[LLVMbugs] [Bug 2803] New: using x86.mmx. emms intrinsic can clobber results returned on the fp stack

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=2803

           Summary: using x86.mmx.emms intrinsic can clobber results
                    returned on the fp stack
           Product: libraries
           Version: trunk
          Platform: PC
        OS/Version: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: Backend: X86
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: paul at rapidmind.com
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=2023)
 --> (http://llvm.org/bugs/attachment.cgi?id=2023)
ll file to reproduce bug

To work around bug 1838 (x86 backend doesn't insert emms when it is needed)
I've manually insert calls to x86.mmx.emms before calling functions that use
the floating point stack.

e.g.,

...
%45 = extractelement <2 x float> %39, i32 1 ; <float> [#uses=1]
%46 = extractelement <2 x float> %40, i32 1 ; <float> [#uses=1]
call void @llvm.x86.mmx.emms( )
%47 = call float @fmaxf( float %45, float %46 ) ; <float> [#uses=1]
%48 = insertelement <2 x float> %44, float %47, i32 1 ; <<2 x float>> [#uses=2]
...

and here is a snippet of generated code:

call    fmaxf
fstpt   88(%esp)
movss   160(%esp), %xmm0
movss   %xmm0, 4(%esp)
movss   144(%esp), %xmm0
movss   %xmm0, (%esp)
call    fminf
emms
fstps   212(%esp)
movss   212(%esp), %xmm0
movss   %xmm0, 128(%esp)
movss   %xmm0, 4(%esp)
fldt    100(%esp)
fstps   232(%esp)
movss   232(%esp), %xmm0
movss   %xmm0, 112(%esp)
movss   %xmm0, (%esp)
call    fminf

I'm not exactly sure what's going on but either
1) emms is moving after the call to fminf
2) emms is moving before the result of a previous fminf is popped from the
floating point stack and clobbering it
3) both?

Sorry for providing such a large test program but in the generated code
sometimes the emms is in the right spot and sometimes not.


-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.