avx10.2-bf16-32.txt | Explore in Territory

# RUN: llvm-mc --disassemble %s -triple=i386 | FileCheck %s --check-prefixes=ATT
# RUN: llvm-mc --disassemble %s -triple=i386 --output-asm-variant=1 | FileCheck %s --check-prefixes=INTEL

# ATT:   vaddnepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vaddnepbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x58,0xd4

# ATT:   vaddnepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vaddnepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x58,0xd4

# ATT:   vaddnepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vaddnepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x58,0xd4

# ATT:   vaddnepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vaddnepbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x58,0xd4

# ATT:   vaddnepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vaddnepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x58,0xd4

# ATT:   vaddnepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vaddnepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x58,0xd4

# ATT:   vaddnepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vaddnepbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x58,0xd4

# ATT:   vaddnepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vaddnepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x58,0xd4

# ATT:   vaddnepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vaddnepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x58,0xd4

# ATT:   vaddnepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vaddnepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x58,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vaddnepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vaddnepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x58,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vaddnepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vaddnepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x58,0x10

# ATT:   vaddnepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vaddnepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x58,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vaddnepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vaddnepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x58,0x51,0x7f

# ATT:   vaddnepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vaddnepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x58,0x52,0x80

# ATT:   vaddnepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vaddnepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x58,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vaddnepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vaddnepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x58,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vaddnepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vaddnepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x58,0x10

# ATT:   vaddnepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vaddnepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x58,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vaddnepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vaddnepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x58,0x51,0x7f

# ATT:   vaddnepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vaddnepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x58,0x52,0x80

# ATT:   vaddnepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vaddnepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x58,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vaddnepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vaddnepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x58,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vaddnepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vaddnepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x58,0x10

# ATT:   vaddnepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vaddnepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x58,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vaddnepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vaddnepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x58,0x51,0x7f

# ATT:   vaddnepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vaddnepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x58,0x52,0x80

# ATT:   vcmppbf16 $123, %ymm4, %ymm3, %k5
# INTEL: vcmppbf16 k5, ymm3, ymm4, 123
0x62,0xf3,0x67,0x28,0xc2,0xec,0x7b

# ATT:   vcmppbf16 $123, %ymm4, %ymm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, ymm3, ymm4, 123
0x62,0xf3,0x67,0x2f,0xc2,0xec,0x7b

# ATT:   vcmppbf16 $123, %xmm4, %xmm3, %k5
# INTEL: vcmppbf16 k5, xmm3, xmm4, 123
0x62,0xf3,0x67,0x08,0xc2,0xec,0x7b

# ATT:   vcmppbf16 $123, %xmm4, %xmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, xmm3, xmm4, 123
0x62,0xf3,0x67,0x0f,0xc2,0xec,0x7b

# ATT:   vcmppbf16 $123, %zmm4, %zmm3, %k5
# INTEL: vcmppbf16 k5, zmm3, zmm4, 123
0x62,0xf3,0x67,0x48,0xc2,0xec,0x7b

# ATT:   vcmppbf16 $123, %zmm4, %zmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, zmm3, zmm4, 123
0x62,0xf3,0x67,0x4f,0xc2,0xec,0x7b

# ATT:   vcmppbf16  $123, 268435456(%esp,%esi,8), %zmm3, %k5
# INTEL: vcmppbf16 k5, zmm3, zmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x67,0x48,0xc2,0xac,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vcmppbf16  $123, 291(%edi,%eax,4), %zmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x67,0x4f,0xc2,0xac,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vcmppbf16  $123, (%eax){1to32}, %zmm3, %k5
# INTEL: vcmppbf16 k5, zmm3, word ptr [eax]{1to32}, 123
0x62,0xf3,0x67,0x58,0xc2,0x28,0x7b

# ATT:   vcmppbf16  $123, -2048(,%ebp,2), %zmm3, %k5
# INTEL: vcmppbf16 k5, zmm3, zmmword ptr [2*ebp - 2048], 123
0x62,0xf3,0x67,0x48,0xc2,0x2c,0x6d,0x00,0xf8,0xff,0xff,0x7b

# ATT:   vcmppbf16  $123, 8128(%ecx), %zmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, zmm3, zmmword ptr [ecx + 8128], 123
0x62,0xf3,0x67,0x4f,0xc2,0x69,0x7f,0x7b

# ATT:   vcmppbf16  $123, -256(%edx){1to32}, %zmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, zmm3, word ptr [edx - 256]{1to32}, 123
0x62,0xf3,0x67,0x5f,0xc2,0x6a,0x80,0x7b

# ATT:   vcmppbf16  $123, 268435456(%esp,%esi,8), %xmm3, %k5
# INTEL: vcmppbf16 k5, xmm3, xmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x67,0x08,0xc2,0xac,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vcmppbf16  $123, 291(%edi,%eax,4), %xmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x67,0x0f,0xc2,0xac,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vcmppbf16  $123, (%eax){1to8}, %xmm3, %k5
# INTEL: vcmppbf16 k5, xmm3, word ptr [eax]{1to8}, 123
0x62,0xf3,0x67,0x18,0xc2,0x28,0x7b

# ATT:   vcmppbf16  $123, -512(,%ebp,2), %xmm3, %k5
# INTEL: vcmppbf16 k5, xmm3, xmmword ptr [2*ebp - 512], 123
0x62,0xf3,0x67,0x08,0xc2,0x2c,0x6d,0x00,0xfe,0xff,0xff,0x7b

# ATT:   vcmppbf16  $123, 2032(%ecx), %xmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, xmm3, xmmword ptr [ecx + 2032], 123
0x62,0xf3,0x67,0x0f,0xc2,0x69,0x7f,0x7b

# ATT:   vcmppbf16  $123, -256(%edx){1to8}, %xmm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, xmm3, word ptr [edx - 256]{1to8}, 123
0x62,0xf3,0x67,0x1f,0xc2,0x6a,0x80,0x7b

# ATT:   vcmppbf16  $123, 268435456(%esp,%esi,8), %ymm3, %k5
# INTEL: vcmppbf16 k5, ymm3, ymmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x67,0x28,0xc2,0xac,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vcmppbf16  $123, 291(%edi,%eax,4), %ymm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x67,0x2f,0xc2,0xac,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vcmppbf16  $123, (%eax){1to16}, %ymm3, %k5
# INTEL: vcmppbf16 k5, ymm3, word ptr [eax]{1to16}, 123
0x62,0xf3,0x67,0x38,0xc2,0x28,0x7b

# ATT:   vcmppbf16  $123, -1024(,%ebp,2), %ymm3, %k5
# INTEL: vcmppbf16 k5, ymm3, ymmword ptr [2*ebp - 1024], 123
0x62,0xf3,0x67,0x28,0xc2,0x2c,0x6d,0x00,0xfc,0xff,0xff,0x7b

# ATT:   vcmppbf16  $123, 4064(%ecx), %ymm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, ymm3, ymmword ptr [ecx + 4064], 123
0x62,0xf3,0x67,0x2f,0xc2,0x69,0x7f,0x7b

# ATT:   vcmppbf16  $123, -256(%edx){1to16}, %ymm3, %k5 {%k7}
# INTEL: vcmppbf16 k5 {k7}, ymm3, word ptr [edx - 256]{1to16}, 123
0x62,0xf3,0x67,0x3f,0xc2,0x6a,0x80,0x7b

# ATT:   vcomsbf16 %xmm3, %xmm2
# INTEL: vcomsbf16 xmm2, xmm3
0x62,0xf5,0x7d,0x08,0x2f,0xd3

# ATT:   vcomsbf16  268435456(%esp,%esi,8), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x08,0x2f,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vcomsbf16  291(%edi,%eax,4), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x08,0x2f,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vcomsbf16  (%eax), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [eax]
0x62,0xf5,0x7d,0x08,0x2f,0x10

# ATT:   vcomsbf16  -64(,%ebp,2), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [2*ebp - 64]
0x62,0xf5,0x7d,0x08,0x2f,0x14,0x6d,0xc0,0xff,0xff,0xff

# ATT:   vcomsbf16  254(%ecx), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [ecx + 254]
0x62,0xf5,0x7d,0x08,0x2f,0x51,0x7f

# ATT:   vcomsbf16  -256(%edx), %xmm2
# INTEL: vcomsbf16 xmm2, word ptr [edx - 256]
0x62,0xf5,0x7d,0x08,0x2f,0x52,0x80

# ATT:   vdivnepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vdivnepbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x5e,0xd4

# ATT:   vdivnepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vdivnepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x5e,0xd4

# ATT:   vdivnepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vdivnepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x5e,0xd4

# ATT:   vdivnepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vdivnepbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x5e,0xd4

# ATT:   vdivnepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vdivnepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x5e,0xd4

# ATT:   vdivnepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vdivnepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x5e,0xd4

# ATT:   vdivnepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vdivnepbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x5e,0xd4

# ATT:   vdivnepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vdivnepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x5e,0xd4

# ATT:   vdivnepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vdivnepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x5e,0xd4

# ATT:   vdivnepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vdivnepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x5e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vdivnepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vdivnepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x5e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vdivnepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vdivnepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x5e,0x10

# ATT:   vdivnepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vdivnepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x5e,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vdivnepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vdivnepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x5e,0x51,0x7f

# ATT:   vdivnepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vdivnepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x5e,0x52,0x80

# ATT:   vdivnepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vdivnepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x5e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vdivnepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vdivnepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x5e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vdivnepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vdivnepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x5e,0x10

# ATT:   vdivnepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vdivnepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x5e,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vdivnepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vdivnepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x5e,0x51,0x7f

# ATT:   vdivnepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vdivnepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x5e,0x52,0x80

# ATT:   vdivnepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vdivnepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x5e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vdivnepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vdivnepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x5e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vdivnepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vdivnepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x5e,0x10

