aboutsummaryrefslogtreecommitdiffstats
path: root/lib
diff options
context:
space:
mode:
authorChris Lattner <sabre@nondot.org>2008-04-01 06:05:21 +0000
committerChris Lattner <sabre@nondot.org>2008-04-01 06:05:21 +0000
commit11da4cf25c3de590ecf434316973e9f25eec34c3 (patch)
treed87f943e29eaabf9e204fd2ab8af186183e915e9 /lib
parent799ed1053b6a4cb2e3a6d588bd24025daee7ad1f (diff)
downloadexternal_llvm-11da4cf25c3de590ecf434316973e9f25eec34c3.zip
external_llvm-11da4cf25c3de590ecf434316973e9f25eec34c3.tar.gz
external_llvm-11da4cf25c3de590ecf434316973e9f25eec34c3.tar.bz2
Reimplement MemoryBuffer::getFile with three enhancements:
1) stop using MappedFile. 2) if profitable use the sys::path::MapInFilePages api to read the file. 3) otherwise fallback to read. When sys::path::MapInFilePages is implemented, this provides several benefits: #1: this avoids fragmenting memory for small files. #2: this avoids extraneous stat calls when the file size is known. #3: this only keeps the file descriptor open while reading the file, not for the duration of the lifetime of the memory buffer. This fixes a serious clang FD 'leak' problem. I believe that this will work on a win32 machine, but I don't have one to test on. I'd appreciate it if someone could check. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49031 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r--lib/Support/MemoryBuffer.cpp149
1 files changed, 64 insertions, 85 deletions
diff --git a/lib/Support/MemoryBuffer.cpp b/lib/Support/MemoryBuffer.cpp
index 1850084..4e4dabc 100644
--- a/lib/Support/MemoryBuffer.cpp
+++ b/lib/Support/MemoryBuffer.cpp
@@ -12,13 +12,24 @@
//===----------------------------------------------------------------------===//
#include "llvm/Support/MemoryBuffer.h"
-#include "llvm/System/MappedFile.h"
+#include "llvm/ADT/OwningPtr.h"
+#include "llvm/ADT/SmallString.h"
+#include "llvm/System/Path.h"
#include "llvm/System/Process.h"
#include "llvm/System/Program.h"
#include <cassert>
#include <cstdio>
#include <cstring>
#include <cerrno>
+#include <sys/types.h>
+#include <sys/stat.h>
+#if !defined(_MSC_VER) && !defined(__MINGW32__)
+#include <unistd.h>
+#include <sys/uio.h>
+#include <sys/fcntl.h>
+#else
+#include <io.h>
+#endif
using namespace llvm;
//===----------------------------------------------------------------------===//
@@ -136,108 +147,78 @@ MemoryBuffer *MemoryBuffer::getFileOrSTDIN(const char *FilenameStart,
}
//===----------------------------------------------------------------------===//
-// MemoryBufferMMapFile implementation.
+// MemoryBuffer::getFile implementation.
//===----------------------------------------------------------------------===//
namespace {
+/// MemoryBufferMMapFile - This represents a file that was mapped in with the
+/// sys::Path::MapInFilePages method. When destroyed, it calls the
+/// sys::Path::UnMapFilePages method.
class MemoryBufferMMapFile : public MemoryBuffer {
- sys::MappedFile File;
+ std::string Filename;
public:
- MemoryBufferMMapFile() {}
-
- bool open(const sys::Path &Filename, std::string *ErrStr);
+ MemoryBufferMMapFile(const char *filename, const char *Pages, uint64_t Size)
+ : Filename(filename) {
+ init(Pages, Pages+Size);
+ }
virtual const char *getBufferIdentifier() const {
- return File.path().c_str();
+ return Filename.c_str();
}
- ~MemoryBufferMMapFile();
+ ~MemoryBufferMMapFile() {
+ sys::Path::UnMapFilePages(getBufferStart(), getBufferSize());
+ }
};
}
-bool MemoryBufferMMapFile::open(const sys::Path &Filename,
- std::string *ErrStr) {
- // FIXME: This does an extra stat syscall to figure out the size, but we
- // already know the size!
- bool Failure = File.open(Filename, ErrStr);
- if (Failure) return true;
-
- if (!File.map(ErrStr))
- return true;
-
- size_t Size = File.size();
-
- static unsigned PageSize = sys::Process::GetPageSize();
- assert(((PageSize & (PageSize-1)) == 0) && PageSize &&
- "Page size is not a power of 2!");
+MemoryBuffer *MemoryBuffer::getFile(const char *FilenameStart, unsigned FnSize,
+ std::string *ErrStr, int64_t FileSize) {
+ // Null terminate the filename.
