Related to bug #16490. This was reported also as a note in bug #18738.

Test:

$ cat test.s
.text                                                                                              
        vpgatherqd      12(,%zmm3,1), %ymm1{%k2}
        vpgatherqd      12(%rax,%zmm3,1), %ymm1{%k2}
        vpgatherqd      512(,%zmm3,1), %ymm1{%k2}
        vpgatherqd      512(%rax,%zmm3,1), %ymm1{%k2}
        vpgatherqd      %xmm2, 12(,%ymm3,1), %xmm1
$ as test.s
$ objdump -dr a.out
a.out:     file format elf64-x86-64


Disassembly of section .text:

0000000000000000 <.text>:
   0:   62 f2 7d 4a 91 0c 1d    vpgatherqd 0x3(,%zmm3,1),%ymm1{%k2}
   7:   03 00 00 00 
   b:   62 f2 7d 4a 91 4c 18    vpgatherqd 0xc(%rax,%zmm3,1),%ymm1{%k2}
  12:   03 
  13:   62 f2 7d 4a 91 0c 1d    vpgatherqd 0x200(,%zmm3,1),%ymm1{%k2}
  1a:   00 02 00 00 
  1e:   62 f2 7d 4a 91 8c 18    vpgatherqd 0x200(%rax,%zmm3,1),%ymm1{%k2}
  25:   00 02 00 00 
  29:   c4 e2 6d 91 0c 1d 0c    vpgatherqd %xmm2,0xc(,%ymm3,1),%xmm1
  30:   00 00 00 

As can be seen on the first instruction, the constant "12" became 0x3. This problem only appears to exist when the displacement is small enough to fit an 8-bit and when the base register is missing (that is, absolute pointers stored in the ZMM register).

This is likely caused by the disp8 scaling performed when the disp8 is used with a base register (see second instruction). It does not happen when the displacement is larger than 8 bits (third and fourth instructions), nor in the AVX2 equivalent.