# ATT:   vdivnepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vdivnepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x5e,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vdivnepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vdivnepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x5e,0x51,0x7f

# ATT:   vdivnepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vdivnepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x5e,0x52,0x80

# ATT:   vfmadd132nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmadd132nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0x98,0xd4

# ATT:   vfmadd132nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmadd132nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0x98,0xd4

# ATT:   vfmadd132nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0x98,0xd4

# ATT:   vfmadd132nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmadd132nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0x98,0xd4

# ATT:   vfmadd132nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmadd132nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0x98,0xd4

# ATT:   vfmadd132nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0x98,0xd4

# ATT:   vfmadd132nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmadd132nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0x98,0xd4

# ATT:   vfmadd132nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmadd132nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0x98,0xd4

# ATT:   vfmadd132nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0x98,0xd4

# ATT:   vfmadd132nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmadd132nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0x98,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd132nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmadd132nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0x98,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd132nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmadd132nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0x98,0x10

# ATT:   vfmadd132nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmadd132nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0x98,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmadd132nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0x98,0x51,0x7f

# ATT:   vfmadd132nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0x98,0x52,0x80

# ATT:   vfmadd132nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmadd132nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0x98,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd132nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmadd132nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0x98,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd132nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmadd132nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0x98,0x10

# ATT:   vfmadd132nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmadd132nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0x98,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmadd132nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0x98,0x51,0x7f

# ATT:   vfmadd132nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0x98,0x52,0x80

# ATT:   vfmadd132nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmadd132nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0x98,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd132nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmadd132nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0x98,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd132nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmadd132nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0x98,0x10

# ATT:   vfmadd132nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmadd132nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0x98,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmadd132nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0x98,0x51,0x7f

# ATT:   vfmadd132nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd132nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0x98,0x52,0x80

# ATT:   vfmadd213nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmadd213nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmadd213nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmadd213nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmadd213nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmadd213nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmadd213nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xa8,0xd4

# ATT:   vfmadd213nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xa8,0xd4

# ATT:   vfmadd213nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmadd213nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xa8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd213nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmadd213nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xa8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd213nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmadd213nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xa8,0x10

# ATT:   vfmadd213nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmadd213nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xa8,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmadd213nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xa8,0x51,0x7f

# ATT:   vfmadd213nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xa8,0x52,0x80

# ATT:   vfmadd213nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmadd213nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xa8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd213nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmadd213nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xa8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd213nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmadd213nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xa8,0x10

# ATT:   vfmadd213nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmadd213nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xa8,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmadd213nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xa8,0x51,0x7f

# ATT:   vfmadd213nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xa8,0x52,0x80

# ATT:   vfmadd213nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmadd213nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xa8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd213nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmadd213nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xa8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd213nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmadd213nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xa8,0x10

# ATT:   vfmadd213nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmadd213nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xa8,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmadd213nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xa8,0x51,0x7f

# ATT:   vfmadd213nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd213nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xa8,0x52,0x80

# ATT:   vfmadd231nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmadd231nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmadd231nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmadd231nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmadd231nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmadd231nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmadd231nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xb8,0xd4

# ATT:   vfmadd231nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xb8,0xd4

# ATT:   vfmadd231nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmadd231nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xb8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd231nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmadd231nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xb8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd231nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmadd231nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xb8,0x10

# ATT:   vfmadd231nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmadd231nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xb8,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmadd231nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xb8,0x51,0x7f

# ATT:   vfmadd231nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xb8,0x52,0x80

# ATT:   vfmadd231nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmadd231nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xb8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd231nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmadd231nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xb8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd231nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmadd231nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xb8,0x10

# ATT:   vfmadd231nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmadd231nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xb8,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmadd231nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xb8,0x51,0x7f

# ATT:   vfmadd231nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xb8,0x52,0x80

# ATT:   vfmadd231nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmadd231nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xb8,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmadd231nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmadd231nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xb8,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmadd231nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmadd231nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xb8,0x10

# ATT:   vfmadd231nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmadd231nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xb8,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmadd231nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xb8,0x51,0x7f

# ATT:   vfmadd231nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmadd231nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xb8,0x52,0x80

# ATT:   vfmsub132nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmsub132nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmsub132nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmsub132nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmsub132nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmsub132nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmsub132nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0x9a,0xd4

# ATT:   vfmsub132nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0x9a,0xd4

# ATT:   vfmsub132nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmsub132nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0x9a,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub132nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmsub132nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0x9a,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub132nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmsub132nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0x9a,0x10

# ATT:   vfmsub132nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmsub132nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0x9a,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmsub132nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0x9a,0x51,0x7f

# ATT:   vfmsub132nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0x9a,0x52,0x80

# ATT:   vfmsub132nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmsub132nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0x9a,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub132nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmsub132nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0x9a,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub132nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmsub132nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0x9a,0x10

# ATT:   vfmsub132nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmsub132nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0x9a,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmsub132nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0x9a,0x51,0x7f

# ATT:   vfmsub132nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0x9a,0x52,0x80

# ATT:   vfmsub132nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmsub132nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0x9a,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub132nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmsub132nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0x9a,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub132nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmsub132nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0x9a,0x10

# ATT:   vfmsub132nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmsub132nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0x9a,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmsub132nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0x9a,0x51,0x7f

# ATT:   vfmsub132nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub132nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0x9a,0x52,0x80

# ATT:   vfmsub213nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmsub213nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmsub213nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmsub213nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmsub213nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmsub213nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmsub213nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xaa,0xd4

# ATT:   vfmsub213nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xaa,0xd4

# ATT:   vfmsub213nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmsub213nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xaa,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub213nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmsub213nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xaa,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub213nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmsub213nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xaa,0x10

# ATT:   vfmsub213nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmsub213nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xaa,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmsub213nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xaa,0x51,0x7f

# ATT:   vfmsub213nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xaa,0x52,0x80

# ATT:   vfmsub213nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmsub213nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xaa,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub213nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmsub213nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xaa,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub213nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmsub213nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xaa,0x10

# ATT:   vfmsub213nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmsub213nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xaa,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmsub213nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xaa,0x51,0x7f

# ATT:   vfmsub213nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xaa,0x52,0x80

# ATT:   vfmsub213nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmsub213nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xaa,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub213nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmsub213nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xaa,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub213nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmsub213nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xaa,0x10

# ATT:   vfmsub213nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmsub213nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xaa,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmsub213nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xaa,0x51,0x7f

# ATT:   vfmsub213nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub213nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xaa,0x52,0x80

# ATT:   vfmsub231nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfmsub231nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xba,0xd4

# ATT:   vfmsub231nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfmsub231nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xba,0xd4

# ATT:   vfmsub231nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xba,0xd4

# ATT:   vfmsub231nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfmsub231nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xba,0xd4

# ATT:   vfmsub231nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfmsub231nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xba,0xd4

# ATT:   vfmsub231nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xba,0xd4

# ATT:   vfmsub231nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfmsub231nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xba,0xd4

# ATT:   vfmsub231nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfmsub231nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xba,0xd4

# ATT:   vfmsub231nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xba,0xd4

# ATT:   vfmsub231nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfmsub231nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xba,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub231nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfmsub231nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xba,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub231nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfmsub231nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xba,0x10

# ATT:   vfmsub231nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfmsub231nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xba,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfmsub231nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xba,0x51,0x7f

# ATT:   vfmsub231nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xba,0x52,0x80

# ATT:   vfmsub231nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfmsub231nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xba,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub231nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfmsub231nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xba,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub231nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfmsub231nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xba,0x10

# ATT:   vfmsub231nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfmsub231nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xba,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfmsub231nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xba,0x51,0x7f

# ATT:   vfmsub231nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xba,0x52,0x80

# ATT:   vfmsub231nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfmsub231nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xba,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfmsub231nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfmsub231nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xba,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfmsub231nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfmsub231nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xba,0x10

# ATT:   vfmsub231nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfmsub231nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xba,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfmsub231nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xba,0x51,0x7f

# ATT:   vfmsub231nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfmsub231nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xba,0x52,0x80

# ATT:   vfnmadd132nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmadd132nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd132nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmadd132nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd132nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmadd132nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd132nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0x9c,0xd4

# ATT:   vfnmadd132nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0x9c,0xd4

# ATT:   vfnmadd132nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmadd132nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0x9c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd132nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd132nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0x9c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd132nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmadd132nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0x9c,0x10

# ATT:   vfnmadd132nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmadd132nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0x9c,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmadd132nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0x9c,0x51,0x7f

# ATT:   vfnmadd132nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0x9c,0x52,0x80

# ATT:   vfnmadd132nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmadd132nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0x9c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd132nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd132nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0x9c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd132nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmadd132nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0x9c,0x10

# ATT:   vfnmadd132nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmadd132nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0x9c,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmadd132nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0x9c,0x51,0x7f

# ATT:   vfnmadd132nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0x9c,0x52,0x80

# ATT:   vfnmadd132nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmadd132nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0x9c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd132nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd132nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0x9c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd132nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmadd132nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0x9c,0x10

# ATT:   vfnmadd132nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmadd132nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0x9c,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmadd132nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0x9c,0x51,0x7f

# ATT:   vfnmadd132nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd132nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0x9c,0x52,0x80

# ATT:   vfnmadd213nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmadd213nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd213nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmadd213nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd213nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmadd213nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd213nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xac,0xd4

# ATT:   vfnmadd213nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xac,0xd4

# ATT:   vfnmadd213nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmadd213nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xac,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd213nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd213nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xac,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd213nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmadd213nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xac,0x10

# ATT:   vfnmadd213nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmadd213nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xac,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmadd213nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xac,0x51,0x7f

# ATT:   vfnmadd213nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xac,0x52,0x80

# ATT:   vfnmadd213nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmadd213nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xac,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd213nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd213nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xac,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd213nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmadd213nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xac,0x10

# ATT:   vfnmadd213nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmadd213nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xac,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmadd213nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xac,0x51,0x7f

# ATT:   vfnmadd213nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xac,0x52,0x80

# ATT:   vfnmadd213nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmadd213nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xac,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd213nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd213nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xac,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd213nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmadd213nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xac,0x10

# ATT:   vfnmadd213nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmadd213nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xac,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmadd213nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xac,0x51,0x7f

