15209 – [3.4/4.0 Regression]Runs out of memory with packed structs (original) (raw)

Description Wolfgang Bangerth 2004-04-29 19:40:12 UTC

Taken from http://gcc.gnu.org/ml/gcc-bugs/2004-04/msg02848.html :

This is somewhat ironic: we run out of memory when using packed structs the wrong way ;-)

I took the files from the package in above mail and tried to compile. Indeed, we get into a memory eating loop with the preprocessed source files I attach (it quickly runs into the 500-600 MB range and then eventually crashes for me). Removing the cause of the warning messages
x.ii:34314: warning: ignoring packed attribute on unpacked non-POD field MD5Hash PACKET_HEADER::hash' x.ii:34315: warning: ignoring packed attribute on unpacked non-POD field MD5Hash PACKET_HEADER::setid' makes the problem to go away, though. Compile line is simply c++ x.ii

With 3.3.4, I get an ICE somewhere in the libstdc++ parts of 3.4, so I can't say from the preprocessed sources (all that I have left on my harddisk right now) whether this is a regression.

I don't have the time to reduce memory hogs at present, so if someone else is interested in picking this up, feel free to.

Wolfgang

Comment 2 Drea Pinski 2004-04-29 20:04:35 UTC

Confirmed, reduced to at least (note this can most likely be reduced down further): typedef unsigned int size_t; typedef unsigned char uint8_t; typedef unsigned short int uint16_t;

typedef unsigned int uint32_t; extension typedef unsigned long long int uint64_t;

typedef uint8_t u8; typedef uint16_t u16; typedef uint32_t u32; typedef uint64_t u64;

struct MAGIC {u8 magic[8];} attribute ((packed)); struct PACKETTYPE {u8 type[16];} attribute ((packed));

typedef u16 leu16; typedef u32 leu32; typedef u64 leu64;

class MD5Hash { public:

MD5Hash(void) {};

void *print(void) const; MD5Hash(const MD5Hash &other); MD5Hash& operator=(const MD5Hash &other);

public: u8 hash[16]; };

struct PACKET_HEADER {

MAGIC magic; leu64 length; MD5Hash hash; MD5Hash setid; PACKETTYPE type; } attribute ((packed));

struct MAINPACKET { PACKET_HEADER header;

leu64 blocksize; leu32 recoverablefilecount; MD5Hash fileid[0];

} attribute ((packed));

struct CriticalPacket { u8 *packetdata; size_t packetlength; };

class MainPacket : public CriticalPacket { const MD5Hash& SetId(void) const;

u64 blocksize; u32 totalfilecount; u32 recoverablefilecount; };

inline const MD5Hash& MainPacket::SetId(void) const { return ((const MAINPACKET*)packetdata)->header.setid; }

Comment 3 Drea Pinski 2004-05-21 19:29:36 UTC

*** Bug 15573 has been marked as a duplicate of this bug. ***

Comment 4 Drea Pinski 2004-05-21 19:44:12 UTC

And here is the most reduced source: struct MD5Hash { MD5Hash(void) {}; void *print(void) const; MD5Hash (const MD5Hash&); }; struct PACKET_HEADER { MD5Hash setid; } attribute ((packed)); const MD5Hash& SetId(unsigned char packetdata) { return ((const PACKET_HEADER)packetdata)->setid; }

Comment 5 Mark Mitchell 2004-05-28 21:03:57 UTC

I cannot reproduce this with the reduced case in Comment #4 using current 3.4.1 sources. Is this still a problem?

Comment 7 Mark Mitchell 2004-05-31 17:58:46 UTC

Working on a fix.

Comment 10 Mark Mitchell 2004-05-31 21:29:20 UTC

Fixed in GCC 3.4.1.