[X87 FPStack] Issues with inline asm

[edwintorok]

Bugzilla Link	2270
Resolution	FIXED
Resolved on	Jun 28, 2011 18:13
Version	unspecified
OS	Linux
Blocks	#1234
Attachments	full testcase
CC	@asl,@efriedma-quic,@pwo,@stoklund

Extended Description

While building glibc this error occurs 3 times:

cc1: /home/edwin/llvm/llvm-svn/trunk/lib/Target/X86/X86FloatingPoint.cpp:858: void::FPS::handleTwoArgFP(llvm::ilist_iteratorllvm::MachineInstr&): Assertion `UpdatedSlot < StackTop && Dest < 7' failed.
../sysdeps/ieee754/ldbl-96/e_asinl.c:161: internal compiler error: Aborted

To reproduce:

$ llc x.bc -f
llc: /home/edwin/llvm/llvm-svn/trunk/lib/Target/X86/X86FloatingPoint.cpp:858: void::FPS::handleTwoArgFP(llvm::ilist_iteratorllvm::MachineInstr&): Assertion `UpdatedSlot < StackTop && Dest < 7' failed.
llc[0x852b506]
/lib/i686/cmov/libc.so.6(abort+0x101)[0x4ecdc871]
/lib/i686/cmov/libc.so.6(__assert_fail+0xee)[0x4ecd40ee]
llc[0x81012cc]
Aborted

$ llc bugpoint-reduced-simplified.bc -f
llc: /home/edwin/llvm/llvm-svn/trunk/lib/Target/X86/X86FloatingPoint.cpp:92: unsigned int::FPS::getStackEntry(unsigned int) const: Assertion `STi < StackTop && "Access past stack top!"' failed.
llc[0x852b506]
/lib/i686/cmov/libc.so.6(abort+0x101)[0x4ecdc871]
/lib/i686/cmov/libc.so.6(__assert_fail+0xee)[0x4ecd40ee]
llc[0x80ff2f4]
Aborted

; ModuleID = 'bugpoint-reduced-simplified.bc'
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-linux-gnu"

define x86_fp80 @​__ieee754_asinl(x86_fp80 %x) nounwind {
entry:
%asmtmp = call x86_fp80 asm sideeffect "fsqrt", "={st},0,{dirflag},{fpsr},~{flags}"( x86_fp80 0xK00000000000000000000 ) nounwind ; <x86_fp80> [#uses=0]
%tmp243 = sub x86_fp80 0xK80000000000000000000, 0xK00000000000000000000 ; <x86_fp80> [#uses=1]
ret x86_fp80 %tmp243
}

[asl]

Edwin,

the FP-related PRs you've just filled looks pretty similar to ones, I filled couple of days ago. Could you please verify (I'm on vacations now)

[edwintorok]

Edwin,

the FP-related PRs you've just filled looks pretty similar to ones, I filled
couple of days ago.

Indeed. I should have done a search before adding new bugs. Thanks for pointing them out Anton.

Could you please verify (I'm on vacations now)

#2641 is indeed a duplicate. I am unsure about 2268 so I just marked it as dependency.

I couldn't find any bug that is similar to the current bug.

[llvmbot]

The reduced testcase now fails with:

llc: llvm/lib/Target/X86/X86FloatingPoint.cpp:1027: void::FPS::handleSpecialFP(llvm::ilist_iteratorllvm::MachineInstr&): Assertion `MO1.isKill()' failed.

[pwo]

I can not reproduce this with r72485. Fixed ?

[edwintorok]

I can not reproduce this with r72485. Fixed ?

The full testcase doesn't fail, however the reduced testcase still fails with the message Duncan posted in comment #​3.

I don't know much about inline asm syntax, is the asm in the reduced testcase invalid?

[edwintorok]

the reduced testcase that still fails
Release/bin/llc /tmp/x.bc
llc: /home/edwin/llvm-svn/llvm/lib/Target/X86/X86FloatingPoint.cpp:1027: void::FPS::handleSpecialFP(llvm::ilist_iteratorllvm::MachineInstr&): Assertion `MO1.isKill()' failed.
0 llc 0x0000000000d5c29f
1 llc 0x0000000000d5c699
2 libpthread.so.0 0x00007fe9006227b0
3 libc.so.6 0x00007fe8ff942065 gsignal + 53
4 libc.so.6 0x00007fe8ff945153 abort + 387
5 libc.so.6 0x00007fe8ff93b159 __assert_fail + 233
6 llc 0x00000000008068de
7 llc 0x000000000080816c
8 llc 0x0000000000808733
9 llc 0x0000000000cf5329 llvm::FPPassManager::runOnFunction(llvm::Function&) + 489
10 llc 0x0000000000cf55e6 llvm::FunctionPassManagerImpl::run(llvm::Function&) + 166
11 llc 0x0000000000cf582b llvm::FunctionPassManager::run(llvm::Function&) + 75
12 llc 0x00000000005172b9 main + 4633
13 libc.so.6 0x00007fe8ff92e5a6 __libc_start_main + 230
14 llc 0x00000000005154e9
Stack dump:
0. Program arguments: Release/bin/llc /tmp/x.bc

1.  Running pass 'X86 FP Stackifier' on function '@__ieee754_asinl'
   

Aborted

[lattner]

Verified, I guess Rafael's magic didn't get this one :)

[llvmbot]

I believe that I have encountered this bug when using llvm-2.7 and gcc-4.5.0 with dragonegg. When compiling some of my code I get the following error:

cc1plus: /home/chris/Desktop/llvm-2.7/lib/Target/X86/X86FloatingPoint.cpp:886: void::FPS::handleTwoArgFP(llvm::ilist_iteratorllvm::MachineInstr&): Assertion `UpdatedSlot < StackTop && Dest < 7' failed.
Stack dump:
0. Running pass 'X86 FP Stackifier' on function '@_ZN8Cylinder17FindCreepingFieldEP9ElecFieldRK8Vector3DS4_S4_S4_ii'
g++-4.5.0: Internal error: Aborted (program cc1plus)
Please submit a full bug report.
See http://gcc.gnu.org/bugs.html for instructions.
make: *** [cylinder.o] Error 1

I have tried using svn versions of llvm, gcc and dragonegg but I still have the same problem. The code does compile with standard gcc though.

Is there any workaround that I can try or has there been any progress in fixing this bug? Let me know if I can help in any way

[llvmbot]

Does it work if you say to target a processor with SSE support?

[llvmbot]

Compiling with -msse seems to solve the problem, thanks!

[lattner]

*** Bug llvm/llvm-bugzilla-archive#7214 has been marked as a duplicate of this bug. ***

[lattner]

*** Bug llvm/llvm-bugzilla-archive#9299 has been marked as a duplicate of this bug. ***

[lattner]

*** Bug llvm/llvm-bugzilla-archive#4434 has been marked as a duplicate of this bug. ***

[llvmbot]

*** Bug llvm/llvm-bugzilla-archive#9948 has been marked as a duplicate of this bug. ***

[llvmbot]

*** Bug llvm/llvm-bugzilla-archive#9948 has been marked as a duplicate of this bug. ***

How can this be solved ?