# ATT:   vfnmadd213nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd213nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xac,0x52,0x80

# ATT:   vfnmadd231nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmadd231nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd231nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmadd231nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd231nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmadd231nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd231nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xbc,0xd4

# ATT:   vfnmadd231nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xbc,0xd4

# ATT:   vfnmadd231nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmadd231nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xbc,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd231nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmadd231nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xbc,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd231nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmadd231nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xbc,0x10

# ATT:   vfnmadd231nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmadd231nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xbc,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmadd231nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xbc,0x51,0x7f

# ATT:   vfnmadd231nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xbc,0x52,0x80

# ATT:   vfnmadd231nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmadd231nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xbc,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd231nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmadd231nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xbc,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd231nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmadd231nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xbc,0x10

# ATT:   vfnmadd231nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmadd231nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xbc,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmadd231nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xbc,0x51,0x7f

# ATT:   vfnmadd231nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xbc,0x52,0x80

# ATT:   vfnmadd231nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmadd231nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xbc,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmadd231nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmadd231nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xbc,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmadd231nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmadd231nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xbc,0x10

# ATT:   vfnmadd231nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmadd231nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xbc,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmadd231nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xbc,0x51,0x7f

# ATT:   vfnmadd231nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmadd231nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xbc,0x52,0x80

# ATT:   vfnmsub132nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmsub132nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub132nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmsub132nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub132nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmsub132nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub132nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0x9e,0xd4

# ATT:   vfnmsub132nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0x9e,0xd4

# ATT:   vfnmsub132nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmsub132nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0x9e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub132nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub132nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0x9e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub132nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmsub132nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0x9e,0x10

# ATT:   vfnmsub132nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmsub132nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0x9e,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmsub132nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0x9e,0x51,0x7f

# ATT:   vfnmsub132nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0x9e,0x52,0x80

# ATT:   vfnmsub132nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmsub132nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0x9e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub132nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub132nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0x9e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub132nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmsub132nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0x9e,0x10

# ATT:   vfnmsub132nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmsub132nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0x9e,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmsub132nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0x9e,0x51,0x7f

# ATT:   vfnmsub132nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0x9e,0x52,0x80

# ATT:   vfnmsub132nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmsub132nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0x9e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub132nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub132nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0x9e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub132nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmsub132nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0x9e,0x10

# ATT:   vfnmsub132nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmsub132nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0x9e,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmsub132nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0x9e,0x51,0x7f

# ATT:   vfnmsub132nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub132nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0x9e,0x52,0x80

# ATT:   vfnmsub213nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmsub213nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub213nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmsub213nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub213nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmsub213nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub213nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xae,0xd4

# ATT:   vfnmsub213nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xae,0xd4

# ATT:   vfnmsub213nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmsub213nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xae,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub213nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub213nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xae,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub213nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmsub213nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xae,0x10

# ATT:   vfnmsub213nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmsub213nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xae,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmsub213nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xae,0x51,0x7f

# ATT:   vfnmsub213nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xae,0x52,0x80

# ATT:   vfnmsub213nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmsub213nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xae,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub213nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub213nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xae,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub213nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmsub213nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xae,0x10

# ATT:   vfnmsub213nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmsub213nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xae,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmsub213nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xae,0x51,0x7f

# ATT:   vfnmsub213nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xae,0x52,0x80

# ATT:   vfnmsub213nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmsub213nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xae,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub213nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub213nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xae,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub213nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmsub213nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xae,0x10

# ATT:   vfnmsub213nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmsub213nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xae,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmsub213nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xae,0x51,0x7f

# ATT:   vfnmsub213nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub213nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xae,0x52,0x80

# ATT:   vfnmsub231nepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vfnmsub231nepbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub231nepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vfnmsub231nepbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub231nepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vfnmsub231nepbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub231nepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0xbe,0xd4

# ATT:   vfnmsub231nepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0xbe,0xd4

# ATT:   vfnmsub231nepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vfnmsub231nepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0xbe,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub231nepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vfnmsub231nepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0xbe,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub231nepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vfnmsub231nepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0xbe,0x10

# ATT:   vfnmsub231nepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vfnmsub231nepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0xbe,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vfnmsub231nepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0xbe,0x51,0x7f

# ATT:   vfnmsub231nepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0xbe,0x52,0x80

# ATT:   vfnmsub231nepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vfnmsub231nepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0xbe,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub231nepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vfnmsub231nepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0xbe,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub231nepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vfnmsub231nepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0xbe,0x10

# ATT:   vfnmsub231nepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vfnmsub231nepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0xbe,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vfnmsub231nepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0xbe,0x51,0x7f

# ATT:   vfnmsub231nepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0xbe,0x52,0x80

# ATT:   vfnmsub231nepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vfnmsub231nepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0xbe,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vfnmsub231nepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vfnmsub231nepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0xbe,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vfnmsub231nepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vfnmsub231nepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0xbe,0x10

# ATT:   vfnmsub231nepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vfnmsub231nepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0xbe,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vfnmsub231nepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0xbe,0x51,0x7f

# ATT:   vfnmsub231nepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vfnmsub231nepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0xbe,0x52,0x80

# ATT:   vfpclasspbf16 $123, %zmm3, %k5
# INTEL: vfpclasspbf16 k5, zmm3, 123
0x62,0xf3,0x7f,0x48,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16 $123, %zmm3, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, zmm3, 123
0x62,0xf3,0x7f,0x4f,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16 $123, %ymm3, %k5
# INTEL: vfpclasspbf16 k5, ymm3, 123
0x62,0xf3,0x7f,0x28,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16 $123, %ymm3, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, ymm3, 123
0x62,0xf3,0x7f,0x2f,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16 $123, %xmm3, %k5
# INTEL: vfpclasspbf16 k5, xmm3, 123
0x62,0xf3,0x7f,0x08,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16 $123, %xmm3, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, xmm3, 123
0x62,0xf3,0x7f,0x0f,0x66,0xeb,0x7b

# ATT:   vfpclasspbf16x  $123, 268435456(%esp,%esi,8), %k5
# INTEL: vfpclasspbf16 k5, xmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x08,0x66,0xac,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vfpclasspbf16x  $123, 291(%edi,%eax,4), %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, xmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x0f,0x66,0xac,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vfpclasspbf16  $123, (%eax){1to8}, %k5
# INTEL: vfpclasspbf16 k5, word ptr [eax]{1to8}, 123
0x62,0xf3,0x7f,0x18,0x66,0x28,0x7b

# ATT:   vfpclasspbf16x  $123, -512(,%ebp,2), %k5
# INTEL: vfpclasspbf16 k5, xmmword ptr [2*ebp - 512], 123
0x62,0xf3,0x7f,0x08,0x66,0x2c,0x6d,0x00,0xfe,0xff,0xff,0x7b

# ATT:   vfpclasspbf16x  $123, 2032(%ecx), %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, xmmword ptr [ecx + 2032], 123
0x62,0xf3,0x7f,0x0f,0x66,0x69,0x7f,0x7b

# ATT:   vfpclasspbf16  $123, -256(%edx){1to8}, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, word ptr [edx - 256]{1to8}, 123
0x62,0xf3,0x7f,0x1f,0x66,0x6a,0x80,0x7b

# ATT:   vfpclasspbf16  $123, (%eax){1to16}, %k5
# INTEL: vfpclasspbf16 k5, word ptr [eax]{1to16}, 123
0x62,0xf3,0x7f,0x38,0x66,0x28,0x7b

# ATT:   vfpclasspbf16y  $123, -1024(,%ebp,2), %k5
# INTEL: vfpclasspbf16 k5, ymmword ptr [2*ebp - 1024], 123
0x62,0xf3,0x7f,0x28,0x66,0x2c,0x6d,0x00,0xfc,0xff,0xff,0x7b

# ATT:   vfpclasspbf16y  $123, 4064(%ecx), %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, ymmword ptr [ecx + 4064], 123
0x62,0xf3,0x7f,0x2f,0x66,0x69,0x7f,0x7b

# ATT:   vfpclasspbf16  $123, -256(%edx){1to16}, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, word ptr [edx - 256]{1to16}, 123
0x62,0xf3,0x7f,0x3f,0x66,0x6a,0x80,0x7b

# ATT:   vfpclasspbf16  $123, (%eax){1to32}, %k5
# INTEL: vfpclasspbf16 k5, word ptr [eax]{1to32}, 123
0x62,0xf3,0x7f,0x58,0x66,0x28,0x7b

# ATT:   vfpclasspbf16z  $123, -2048(,%ebp,2), %k5
# INTEL: vfpclasspbf16 k5, zmmword ptr [2*ebp - 2048], 123
0x62,0xf3,0x7f,0x48,0x66,0x2c,0x6d,0x00,0xf8,0xff,0xff,0x7b

# ATT:   vfpclasspbf16z  $123, 8128(%ecx), %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, zmmword ptr [ecx + 8128], 123
0x62,0xf3,0x7f,0x4f,0x66,0x69,0x7f,0x7b

# ATT:   vfpclasspbf16  $123, -256(%edx){1to32}, %k5 {%k7}
# INTEL: vfpclasspbf16 k5 {k7}, word ptr [edx - 256]{1to32}, 123
0x62,0xf3,0x7f,0x5f,0x66,0x6a,0x80,0x7b