+ SmallString<1000> Filename(FilenameStart, FilenameStart+FnSize);
+ Filename.push_back(0);
- // If this file is not an exact multiple of the system page size (common
- // case), then the OS has zero terminated the buffer for us.
- const char *FileBase = static_cast<const char*>(File.getBase());
- if ((Size & (PageSize-1)) != 0) {
- init(FileBase, FileBase+Size);
- } else {
- // Otherwise, we allocate a new memory buffer and copy the data over
- initCopyOf(FileBase, FileBase+Size);
-
- // No need to keep the file mapped any longer.
- File.unmap();
+ int OpenFlags = 0;
+#ifdef O_BINARY
+ Flags |= O_BINARY; // Open input file in binary mode on win32.
+#endif
+ int FD = ::open(&Filename[0], O_RDONLY|OpenFlags);
+ if (FD == -1) {
+ if (ErrStr) *ErrStr = "could not open file";
+ return 0;
}
- return false;
-}
-
-MemoryBufferMMapFile::~MemoryBufferMMapFile() {
- if (File.isMapped())
- File.unmap();
-}
-
-//===----------------------------------------------------------------------===//
-// MemoryBuffer::getFile implementation.
-//===----------------------------------------------------------------------===//
-
-MemoryBuffer *MemoryBuffer::getFile(const char *FilenameStart, unsigned FnSize,
- std::string *ErrStr, int64_t FileSize){
- // FIXME: it would be nice if PathWithStatus didn't copy the filename into a
- // temporary string. :(
- sys::PathWithStatus P(FilenameStart, FnSize);
-#if 1
- MemoryBufferMMapFile *M = new MemoryBufferMMapFile();
- if (!M->open(P, ErrStr))
- return M;
- delete M;
- return 0;
-#else
- // FIXME: We need an efficient and portable method to open a file and then use
- // 'read' to copy the bits out. The unix implementation is below. This is
- // an important optimization for clients that want to open large numbers of
- // small files (using mmap on everything can easily exhaust address space!).
- // If the user didn't specify a filesize, do a stat to find it.
+ // If we don't know the file size, use fstat to find out. fstat on an open
+ // file descriptor is cheaper than stat on a random path.
if (FileSize == -1) {
- const sys::FileStatus *FS = P.getFileStatus();
- if (FS == 0) return 0; // Error stat'ing file.
-
- FileSize = FS->fileSize;
+ struct stat FileInfo;
+ // TODO: This should use fstat64 when available.
+ if (fstat(FD, &FileInfo) == -1) {
+ if (ErrStr) *ErrStr = "could not get file length";
+ ::close(FD);
+ return 0;
+ }
+ FileSize = FileInfo.st_size;
}
- // If the file is larger than some threshold, use mmap, otherwise use 'read'.
- if (FileSize >= 4096*4) {
- MemoryBufferMMapFile *M = new MemoryBufferMMapFile();
- if (!M->open(P, ErrStr))
- return M;
- delete M;
- return 0;
+
+ // If the file is large, try to use mmap to read it in. We don't use mmap
+ // for small files, because this can severely fragment our address space. Also
+ // don't try to map files that are exactly a multiple of the system page size,
+ // as the file would not have the required null terminator.
+ if (FileSize >= 4096*4 &&
+ (FileSize & (sys::Process::GetPageSize()-1)) != 0) {
+ if (const char *Pages = sys::Path::MapInFilePages(FD, FileSize)) {
+ // Close the file descriptor, now that the whole file is in memory.
+ ::close(FD);
+ return new MemoryBufferMMapFile(&Filename[0], Pages, FileSize);
+ }
}
- MemoryBuffer *SB = getNewUninitMemBuffer(FileSize, FilenameStart);
+ OwningPtr<MemoryBuffer> SB;
+ SB.reset(MemoryBuffer::getNewUninitMemBuffer(FileSize, &Filename[0]));
char *BufPtr = const_cast<char*>(SB->getBufferStart());
- int FD = ::open(FilenameStart, O_RDONLY);
- if (FD == -1) {
- delete SB;
- return 0;
- }
-
unsigned BytesLeft = FileSize;
while (BytesLeft) {
ssize_t NumRead = ::read(FD, BufPtr, BytesLeft);
@@ -249,17 +230,15 @@ MemoryBuffer *MemoryBuffer::getFile(const char *FilenameStart, unsigned FnSize,
} else {
// error reading.
close(FD);
- delete SB;
+ if (ErrStr) *ErrStr = "error reading file data";
return 0;
}
}
close(FD);
- return SB;
-#endif
+ return SB.take();
}
-
//===----------------------------------------------------------------------===//
// MemoryBuffer::getSTDIN implementation.
//===----------------------------------------------------------------------===//