Fred / Kettch
2010-Mar-16 07:30 UTC
[LLVMdev] LLVM-GCC generating too much code from inline assembly
Hi, I recently switched to LLVM-GCC 4.2 on OS X, to go around a bug caused by gcc with optimized code. Unfortunately, I ran into another weird problem on LLVM-GCC. In my code, there's a file with a bunch of inline assembly blocks, that worked fine with GCC 4.2. Now, when compiling with LLVM-GCC 4.2, weird things happen. Here's an example: (the blocks are larger than that, but a single line will do to show you what's happening) __asm { mov ebx, iObject } When compiling in release, everything goes fine, and the produced assembly looks like this : ## InlineAsm Start .file "... my filename" ## InlineAsm End ## InlineAsm Start .line 1454 ## InlineAsm End ## InlineAsm Start movl -20(%ebp), $ebx ## InlineAsm End The problem occurs in debug, and the code looks like this : ## InlineAsm Start .file "... my filename" ## InlineAsm End ## InlineAsm Start .line 1454 ## InlineAsm End ## InlineAsm Start movl -20(%ebp), $ebx ## InlineAsm End movl %ebx, %eax movl %eax, -52(%ebp) Only the last two lines are added, but since I'm using the eax register for other things in the block, the code keeps writing over it, and end up crashing when I try to really use whatever I did put in eax.>From the looks of it, those lines must be here for a reason, but I'mwondering how to avoid generating them, as it's causing some major problems in my code (at least in this specific file). Is there some way to disable that ? Thanks, Fred -------------- next part -------------- An HTML attachment was scrubbed... URL: <lists.llvm.org/pipermail/llvm-dev/attachments/20100316/3e525a54/attachment.html>
Jim Grosbach
2010-Mar-16 15:35 UTC
[LLVMdev] LLVM-GCC generating too much code from inline assembly
You may find it helpful to reference ibiblio.org/gferg/ldp/GCC-Inline-Assembly-HOWTO.html. In particular, the information regarding clobbers and constraints. Generally speaking, it's best not to use inline assembly at all. What are you trying to do that you find it necessary? On Mar 16, 2010, at 12:30 AM, Fred / Kettch wrote:> Hi, > > I recently switched to LLVM-GCC 4.2 on OS X, to go around a bug caused by gcc with optimized code. Unfortunately, I ran into another weird problem on LLVM-GCC. In my code, there's a file with a bunch of inline assembly blocks, that worked fine with GCC 4.2. Now, when compiling with LLVM-GCC 4.2, weird things happen. > > Here's an example: (the blocks are larger than that, but a single line will do to show you what's happening) > __asm { > mov ebx, iObject > } > > When compiling in release, everything goes fine, and the produced assembly looks like this : > ## InlineAsm Start > .file "... my filename" > ## InlineAsm End > ## InlineAsm Start > .line 1454 > ## InlineAsm End > ## InlineAsm Start > movl -20(%ebp), $ebx > ## InlineAsm End > > The problem occurs in debug, and the code looks like this : > ## InlineAsm Start > .file "... my filename" > ## InlineAsm End > ## InlineAsm Start > .line 1454 > ## InlineAsm End > ## InlineAsm Start > movl -20(%ebp), $ebx > ## InlineAsm End > movl %ebx, %eax > movl %eax, -52(%ebp) > > Only the last two lines are added, but since I'm using the eax register for other things in the block, the code keeps writing over it, and end up crashing when I try to really use whatever I did put in eax. > >From the looks of it, those lines must be here for a reason, but I'm wondering how to avoid generating them, as it's causing some major problems in my code (at least in this specific file). Is there some way to disable that ? > > Thanks, > Fred > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu llvm.cs.uiuc.edu > lists.cs.uiuc.edu/mailman/listinfo/llvmdev
Fred / Kettch
2010-Mar-16 15:55 UTC
[LLVMdev] LLVM-GCC generating too much code from inline assembly