# ATT:   vgetexppbf16 %xmm3, %xmm2
# INTEL: vgetexppbf16 xmm2, xmm3
0x62,0xf5,0x7d,0x08,0x42,0xd3

# ATT:   vgetexppbf16 %xmm3, %xmm2 {%k7}
# INTEL: vgetexppbf16 xmm2 {k7}, xmm3
0x62,0xf5,0x7d,0x0f,0x42,0xd3

# ATT:   vgetexppbf16 %xmm3, %xmm2 {%k7} {z}
# INTEL: vgetexppbf16 xmm2 {k7} {z}, xmm3
0x62,0xf5,0x7d,0x8f,0x42,0xd3

# ATT:   vgetexppbf16 %zmm3, %zmm2
# INTEL: vgetexppbf16 zmm2, zmm3
0x62,0xf5,0x7d,0x48,0x42,0xd3

# ATT:   vgetexppbf16 %zmm3, %zmm2 {%k7}
# INTEL: vgetexppbf16 zmm2 {k7}, zmm3
0x62,0xf5,0x7d,0x4f,0x42,0xd3

# ATT:   vgetexppbf16 %zmm3, %zmm2 {%k7} {z}
# INTEL: vgetexppbf16 zmm2 {k7} {z}, zmm3
0x62,0xf5,0x7d,0xcf,0x42,0xd3

# ATT:   vgetexppbf16 %ymm3, %ymm2
# INTEL: vgetexppbf16 ymm2, ymm3
0x62,0xf5,0x7d,0x28,0x42,0xd3

# ATT:   vgetexppbf16 %ymm3, %ymm2 {%k7}
# INTEL: vgetexppbf16 ymm2 {k7}, ymm3
0x62,0xf5,0x7d,0x2f,0x42,0xd3

# ATT:   vgetexppbf16 %ymm3, %ymm2 {%k7} {z}
# INTEL: vgetexppbf16 ymm2 {k7} {z}, ymm3
0x62,0xf5,0x7d,0xaf,0x42,0xd3

# ATT:   vgetexppbf16  268435456(%esp,%esi,8), %xmm2
# INTEL: vgetexppbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x08,0x42,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vgetexppbf16  291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vgetexppbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x0f,0x42,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vgetexppbf16  (%eax){1to8}, %xmm2
# INTEL: vgetexppbf16 xmm2, word ptr [eax]{1to8}
0x62,0xf5,0x7d,0x18,0x42,0x10

# ATT:   vgetexppbf16  -512(,%ebp,2), %xmm2
# INTEL: vgetexppbf16 xmm2, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x7d,0x08,0x42,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vgetexppbf16  2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vgetexppbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032]
0x62,0xf5,0x7d,0x8f,0x42,0x51,0x7f

# ATT:   vgetexppbf16  -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vgetexppbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}
0x62,0xf5,0x7d,0x9f,0x42,0x52,0x80

# ATT:   vgetexppbf16  268435456(%esp,%esi,8), %ymm2
# INTEL: vgetexppbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x28,0x42,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vgetexppbf16  291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vgetexppbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x2f,0x42,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vgetexppbf16  (%eax){1to16}, %ymm2
# INTEL: vgetexppbf16 ymm2, word ptr [eax]{1to16}
0x62,0xf5,0x7d,0x38,0x42,0x10

# ATT:   vgetexppbf16  -1024(,%ebp,2), %ymm2
# INTEL: vgetexppbf16 ymm2, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x7d,0x28,0x42,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vgetexppbf16  4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vgetexppbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064]
0x62,0xf5,0x7d,0xaf,0x42,0x51,0x7f

# ATT:   vgetexppbf16  -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vgetexppbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}
0x62,0xf5,0x7d,0xbf,0x42,0x52,0x80

# ATT:   vgetexppbf16  268435456(%esp,%esi,8), %zmm2
# INTEL: vgetexppbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x48,0x42,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vgetexppbf16  291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vgetexppbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x4f,0x42,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vgetexppbf16  (%eax){1to32}, %zmm2
# INTEL: vgetexppbf16 zmm2, word ptr [eax]{1to32}
0x62,0xf5,0x7d,0x58,0x42,0x10

# ATT:   vgetexppbf16  -2048(,%ebp,2), %zmm2
# INTEL: vgetexppbf16 zmm2, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x7d,0x48,0x42,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vgetexppbf16  8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vgetexppbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128]
0x62,0xf5,0x7d,0xcf,0x42,0x51,0x7f

# ATT:   vgetexppbf16  -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vgetexppbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}
0x62,0xf5,0x7d,0xdf,0x42,0x52,0x80

# ATT:   vgetmantpbf16 $123, %zmm3, %zmm2
# INTEL: vgetmantpbf16 zmm2, zmm3, 123
0x62,0xf3,0x7f,0x48,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %zmm3, %zmm2 {%k7}
# INTEL: vgetmantpbf16 zmm2 {k7}, zmm3, 123
0x62,0xf3,0x7f,0x4f,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %zmm3, %zmm2 {%k7} {z}
# INTEL: vgetmantpbf16 zmm2 {k7} {z}, zmm3, 123
0x62,0xf3,0x7f,0xcf,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %ymm3, %ymm2
# INTEL: vgetmantpbf16 ymm2, ymm3, 123
0x62,0xf3,0x7f,0x28,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %ymm3, %ymm2 {%k7}
# INTEL: vgetmantpbf16 ymm2 {k7}, ymm3, 123
0x62,0xf3,0x7f,0x2f,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %ymm3, %ymm2 {%k7} {z}
# INTEL: vgetmantpbf16 ymm2 {k7} {z}, ymm3, 123
0x62,0xf3,0x7f,0xaf,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %xmm3, %xmm2
# INTEL: vgetmantpbf16 xmm2, xmm3, 123
0x62,0xf3,0x7f,0x08,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %xmm3, %xmm2 {%k7}
# INTEL: vgetmantpbf16 xmm2 {k7}, xmm3, 123
0x62,0xf3,0x7f,0x0f,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16 $123, %xmm3, %xmm2 {%k7} {z}
# INTEL: vgetmantpbf16 xmm2 {k7} {z}, xmm3, 123
0x62,0xf3,0x7f,0x8f,0x26,0xd3,0x7b

# ATT:   vgetmantpbf16  $123, 268435456(%esp,%esi,8), %xmm2
# INTEL: vgetmantpbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x08,0x26,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vgetmantpbf16  $123, 291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vgetmantpbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x0f,0x26,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vgetmantpbf16  $123, (%eax){1to8}, %xmm2
# INTEL: vgetmantpbf16 xmm2, word ptr [eax]{1to8}, 123
0x62,0xf3,0x7f,0x18,0x26,0x10,0x7b

# ATT:   vgetmantpbf16  $123, -512(,%ebp,2), %xmm2
# INTEL: vgetmantpbf16 xmm2, xmmword ptr [2*ebp - 512], 123
0x62,0xf3,0x7f,0x08,0x26,0x14,0x6d,0x00,0xfe,0xff,0xff,0x7b

# ATT:   vgetmantpbf16  $123, 2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vgetmantpbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032], 123
0x62,0xf3,0x7f,0x8f,0x26,0x51,0x7f,0x7b

# ATT:   vgetmantpbf16  $123, -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vgetmantpbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}, 123
0x62,0xf3,0x7f,0x9f,0x26,0x52,0x80,0x7b

# ATT:   vgetmantpbf16  $123, 268435456(%esp,%esi,8), %ymm2
# INTEL: vgetmantpbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x28,0x26,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vgetmantpbf16  $123, 291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vgetmantpbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x2f,0x26,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vgetmantpbf16  $123, (%eax){1to16}, %ymm2
# INTEL: vgetmantpbf16 ymm2, word ptr [eax]{1to16}, 123
0x62,0xf3,0x7f,0x38,0x26,0x10,0x7b

# ATT:   vgetmantpbf16  $123, -1024(,%ebp,2), %ymm2
# INTEL: vgetmantpbf16 ymm2, ymmword ptr [2*ebp - 1024], 123
0x62,0xf3,0x7f,0x28,0x26,0x14,0x6d,0x00,0xfc,0xff,0xff,0x7b

# ATT:   vgetmantpbf16  $123, 4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vgetmantpbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064], 123
0x62,0xf3,0x7f,0xaf,0x26,0x51,0x7f,0x7b

# ATT:   vgetmantpbf16  $123, -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vgetmantpbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}, 123
0x62,0xf3,0x7f,0xbf,0x26,0x52,0x80,0x7b

# ATT:   vgetmantpbf16  $123, 268435456(%esp,%esi,8), %zmm2
# INTEL: vgetmantpbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x48,0x26,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vgetmantpbf16  $123, 291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vgetmantpbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x4f,0x26,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vgetmantpbf16  $123, (%eax){1to32}, %zmm2
# INTEL: vgetmantpbf16 zmm2, word ptr [eax]{1to32}, 123
0x62,0xf3,0x7f,0x58,0x26,0x10,0x7b

# ATT:   vgetmantpbf16  $123, -2048(,%ebp,2), %zmm2
# INTEL: vgetmantpbf16 zmm2, zmmword ptr [2*ebp - 2048], 123
0x62,0xf3,0x7f,0x48,0x26,0x14,0x6d,0x00,0xf8,0xff,0xff,0x7b

# ATT:   vgetmantpbf16  $123, 8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vgetmantpbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128], 123
0x62,0xf3,0x7f,0xcf,0x26,0x51,0x7f,0x7b

# ATT:   vgetmantpbf16  $123, -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vgetmantpbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}, 123
0x62,0xf3,0x7f,0xdf,0x26,0x52,0x80,0x7b

# ATT:   vmaxpbf16 %ymm4, %ymm3, %ymm2
# INTEL: vmaxpbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x5f,0xd4

# ATT:   vmaxpbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vmaxpbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x5f,0xd4

# ATT:   vmaxpbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vmaxpbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x5f,0xd4

# ATT:   vmaxpbf16 %zmm4, %zmm3, %zmm2
# INTEL: vmaxpbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x5f,0xd4

# ATT:   vmaxpbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vmaxpbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x5f,0xd4

# ATT:   vmaxpbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vmaxpbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x5f,0xd4

# ATT:   vmaxpbf16 %xmm4, %xmm3, %xmm2
# INTEL: vmaxpbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x5f,0xd4

# ATT:   vmaxpbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vmaxpbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x5f,0xd4

# ATT:   vmaxpbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vmaxpbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x5f,0xd4

