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx) |
| 36 |
: "0"(0x80000001)); |
| 37 |
if (!(edx & (1u << 27))) { |
| 38 |
return createStringError(inconvertibleErrorCode(), |
| 39 |
"Missing rdtscp support."); |
| 40 |
} |
| 41 |
|
| 42 |
unsigned cpu; |
| 43 |
unsigned long rax, rdx; |
| 44 |
|
| 45 |
__asm__ __volatile__("rdtscp\n" : "=a"(rax), "=d"(rdx), "=c"(cpu)::); |
| 46 |
|
| 47 |
return (rdx << 32) + rax; |
| 48 |
} |
| 49 |
|
| 50 |
// Test TSC to walltime conversion based on perf conversion values. |
| 51 |
TEST(Perf, TscConversion) { |
| 52 |
// This test works by first reading the TSC value directly before |
| 53 |
// and after sleeping, then converting these values to nanoseconds, and |
| 54 |
// finally ensuring the difference is approximately equal to the sleep time. |
| 55 |
// |
| 56 |
// There will be slight overhead associated with the sleep call, so it isn't |
| 57 |
// reasonable to expect the difference to be exactly equal to the sleep time. |
| 58 |
|
| 59 |
const int SLEEP_SECS = 1; |
| 60 |
std::chrono::nanoseconds SLEEP_NANOS{std::chrono::seconds(SLEEP_SECS)}; |
| 61 |
|
| 62 |
Expected<LinuxPerfZeroTscConversion> params = |
| 63 |
LoadPerfTscConversionParameters(); |
| 64 |
|
| 65 |
// Skip the test if the conversion parameters aren't available. |
| 66 |
if (!params) |
| 67 |
GTEST_SKIP() << toString(params.takeError()); |
| 68 |
|
| 69 |
Expected<uint64_t> tsc_before_sleep = readTsc(); |
| 70 |
sleep(SLEEP_SECS); |
| 71 |
Expected<uint64_t> tsc_after_sleep = readTsc(); |
| 72 |
|
| 73 |
// Skip the test if we are unable to read the TSC value. |
| 74 |
if (!tsc_before_sleep) |
| 75 |
GTEST_SKIP() << toString(tsc_before_sleep.takeError()); |
| 76 |
if (!tsc_after_sleep) |
| 77 |
GTEST_SKIP() << toString(tsc_after_sleep.takeError()); |
| 78 |
|
| 79 |
uint64_t converted_tsc_diff = |
| 80 |
params->ToNanos(*tsc_after_sleep) - params->ToNanos(*tsc_before_sleep); |
| 81 |
|
| 82 |
std::chrono::microseconds acceptable_overhead(500); |
| 83 |
|
| 84 |
ASSERT_GE(converted_tsc_diff, static_cast<uint64_t>(SLEEP_NANOS.count())); |
| 85 |
ASSERT_LT(converted_tsc_diff, |
| 86 |
static_cast<uint64_t>((SLEEP_NANOS + acceptable_overhead).count())); |
| 87 |
} |
| 88 |
|
| 89 |
#endif // __x86_64__ |