Yeah, I used that document as a reference to rewrite all the assembly blocks that were failing, but since the code is pretty central to my software, I'd like to be able to revert to the old code (that GCC 4.2 was producing right) if I need to check for differences between the two, or just if other problems arise... In my case, I absolutely need inline asm, since I'm writing functions that do what objc_msgSend does for Obj-C, but for C++. Methods are binded at the start and then I can have C++ setters/getters, and so on... It's a pretty hard thing to do without asm ;) But right now I managed to get it done with small asm blocks, using constraints and clobbers on each of them. And now that it's all working again, I'm finding performance took a big hit since I went to using LLVM-GCC in debug. Are there any other surprises like the one I found on asm blocks that are disseminated all around, making my app be 3x times slower in debug ? Fred On Mar 16, 2010, at 4:35 PM, Jim Grosbach wrote:> You may find it helpful to reference ibiblio.org/gferg > ldp/GCC-Inline-Assembly-HOWTO.html. In particular, the information > regarding clobbers and constraints. > > Generally speaking, it's best not to use inline assembly at all. > What are you trying to do that you find it necessary? > > > On Mar 16, 2010, at 12:30 AM, Fred / Kettch wrote: > >> Hi, >> >> I recently switched to LLVM-GCC 4.2 on OS X, to go around a bug >> caused by gcc with optimized code. Unfortunately, I ran into >> another weird problem on LLVM-GCC. In my code, there's a file with >> a bunch of inline assembly blocks, that worked fine with GCC 4.2. >> Now, when compiling with LLVM-GCC 4.2, weird things happen. >> >> Here's an example: (the blocks are larger than that, but a single >> line will do to show you what's happening) >> __asm { >> mov ebx, iObject >> } >> >> When compiling in release, everything goes fine, and the produced >> assembly looks like this : >> ## InlineAsm Start >> .file "... my filename" >> ## InlineAsm End >> ## InlineAsm Start >> .line 1454 >> ## InlineAsm End >> ## InlineAsm Start >> movl -20(%ebp), $ebx >> ## InlineAsm End >> >> The problem occurs in debug, and the code looks like this : >> ## InlineAsm Start >> .file "... my filename" >> ## InlineAsm End >> ## InlineAsm Start >> .line 1454 >> ## InlineAsm End >> ## InlineAsm Start >> movl -20(%ebp), $ebx >> ## InlineAsm End >> movl %ebx, %eax >> movl %eax, -52(%ebp) >> >> Only the last two lines are added, but since I'm using the eax >> register for other things in the block, the code keeps writing >> over it, and end up crashing when I try to really use whatever I >> did put in eax. >>> From the looks of it, those lines must be here for a reason, but >>> I'm wondering how to avoid generating them, as it's causing some >>> major problems in my code (at least in this specific file). Is >>> there some way to disable that ? >> >> Thanks, >> Fred >> >> >> _______________________________________________ >> LLVM Developers mailing list >> LLVMdev at cs.uiuc.edu llvm.cs.uiuc.edu >> lists.cs.uiuc.edu/mailman/listinfo/llvmdev >
Dale Johannesen
2010-Mar-23 22:11 UTC
[LLVMdev] LLVM-GCC generating too much code from inline assembly
On Mar 16, 2010, at 12:30 AMPDT, Fred / Kettch wrote:> > I recently switched to LLVM-GCC 4.2 on OS X, to go around a bug caused by gcc with optimized code. Unfortunately, I ran into another weird problem on LLVM-GCC. In my code, there's a file with a bunch of inline assembly blocks, that worked fine with GCC 4.2. Now, when compiling with LLVM-GCC 4.2, weird things happen. > > Here's an example: (the blocks are larger than that, but a single line will do to show you what's happening) > __asm { > mov ebx, iObject > } > > The problem occurs in debug, and the code looks like this : > ## InlineAsm Start > .file "... my filename" > ## InlineAsm End > ## InlineAsm Start > .line 1454 > ## InlineAsm End > ## InlineAsm Start > movl -20(%ebp), $ebx > ## InlineAsm End > movl %ebx, %eax > movl %eax, -52(%ebp)This is something I fixed recently; I think it produces functionally correct code at -O0 now in TOT. You will still see interpolated copies, and stack loads and stores, but it should no longer clobber registers that are explicitly used for something else. You need to use -O to get good performance though. -------------- next part -------------- An HTML attachment was scrubbed... URL: <lists.llvm.org/pipermail/llvm-dev/attachments/20100323/b0352560/attachment.html>
Maybe Matching Threads
- [LLVMdev] LLVM-GCC generating too much code from inline assembly
- [LLVMdev] LLVM-GCC generating too much code from inline assembly
- Help required regarding IPRA and Local Function optimization
- [LLVMdev] Build issues on Solaris
- [LLVMdev] [RFC] [X86] Mov to push transformation in x86-32 call sequences