# ATT:   vmaxpbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vmaxpbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x5f,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmaxpbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vmaxpbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x5f,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmaxpbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vmaxpbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x5f,0x10

# ATT:   vmaxpbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vmaxpbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x5f,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vmaxpbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vmaxpbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x5f,0x51,0x7f

# ATT:   vmaxpbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vmaxpbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x5f,0x52,0x80

# ATT:   vmaxpbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vmaxpbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x5f,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmaxpbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vmaxpbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x5f,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmaxpbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vmaxpbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x5f,0x10

# ATT:   vmaxpbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vmaxpbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x5f,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vmaxpbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vmaxpbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x5f,0x51,0x7f

# ATT:   vmaxpbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vmaxpbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x5f,0x52,0x80

# ATT:   vmaxpbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vmaxpbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x5f,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmaxpbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vmaxpbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x5f,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmaxpbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vmaxpbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x5f,0x10

# ATT:   vmaxpbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vmaxpbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x5f,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vmaxpbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vmaxpbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x5f,0x51,0x7f

# ATT:   vmaxpbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vmaxpbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x5f,0x52,0x80

# ATT:   vminpbf16 %ymm4, %ymm3, %ymm2
# INTEL: vminpbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x5d,0xd4

# ATT:   vminpbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vminpbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x5d,0xd4

# ATT:   vminpbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vminpbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x5d,0xd4

# ATT:   vminpbf16 %zmm4, %zmm3, %zmm2
# INTEL: vminpbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x5d,0xd4

# ATT:   vminpbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vminpbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x5d,0xd4

# ATT:   vminpbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vminpbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x5d,0xd4

# ATT:   vminpbf16 %xmm4, %xmm3, %xmm2
# INTEL: vminpbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x5d,0xd4

# ATT:   vminpbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vminpbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x5d,0xd4

# ATT:   vminpbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vminpbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x5d,0xd4

# ATT:   vminpbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vminpbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x5d,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vminpbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vminpbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x5d,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vminpbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vminpbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x5d,0x10

# ATT:   vminpbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vminpbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x5d,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vminpbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vminpbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x5d,0x51,0x7f

# ATT:   vminpbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vminpbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x5d,0x52,0x80

# ATT:   vminpbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vminpbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x5d,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vminpbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vminpbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x5d,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vminpbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vminpbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x5d,0x10

# ATT:   vminpbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vminpbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x5d,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vminpbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vminpbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x5d,0x51,0x7f

# ATT:   vminpbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vminpbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x5d,0x52,0x80

# ATT:   vminpbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vminpbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x5d,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vminpbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vminpbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x5d,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vminpbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vminpbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x5d,0x10

# ATT:   vminpbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vminpbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x5d,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vminpbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vminpbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x5d,0x51,0x7f

# ATT:   vminpbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vminpbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x5d,0x52,0x80

# ATT:   vmulnepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vmulnepbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x59,0xd4

# ATT:   vmulnepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vmulnepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x59,0xd4

# ATT:   vmulnepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vmulnepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x59,0xd4

# ATT:   vmulnepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vmulnepbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x59,0xd4

# ATT:   vmulnepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vmulnepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x59,0xd4

# ATT:   vmulnepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vmulnepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x59,0xd4

# ATT:   vmulnepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vmulnepbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x59,0xd4

# ATT:   vmulnepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vmulnepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x59,0xd4

# ATT:   vmulnepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vmulnepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x59,0xd4

# ATT:   vmulnepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vmulnepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x59,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmulnepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vmulnepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x59,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmulnepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vmulnepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x59,0x10

# ATT:   vmulnepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vmulnepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x59,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vmulnepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vmulnepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x59,0x51,0x7f

# ATT:   vmulnepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vmulnepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x59,0x52,0x80

# ATT:   vmulnepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vmulnepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x59,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmulnepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vmulnepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x59,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmulnepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vmulnepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x59,0x10

# ATT:   vmulnepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vmulnepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x59,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vmulnepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vmulnepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x59,0x51,0x7f

# ATT:   vmulnepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vmulnepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x59,0x52,0x80

# ATT:   vmulnepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vmulnepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x59,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vmulnepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vmulnepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x59,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vmulnepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vmulnepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x59,0x10

# ATT:   vmulnepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vmulnepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x59,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vmulnepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vmulnepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x59,0x51,0x7f

# ATT:   vmulnepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vmulnepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x59,0x52,0x80

# ATT:   vrcppbf16 %xmm3, %xmm2
# INTEL: vrcppbf16 xmm2, xmm3
0x62,0xf6,0x7c,0x08,0x4c,0xd3

# ATT:   vrcppbf16 %xmm3, %xmm2 {%k7}
# INTEL: vrcppbf16 xmm2 {k7}, xmm3
0x62,0xf6,0x7c,0x0f,0x4c,0xd3

# ATT:   vrcppbf16 %xmm3, %xmm2 {%k7} {z}
# INTEL: vrcppbf16 xmm2 {k7} {z}, xmm3
0x62,0xf6,0x7c,0x8f,0x4c,0xd3

# ATT:   vrcppbf16 %zmm3, %zmm2
# INTEL: vrcppbf16 zmm2, zmm3
0x62,0xf6,0x7c,0x48,0x4c,0xd3

# ATT:   vrcppbf16 %zmm3, %zmm2 {%k7}
# INTEL: vrcppbf16 zmm2 {k7}, zmm3
0x62,0xf6,0x7c,0x4f,0x4c,0xd3

# ATT:   vrcppbf16 %zmm3, %zmm2 {%k7} {z}
# INTEL: vrcppbf16 zmm2 {k7} {z}, zmm3
0x62,0xf6,0x7c,0xcf,0x4c,0xd3

# ATT:   vrcppbf16 %ymm3, %ymm2
# INTEL: vrcppbf16 ymm2, ymm3
0x62,0xf6,0x7c,0x28,0x4c,0xd3

# ATT:   vrcppbf16 %ymm3, %ymm2 {%k7}
# INTEL: vrcppbf16 ymm2 {k7}, ymm3
0x62,0xf6,0x7c,0x2f,0x4c,0xd3

# ATT:   vrcppbf16 %ymm3, %ymm2 {%k7} {z}
# INTEL: vrcppbf16 ymm2 {k7} {z}, ymm3
0x62,0xf6,0x7c,0xaf,0x4c,0xd3

# ATT:   vrcppbf16  268435456(%esp,%esi,8), %xmm2
# INTEL: vrcppbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x08,0x4c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrcppbf16  291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vrcppbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x0f,0x4c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrcppbf16  (%eax){1to8}, %xmm2
# INTEL: vrcppbf16 xmm2, word ptr [eax]{1to8}
0x62,0xf6,0x7c,0x18,0x4c,0x10

# ATT:   vrcppbf16  -512(,%ebp,2), %xmm2
# INTEL: vrcppbf16 xmm2, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x7c,0x08,0x4c,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vrcppbf16  2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vrcppbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032]
0x62,0xf6,0x7c,0x8f,0x4c,0x51,0x7f

# ATT:   vrcppbf16  -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vrcppbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}
0x62,0xf6,0x7c,0x9f,0x4c,0x52,0x80

# ATT:   vrcppbf16  268435456(%esp,%esi,8), %ymm2
# INTEL: vrcppbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x28,0x4c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrcppbf16  291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vrcppbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x2f,0x4c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrcppbf16  (%eax){1to16}, %ymm2
# INTEL: vrcppbf16 ymm2, word ptr [eax]{1to16}
0x62,0xf6,0x7c,0x38,0x4c,0x10

# ATT:   vrcppbf16  -1024(,%ebp,2), %ymm2
# INTEL: vrcppbf16 ymm2, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x7c,0x28,0x4c,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vrcppbf16  4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vrcppbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064]
0x62,0xf6,0x7c,0xaf,0x4c,0x51,0x7f

# ATT:   vrcppbf16  -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vrcppbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}
0x62,0xf6,0x7c,0xbf,0x4c,0x52,0x80

# ATT:   vrcppbf16  268435456(%esp,%esi,8), %zmm2
# INTEL: vrcppbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x48,0x4c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrcppbf16  291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vrcppbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x4f,0x4c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrcppbf16  (%eax){1to32}, %zmm2
# INTEL: vrcppbf16 zmm2, word ptr [eax]{1to32}
0x62,0xf6,0x7c,0x58,0x4c,0x10

# ATT:   vrcppbf16  -2048(,%ebp,2), %zmm2
# INTEL: vrcppbf16 zmm2, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x7c,0x48,0x4c,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vrcppbf16  8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vrcppbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128]
0x62,0xf6,0x7c,0xcf,0x4c,0x51,0x7f

# ATT:   vrcppbf16  -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vrcppbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}
0x62,0xf6,0x7c,0xdf,0x4c,0x52,0x80

# ATT:   vreducenepbf16 $123, %zmm3, %zmm2
# INTEL: vreducenepbf16 zmm2, zmm3, 123
0x62,0xf3,0x7f,0x48,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %zmm3, %zmm2 {%k7}
# INTEL: vreducenepbf16 zmm2 {k7}, zmm3, 123
0x62,0xf3,0x7f,0x4f,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %zmm3, %zmm2 {%k7} {z}
# INTEL: vreducenepbf16 zmm2 {k7} {z}, zmm3, 123
0x62,0xf3,0x7f,0xcf,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %ymm3, %ymm2
# INTEL: vreducenepbf16 ymm2, ymm3, 123
0x62,0xf3,0x7f,0x28,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %ymm3, %ymm2 {%k7}
# INTEL: vreducenepbf16 ymm2 {k7}, ymm3, 123
0x62,0xf3,0x7f,0x2f,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %ymm3, %ymm2 {%k7} {z}
# INTEL: vreducenepbf16 ymm2 {k7} {z}, ymm3, 123
0x62,0xf3,0x7f,0xaf,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %xmm3, %xmm2
# INTEL: vreducenepbf16 xmm2, xmm3, 123
0x62,0xf3,0x7f,0x08,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %xmm3, %xmm2 {%k7}
# INTEL: vreducenepbf16 xmm2 {k7}, xmm3, 123
0x62,0xf3,0x7f,0x0f,0x56,0xd3,0x7b

