http://bugzilla.novell.com/show_bug.cgi?id=933903 Richard Biener changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |schwab@suse.com Flags|needinfo?(avi.cohenstuart@i | |nfor.com) | --- Comment #7 from Richard Biener --- memset cannot "corrupt" xmm0 as xmm0 is not saved across calls. But indeed on SLE11 I get a SIGFPE with the test program in Program received signal SIGFPE, Arithmetic exception. 0x00007ffff7b799d8 in __amd_bas64_exp () from /lib64/libm.so.6 the bug is Dump of assembler code for function __amd_bas64_exp: 0x00007ffff7b79990 <+0>: ucomisd 0x246738(%rip),%xmm0 # 0x7ffff7dc00d0 0x00007ffff7b79998 <+8>: ja 0x7ffff7b79ad0 <__amd_bas64_exp+320> 0x00007ffff7b7999e <+14>: jp 0x7ffff7b79b00 <__amd_bas64_exp+368> 0x00007ffff7b799a4 <+20>: ucomisd 0x24672c(%rip),%xmm0 # 0x7ffff7dc00d8 0x00007ffff7b799ac <+28>: jbe 0x7ffff7b79b20 <__amd_bas64_exp+400> 0x00007ffff7b799b2 <+34>: movdqa %xmm0,%xmm1 0x00007ffff7b799b6 <+38>: pand 0x246722(%rip),%xmm1 # 0x7ffff7dc00e0 0x00007ffff7b799be <+46>: ucomisd 0x246722(%rip),%xmm1 # 0x7ffff7dc00e8 0x00007ffff7b799c6 <+54>: jb 0x7ffff7b79b70 <__amd_bas64_exp+480> 0x00007ffff7b799cc <+60>: movapd %xmm0,%xmm1 0x00007ffff7b799d0 <+64>: mulsd 0x246720(%rip),%xmm1 # 0x7ffff7dc00f8 => 0x00007ffff7b799d8 <+72>: cvttpd2dq %xmm1,%xmm2 0x00007ffff7b799dc <+76>: cvtdq2pd %xmm2,%xmm1 the use of movapd here instead of movq or rather that __amd_base64_exp expects %xmm0 upper half to be zeroed? Not sure why it uses cvttpd2dq here anyway rather than the sd variants. 0x00000000004007f5 <+123>: movq $0x2,-0x10(%rbp) => 0x00000000004007fd <+131>: cvtsi2sdq -0x10(%rbp),%xmm0 cvtsi2sdq doesn't clear the upper bits, only moves do. glibc bug IMHO. Index: sysdeps/x86_64/fpu/exp2_bdozr.S =================================================================== --- /dev/null 1970-01-01 00:00:00.000000000 +0000 +++ sysdeps/x86_64/fpu/exp2_bdozr.S 2011-09-02 16:00:58.000000000 +0200 @@ -0,0 +1,182 @@ +# ... + # n = int( x * (64)) + vcvttpd2dq %xmm1, %xmm2 #xmm2 = (int)n + vcvtdq2pd %xmm2, %xmm1 #xmm1 = (double)n + vmovd %xmm2, %ecx the comment doesn't suggest why the packed variants are used either. -- You are receiving this mail because: You are on the CC list for the bug.