I am new to this list, so I hope I'm posting new information in the correct way. The problem described below has been solved and I'm posting the answer here for others in the future. The problem was a bug in the gdb debugger -- apparently it cannot show the upper 256 bits of a zmm register with the command i r zmm0. All 8 values are actually in the zmm register, but gdb can't show it all.
The Intel Software Developers Manual describes VCVTTPD2QQ as:
"Convert eight packed double-precision floating-point values from zmm2/m512 to eight packed quadword integers in zmm1 using truncation with writemask k1."
I am using VCVTTPD2QQ to load eight double-precision 64-bit floats into zmm1, encoded as follows:
mov rax,18446744073709551615
KMOVQ k1,rax
EVEX.512.66.0F.W1 VCVTTPD2QQ zmm1 {k1}{z},[r11+r15]
where r11 is a pointer to an array of 10,000 64-bit double-precision floats and r15 is 0 (the base address of the array). The data are read in from an external source. I want to load the first 8 into zmm1, but it loads only 4.
I fill k1 with all 1's to indicate to move all 8 numbers.
The Intel manual distinguishes the three possible encodings by the register names; if the name is a zmm register, then it should move 8.
What is wrong with my encoding that I get only 4 but not 8 data points loaded into zmm1?