# ATT:   vreducenepbf16 $123, %xmm3, %xmm2 {%k7} {z}
# INTEL: vreducenepbf16 xmm2 {k7} {z}, xmm3, 123
0x62,0xf3,0x7f,0x8f,0x56,0xd3,0x7b

# ATT:   vreducenepbf16  $123, 268435456(%esp,%esi,8), %xmm2
# INTEL: vreducenepbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x08,0x56,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vreducenepbf16  $123, 291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vreducenepbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x0f,0x56,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vreducenepbf16  $123, (%eax){1to8}, %xmm2
# INTEL: vreducenepbf16 xmm2, word ptr [eax]{1to8}, 123
0x62,0xf3,0x7f,0x18,0x56,0x10,0x7b

# ATT:   vreducenepbf16  $123, -512(,%ebp,2), %xmm2
# INTEL: vreducenepbf16 xmm2, xmmword ptr [2*ebp - 512], 123
0x62,0xf3,0x7f,0x08,0x56,0x14,0x6d,0x00,0xfe,0xff,0xff,0x7b

# ATT:   vreducenepbf16  $123, 2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vreducenepbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032], 123
0x62,0xf3,0x7f,0x8f,0x56,0x51,0x7f,0x7b

# ATT:   vreducenepbf16  $123, -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vreducenepbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}, 123
0x62,0xf3,0x7f,0x9f,0x56,0x52,0x80,0x7b

# ATT:   vreducenepbf16  $123, 268435456(%esp,%esi,8), %ymm2
# INTEL: vreducenepbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x28,0x56,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vreducenepbf16  $123, 291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vreducenepbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x2f,0x56,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vreducenepbf16  $123, (%eax){1to16}, %ymm2
# INTEL: vreducenepbf16 ymm2, word ptr [eax]{1to16}, 123
0x62,0xf3,0x7f,0x38,0x56,0x10,0x7b

# ATT:   vreducenepbf16  $123, -1024(,%ebp,2), %ymm2
# INTEL: vreducenepbf16 ymm2, ymmword ptr [2*ebp - 1024], 123
0x62,0xf3,0x7f,0x28,0x56,0x14,0x6d,0x00,0xfc,0xff,0xff,0x7b

# ATT:   vreducenepbf16  $123, 4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vreducenepbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064], 123
0x62,0xf3,0x7f,0xaf,0x56,0x51,0x7f,0x7b

# ATT:   vreducenepbf16  $123, -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vreducenepbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}, 123
0x62,0xf3,0x7f,0xbf,0x56,0x52,0x80,0x7b

# ATT:   vreducenepbf16  $123, 268435456(%esp,%esi,8), %zmm2
# INTEL: vreducenepbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x48,0x56,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vreducenepbf16  $123, 291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vreducenepbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x4f,0x56,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vreducenepbf16  $123, (%eax){1to32}, %zmm2
# INTEL: vreducenepbf16 zmm2, word ptr [eax]{1to32}, 123
0x62,0xf3,0x7f,0x58,0x56,0x10,0x7b

# ATT:   vreducenepbf16  $123, -2048(,%ebp,2), %zmm2
# INTEL: vreducenepbf16 zmm2, zmmword ptr [2*ebp - 2048], 123
0x62,0xf3,0x7f,0x48,0x56,0x14,0x6d,0x00,0xf8,0xff,0xff,0x7b

# ATT:   vreducenepbf16  $123, 8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vreducenepbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128], 123
0x62,0xf3,0x7f,0xcf,0x56,0x51,0x7f,0x7b

# ATT:   vreducenepbf16  $123, -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vreducenepbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}, 123
0x62,0xf3,0x7f,0xdf,0x56,0x52,0x80,0x7b

# ATT:   vrndscalenepbf16 $123, %zmm3, %zmm2
# INTEL: vrndscalenepbf16 zmm2, zmm3, 123
0x62,0xf3,0x7f,0x48,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %zmm3, %zmm2 {%k7}
# INTEL: vrndscalenepbf16 zmm2 {k7}, zmm3, 123
0x62,0xf3,0x7f,0x4f,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %zmm3, %zmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 zmm2 {k7} {z}, zmm3, 123
0x62,0xf3,0x7f,0xcf,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %ymm3, %ymm2
# INTEL: vrndscalenepbf16 ymm2, ymm3, 123
0x62,0xf3,0x7f,0x28,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %ymm3, %ymm2 {%k7}
# INTEL: vrndscalenepbf16 ymm2 {k7}, ymm3, 123
0x62,0xf3,0x7f,0x2f,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %ymm3, %ymm2 {%k7} {z}
# INTEL: vrndscalenepbf16 ymm2 {k7} {z}, ymm3, 123
0x62,0xf3,0x7f,0xaf,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %xmm3, %xmm2
# INTEL: vrndscalenepbf16 xmm2, xmm3, 123
0x62,0xf3,0x7f,0x08,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %xmm3, %xmm2 {%k7}
# INTEL: vrndscalenepbf16 xmm2 {k7}, xmm3, 123
0x62,0xf3,0x7f,0x0f,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16 $123, %xmm3, %xmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 xmm2 {k7} {z}, xmm3, 123
0x62,0xf3,0x7f,0x8f,0x08,0xd3,0x7b

# ATT:   vrndscalenepbf16  $123, 268435456(%esp,%esi,8), %xmm2
# INTEL: vrndscalenepbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x08,0x08,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, 291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vrndscalenepbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x0f,0x08,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vrndscalenepbf16  $123, (%eax){1to8}, %xmm2
# INTEL: vrndscalenepbf16 xmm2, word ptr [eax]{1to8}, 123
0x62,0xf3,0x7f,0x18,0x08,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, -512(,%ebp,2), %xmm2
# INTEL: vrndscalenepbf16 xmm2, xmmword ptr [2*ebp - 512], 123
0x62,0xf3,0x7f,0x08,0x08,0x14,0x6d,0x00,0xfe,0xff,0xff,0x7b

# ATT:   vrndscalenepbf16  $123, 2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032], 123
0x62,0xf3,0x7f,0x8f,0x08,0x51,0x7f,0x7b

# ATT:   vrndscalenepbf16  $123, -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}, 123
0x62,0xf3,0x7f,0x9f,0x08,0x52,0x80,0x7b

# ATT:   vrndscalenepbf16  $123, 268435456(%esp,%esi,8), %ymm2
# INTEL: vrndscalenepbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x28,0x08,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, 291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vrndscalenepbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x2f,0x08,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vrndscalenepbf16  $123, (%eax){1to16}, %ymm2
# INTEL: vrndscalenepbf16 ymm2, word ptr [eax]{1to16}, 123
0x62,0xf3,0x7f,0x38,0x08,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, -1024(,%ebp,2), %ymm2
# INTEL: vrndscalenepbf16 ymm2, ymmword ptr [2*ebp - 1024], 123
0x62,0xf3,0x7f,0x28,0x08,0x14,0x6d,0x00,0xfc,0xff,0xff,0x7b

# ATT:   vrndscalenepbf16  $123, 4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vrndscalenepbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064], 123
0x62,0xf3,0x7f,0xaf,0x08,0x51,0x7f,0x7b

# ATT:   vrndscalenepbf16  $123, -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vrndscalenepbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}, 123
0x62,0xf3,0x7f,0xbf,0x08,0x52,0x80,0x7b

# ATT:   vrndscalenepbf16  $123, 268435456(%esp,%esi,8), %zmm2
# INTEL: vrndscalenepbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456], 123
0x62,0xf3,0x7f,0x48,0x08,0x94,0xf4,0x00,0x00,0x00,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, 291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vrndscalenepbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291], 123
0x62,0xf3,0x7f,0x4f,0x08,0x94,0x87,0x23,0x01,0x00,0x00,0x7b

# ATT:   vrndscalenepbf16  $123, (%eax){1to32}, %zmm2
# INTEL: vrndscalenepbf16 zmm2, word ptr [eax]{1to32}, 123
0x62,0xf3,0x7f,0x58,0x08,0x10,0x7b

# ATT:   vrndscalenepbf16  $123, -2048(,%ebp,2), %zmm2
# INTEL: vrndscalenepbf16 zmm2, zmmword ptr [2*ebp - 2048], 123
0x62,0xf3,0x7f,0x48,0x08,0x14,0x6d,0x00,0xf8,0xff,0xff,0x7b

# ATT:   vrndscalenepbf16  $123, 8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128], 123
0x62,0xf3,0x7f,0xcf,0x08,0x51,0x7f,0x7b

# ATT:   vrndscalenepbf16  $123, -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vrndscalenepbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}, 123
0x62,0xf3,0x7f,0xdf,0x08,0x52,0x80,0x7b

# ATT:   vrsqrtpbf16 %xmm3, %xmm2
# INTEL: vrsqrtpbf16 xmm2, xmm3
0x62,0xf6,0x7c,0x08,0x4e,0xd3

# ATT:   vrsqrtpbf16 %xmm3, %xmm2 {%k7}
# INTEL: vrsqrtpbf16 xmm2 {k7}, xmm3
0x62,0xf6,0x7c,0x0f,0x4e,0xd3

# ATT:   vrsqrtpbf16 %xmm3, %xmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 xmm2 {k7} {z}, xmm3
0x62,0xf6,0x7c,0x8f,0x4e,0xd3

# ATT:   vrsqrtpbf16 %zmm3, %zmm2
# INTEL: vrsqrtpbf16 zmm2, zmm3
0x62,0xf6,0x7c,0x48,0x4e,0xd3

# ATT:   vrsqrtpbf16 %zmm3, %zmm2 {%k7}
# INTEL: vrsqrtpbf16 zmm2 {k7}, zmm3
0x62,0xf6,0x7c,0x4f,0x4e,0xd3

# ATT:   vrsqrtpbf16 %zmm3, %zmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 zmm2 {k7} {z}, zmm3
0x62,0xf6,0x7c,0xcf,0x4e,0xd3

# ATT:   vrsqrtpbf16 %ymm3, %ymm2
# INTEL: vrsqrtpbf16 ymm2, ymm3
0x62,0xf6,0x7c,0x28,0x4e,0xd3

# ATT:   vrsqrtpbf16 %ymm3, %ymm2 {%k7}
# INTEL: vrsqrtpbf16 ymm2 {k7}, ymm3
0x62,0xf6,0x7c,0x2f,0x4e,0xd3

# ATT:   vrsqrtpbf16 %ymm3, %ymm2 {%k7} {z}
# INTEL: vrsqrtpbf16 ymm2 {k7} {z}, ymm3
0x62,0xf6,0x7c,0xaf,0x4e,0xd3

# ATT:   vrsqrtpbf16  268435456(%esp,%esi,8), %xmm2
# INTEL: vrsqrtpbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x08,0x4e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrsqrtpbf16  291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vrsqrtpbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x0f,0x4e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrsqrtpbf16  (%eax){1to8}, %xmm2
# INTEL: vrsqrtpbf16 xmm2, word ptr [eax]{1to8}
0x62,0xf6,0x7c,0x18,0x4e,0x10

# ATT:   vrsqrtpbf16  -512(,%ebp,2), %xmm2
# INTEL: vrsqrtpbf16 xmm2, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x7c,0x08,0x4e,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vrsqrtpbf16  2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032]
0x62,0xf6,0x7c,0x8f,0x4e,0x51,0x7f

# ATT:   vrsqrtpbf16  -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}
0x62,0xf6,0x7c,0x9f,0x4e,0x52,0x80

# ATT:   vrsqrtpbf16  268435456(%esp,%esi,8), %ymm2
# INTEL: vrsqrtpbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x28,0x4e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrsqrtpbf16  291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vrsqrtpbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x2f,0x4e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrsqrtpbf16  (%eax){1to16}, %ymm2
# INTEL: vrsqrtpbf16 ymm2, word ptr [eax]{1to16}
0x62,0xf6,0x7c,0x38,0x4e,0x10

# ATT:   vrsqrtpbf16  -1024(,%ebp,2), %ymm2
# INTEL: vrsqrtpbf16 ymm2, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x7c,0x28,0x4e,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vrsqrtpbf16  4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vrsqrtpbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064]
0x62,0xf6,0x7c,0xaf,0x4e,0x51,0x7f

# ATT:   vrsqrtpbf16  -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vrsqrtpbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}
0x62,0xf6,0x7c,0xbf,0x4e,0x52,0x80

# ATT:   vrsqrtpbf16  268435456(%esp,%esi,8), %zmm2
# INTEL: vrsqrtpbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x7c,0x48,0x4e,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vrsqrtpbf16  291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vrsqrtpbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x7c,0x4f,0x4e,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vrsqrtpbf16  (%eax){1to32}, %zmm2
# INTEL: vrsqrtpbf16 zmm2, word ptr [eax]{1to32}
0x62,0xf6,0x7c,0x58,0x4e,0x10

# ATT:   vrsqrtpbf16  -2048(,%ebp,2), %zmm2
# INTEL: vrsqrtpbf16 zmm2, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x7c,0x48,0x4e,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vrsqrtpbf16  8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128]
0x62,0xf6,0x7c,0xcf,0x4e,0x51,0x7f

# ATT:   vrsqrtpbf16  -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vrsqrtpbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}
0x62,0xf6,0x7c,0xdf,0x4e,0x52,0x80

# ATT:   vscalefpbf16 %ymm4, %ymm3, %ymm2
# INTEL: vscalefpbf16 ymm2, ymm3, ymm4
0x62,0xf6,0x64,0x28,0x2c,0xd4

# ATT:   vscalefpbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vscalefpbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf6,0x64,0x2f,0x2c,0xd4

# ATT:   vscalefpbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vscalefpbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf6,0x64,0xaf,0x2c,0xd4

# ATT:   vscalefpbf16 %zmm4, %zmm3, %zmm2
# INTEL: vscalefpbf16 zmm2, zmm3, zmm4
0x62,0xf6,0x64,0x48,0x2c,0xd4

# ATT:   vscalefpbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vscalefpbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf6,0x64,0x4f,0x2c,0xd4

# ATT:   vscalefpbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vscalefpbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf6,0x64,0xcf,0x2c,0xd4

# ATT:   vscalefpbf16 %xmm4, %xmm3, %xmm2
# INTEL: vscalefpbf16 xmm2, xmm3, xmm4
0x62,0xf6,0x64,0x08,0x2c,0xd4

# ATT:   vscalefpbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vscalefpbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf6,0x64,0x0f,0x2c,0xd4

# ATT:   vscalefpbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vscalefpbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf6,0x64,0x8f,0x2c,0xd4

# ATT:   vscalefpbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vscalefpbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x48,0x2c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vscalefpbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vscalefpbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x4f,0x2c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vscalefpbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vscalefpbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf6,0x64,0x58,0x2c,0x10

# ATT:   vscalefpbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vscalefpbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf6,0x64,0x48,0x2c,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vscalefpbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vscalefpbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf6,0x64,0xcf,0x2c,0x51,0x7f

# ATT:   vscalefpbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vscalefpbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf6,0x64,0xdf,0x2c,0x52,0x80

# ATT:   vscalefpbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vscalefpbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x28,0x2c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vscalefpbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vscalefpbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x2f,0x2c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vscalefpbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vscalefpbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf6,0x64,0x38,0x2c,0x10

# ATT:   vscalefpbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vscalefpbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf6,0x64,0x28,0x2c,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vscalefpbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vscalefpbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf6,0x64,0xaf,0x2c,0x51,0x7f

# ATT:   vscalefpbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vscalefpbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf6,0x64,0xbf,0x2c,0x52,0x80

# ATT:   vscalefpbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vscalefpbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf6,0x64,0x08,0x2c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vscalefpbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vscalefpbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf6,0x64,0x0f,0x2c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vscalefpbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vscalefpbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf6,0x64,0x18,0x2c,0x10

# ATT:   vscalefpbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vscalefpbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf6,0x64,0x08,0x2c,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vscalefpbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vscalefpbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf6,0x64,0x8f,0x2c,0x51,0x7f

# ATT:   vscalefpbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vscalefpbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf6,0x64,0x9f,0x2c,0x52,0x80

# ATT:   vsqrtnepbf16 %xmm3, %xmm2
# INTEL: vsqrtnepbf16 xmm2, xmm3
0x62,0xf5,0x7d,0x08,0x51,0xd3

# ATT:   vsqrtnepbf16 %xmm3, %xmm2 {%k7}
# INTEL: vsqrtnepbf16 xmm2 {k7}, xmm3
0x62,0xf5,0x7d,0x0f,0x51,0xd3

# ATT:   vsqrtnepbf16 %xmm3, %xmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 xmm2 {k7} {z}, xmm3
0x62,0xf5,0x7d,0x8f,0x51,0xd3

# ATT:   vsqrtnepbf16 %zmm3, %zmm2
# INTEL: vsqrtnepbf16 zmm2, zmm3
0x62,0xf5,0x7d,0x48,0x51,0xd3

# ATT:   vsqrtnepbf16 %zmm3, %zmm2 {%k7}
# INTEL: vsqrtnepbf16 zmm2 {k7}, zmm3
0x62,0xf5,0x7d,0x4f,0x51,0xd3

# ATT:   vsqrtnepbf16 %zmm3, %zmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 zmm2 {k7} {z}, zmm3
0x62,0xf5,0x7d,0xcf,0x51,0xd3

# ATT:   vsqrtnepbf16 %ymm3, %ymm2
# INTEL: vsqrtnepbf16 ymm2, ymm3
0x62,0xf5,0x7d,0x28,0x51,0xd3

# ATT:   vsqrtnepbf16 %ymm3, %ymm2 {%k7}
# INTEL: vsqrtnepbf16 ymm2 {k7}, ymm3
0x62,0xf5,0x7d,0x2f,0x51,0xd3

# ATT:   vsqrtnepbf16 %ymm3, %ymm2 {%k7} {z}
# INTEL: vsqrtnepbf16 ymm2 {k7} {z}, ymm3
0x62,0xf5,0x7d,0xaf,0x51,0xd3

# ATT:   vsqrtnepbf16  268435456(%esp,%esi,8), %xmm2
# INTEL: vsqrtnepbf16 xmm2, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x08,0x51,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsqrtnepbf16  291(%edi,%eax,4), %xmm2 {%k7}
# INTEL: vsqrtnepbf16 xmm2 {k7}, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x0f,0x51,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsqrtnepbf16  (%eax){1to8}, %xmm2
# INTEL: vsqrtnepbf16 xmm2, word ptr [eax]{1to8}
0x62,0xf5,0x7d,0x18,0x51,0x10

# ATT:   vsqrtnepbf16  -512(,%ebp,2), %xmm2
# INTEL: vsqrtnepbf16 xmm2, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x7d,0x08,0x51,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vsqrtnepbf16  2032(%ecx), %xmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 xmm2 {k7} {z}, xmmword ptr [ecx + 2032]
0x62,0xf5,0x7d,0x8f,0x51,0x51,0x7f

# ATT:   vsqrtnepbf16  -256(%edx){1to8}, %xmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 xmm2 {k7} {z}, word ptr [edx - 256]{1to8}
0x62,0xf5,0x7d,0x9f,0x51,0x52,0x80

# ATT:   vsqrtnepbf16  268435456(%esp,%esi,8), %ymm2
# INTEL: vsqrtnepbf16 ymm2, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x28,0x51,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsqrtnepbf16  291(%edi,%eax,4), %ymm2 {%k7}
# INTEL: vsqrtnepbf16 ymm2 {k7}, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x2f,0x51,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsqrtnepbf16  (%eax){1to16}, %ymm2
# INTEL: vsqrtnepbf16 ymm2, word ptr [eax]{1to16}
0x62,0xf5,0x7d,0x38,0x51,0x10

# ATT:   vsqrtnepbf16  -1024(,%ebp,2), %ymm2
# INTEL: vsqrtnepbf16 ymm2, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x7d,0x28,0x51,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vsqrtnepbf16  4064(%ecx), %ymm2 {%k7} {z}
# INTEL: vsqrtnepbf16 ymm2 {k7} {z}, ymmword ptr [ecx + 4064]
0x62,0xf5,0x7d,0xaf,0x51,0x51,0x7f

# ATT:   vsqrtnepbf16  -256(%edx){1to16}, %ymm2 {%k7} {z}
# INTEL: vsqrtnepbf16 ymm2 {k7} {z}, word ptr [edx - 256]{1to16}
0x62,0xf5,0x7d,0xbf,0x51,0x52,0x80

# ATT:   vsqrtnepbf16  268435456(%esp,%esi,8), %zmm2
# INTEL: vsqrtnepbf16 zmm2, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x7d,0x48,0x51,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsqrtnepbf16  291(%edi,%eax,4), %zmm2 {%k7}
# INTEL: vsqrtnepbf16 zmm2 {k7}, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x7d,0x4f,0x51,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsqrtnepbf16  (%eax){1to32}, %zmm2
# INTEL: vsqrtnepbf16 zmm2, word ptr [eax]{1to32}
0x62,0xf5,0x7d,0x58,0x51,0x10

# ATT:   vsqrtnepbf16  -2048(,%ebp,2), %zmm2
# INTEL: vsqrtnepbf16 zmm2, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x7d,0x48,0x51,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vsqrtnepbf16  8128(%ecx), %zmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 zmm2 {k7} {z}, zmmword ptr [ecx + 8128]
0x62,0xf5,0x7d,0xcf,0x51,0x51,0x7f

# ATT:   vsqrtnepbf16  -256(%edx){1to32}, %zmm2 {%k7} {z}
# INTEL: vsqrtnepbf16 zmm2 {k7} {z}, word ptr [edx - 256]{1to32}
0x62,0xf5,0x7d,0xdf,0x51,0x52,0x80

# ATT:   vsubnepbf16 %ymm4, %ymm3, %ymm2
# INTEL: vsubnepbf16 ymm2, ymm3, ymm4
0x62,0xf5,0x65,0x28,0x5c,0xd4

# ATT:   vsubnepbf16 %ymm4, %ymm3, %ymm2 {%k7}
# INTEL: vsubnepbf16 ymm2 {k7}, ymm3, ymm4
0x62,0xf5,0x65,0x2f,0x5c,0xd4

# ATT:   vsubnepbf16 %ymm4, %ymm3, %ymm2 {%k7} {z}
# INTEL: vsubnepbf16 ymm2 {k7} {z}, ymm3, ymm4
0x62,0xf5,0x65,0xaf,0x5c,0xd4

# ATT:   vsubnepbf16 %zmm4, %zmm3, %zmm2
# INTEL: vsubnepbf16 zmm2, zmm3, zmm4
0x62,0xf5,0x65,0x48,0x5c,0xd4

# ATT:   vsubnepbf16 %zmm4, %zmm3, %zmm2 {%k7}
# INTEL: vsubnepbf16 zmm2 {k7}, zmm3, zmm4
0x62,0xf5,0x65,0x4f,0x5c,0xd4

# ATT:   vsubnepbf16 %zmm4, %zmm3, %zmm2 {%k7} {z}
# INTEL: vsubnepbf16 zmm2 {k7} {z}, zmm3, zmm4
0x62,0xf5,0x65,0xcf,0x5c,0xd4

# ATT:   vsubnepbf16 %xmm4, %xmm3, %xmm2
# INTEL: vsubnepbf16 xmm2, xmm3, xmm4
0x62,0xf5,0x65,0x08,0x5c,0xd4

# ATT:   vsubnepbf16 %xmm4, %xmm3, %xmm2 {%k7}
# INTEL: vsubnepbf16 xmm2 {k7}, xmm3, xmm4
0x62,0xf5,0x65,0x0f,0x5c,0xd4

# ATT:   vsubnepbf16 %xmm4, %xmm3, %xmm2 {%k7} {z}
# INTEL: vsubnepbf16 xmm2 {k7} {z}, xmm3, xmm4
0x62,0xf5,0x65,0x8f,0x5c,0xd4

# ATT:   vsubnepbf16  268435456(%esp,%esi,8), %zmm3, %zmm2
# INTEL: vsubnepbf16 zmm2, zmm3, zmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x48,0x5c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsubnepbf16  291(%edi,%eax,4), %zmm3, %zmm2 {%k7}
# INTEL: vsubnepbf16 zmm2 {k7}, zmm3, zmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x4f,0x5c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsubnepbf16  (%eax){1to32}, %zmm3, %zmm2
# INTEL: vsubnepbf16 zmm2, zmm3, word ptr [eax]{1to32}
0x62,0xf5,0x65,0x58,0x5c,0x10

# ATT:   vsubnepbf16  -2048(,%ebp,2), %zmm3, %zmm2
# INTEL: vsubnepbf16 zmm2, zmm3, zmmword ptr [2*ebp - 2048]
0x62,0xf5,0x65,0x48,0x5c,0x14,0x6d,0x00,0xf8,0xff,0xff

# ATT:   vsubnepbf16  8128(%ecx), %zmm3, %zmm2 {%k7} {z}
# INTEL: vsubnepbf16 zmm2 {k7} {z}, zmm3, zmmword ptr [ecx + 8128]
0x62,0xf5,0x65,0xcf,0x5c,0x51,0x7f

# ATT:   vsubnepbf16  -256(%edx){1to32}, %zmm3, %zmm2 {%k7} {z}
# INTEL: vsubnepbf16 zmm2 {k7} {z}, zmm3, word ptr [edx - 256]{1to32}
0x62,0xf5,0x65,0xdf,0x5c,0x52,0x80

# ATT:   vsubnepbf16  268435456(%esp,%esi,8), %ymm3, %ymm2
# INTEL: vsubnepbf16 ymm2, ymm3, ymmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x28,0x5c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsubnepbf16  291(%edi,%eax,4), %ymm3, %ymm2 {%k7}
# INTEL: vsubnepbf16 ymm2 {k7}, ymm3, ymmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x2f,0x5c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsubnepbf16  (%eax){1to16}, %ymm3, %ymm2
# INTEL: vsubnepbf16 ymm2, ymm3, word ptr [eax]{1to16}
0x62,0xf5,0x65,0x38,0x5c,0x10

# ATT:   vsubnepbf16  -1024(,%ebp,2), %ymm3, %ymm2
# INTEL: vsubnepbf16 ymm2, ymm3, ymmword ptr [2*ebp - 1024]
0x62,0xf5,0x65,0x28,0x5c,0x14,0x6d,0x00,0xfc,0xff,0xff

# ATT:   vsubnepbf16  4064(%ecx), %ymm3, %ymm2 {%k7} {z}
# INTEL: vsubnepbf16 ymm2 {k7} {z}, ymm3, ymmword ptr [ecx + 4064]
0x62,0xf5,0x65,0xaf,0x5c,0x51,0x7f

# ATT:   vsubnepbf16  -256(%edx){1to16}, %ymm3, %ymm2 {%k7} {z}
# INTEL: vsubnepbf16 ymm2 {k7} {z}, ymm3, word ptr [edx - 256]{1to16}
0x62,0xf5,0x65,0xbf,0x5c,0x52,0x80

# ATT:   vsubnepbf16  268435456(%esp,%esi,8), %xmm3, %xmm2
# INTEL: vsubnepbf16 xmm2, xmm3, xmmword ptr [esp + 8*esi + 268435456]
0x62,0xf5,0x65,0x08,0x5c,0x94,0xf4,0x00,0x00,0x00,0x10

# ATT:   vsubnepbf16  291(%edi,%eax,4), %xmm3, %xmm2 {%k7}
# INTEL: vsubnepbf16 xmm2 {k7}, xmm3, xmmword ptr [edi + 4*eax + 291]
0x62,0xf5,0x65,0x0f,0x5c,0x94,0x87,0x23,0x01,0x00,0x00

# ATT:   vsubnepbf16  (%eax){1to8}, %xmm3, %xmm2
# INTEL: vsubnepbf16 xmm2, xmm3, word ptr [eax]{1to8}
0x62,0xf5,0x65,0x18,0x5c,0x10

# ATT:   vsubnepbf16  -512(,%ebp,2), %xmm3, %xmm2
# INTEL: vsubnepbf16 xmm2, xmm3, xmmword ptr [2*ebp - 512]
0x62,0xf5,0x65,0x08,0x5c,0x14,0x6d,0x00,0xfe,0xff,0xff

# ATT:   vsubnepbf16  2032(%ecx), %xmm3, %xmm2 {%k7} {z}
# INTEL: vsubnepbf16 xmm2 {k7} {z}, xmm3, xmmword ptr [ecx + 2032]
0x62,0xf5,0x65,0x8f,0x5c,0x51,0x7f

# ATT:   vsubnepbf16  -256(%edx){1to8}, %xmm3, %xmm2 {%k7} {z}
# INTEL: vsubnepbf16 xmm2 {k7} {z}, xmm3, word ptr [edx - 256]{1to8}
0x62,0xf5,0x65,0x9f,0x5c,0x52,0x80
llvm/llvm/test/MC/Disassembler/X86/avx10.2-bf16-32.txt
