AMDGPU: Fix v_dot{4, 8}* instruction encoding

Differential Revision: https://reviews.llvm.org/D46848

llvm-svn: 332387

1 files changed, 308 insertions, 8 deletions

diff --git a/llvm/test/MC/Disassembler/AMDGPU/dl-insts.txt b/llvm/test/MC/Disassembler/AMDGPU/dl-insts.txt
index 058f3ce3eaa..38ddd1f327e 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/dl-insts.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/dl-insts.txt
@@ -546,17 +546,17 @@
 # CHECK: v_dot2_u32_u16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa7,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0x40,0xa7,0xd3,0x01,0x05,0x0e,0x1c
 
-# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04]
-0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xa8,0xd3,0x01,0x05,0x0e,0x1c
 
-# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04]
-0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xa9,0xd3,0x01,0x05,0x0e,0x1c
 
-# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04]
-0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xaa,0xd3,0x01,0x05,0x0e,0x1c
 
-# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04]
-0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xab,0xd3,0x01,0x05,0x0e,0x1c
 
 # CHECK: v_dot2_f32_f16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa3,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0x40,0xa3,0xd3,0x01,0x05,0x0e,0x1c
@@ -774,6 +774,294 @@
 # CHECK: v_dot2_u32_u16 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x18,0xa7,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0x18,0xa7,0xd3,0x01,0x05,0x0e,0x1c
 
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[0,1,0] ; encoding: [0x00,0x50,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x50,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,0,0] ; encoding: [0x00,0x48,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x48,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,1,0] ; encoding: [0x00,0x58,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x58,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x10,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x08,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x18,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[0,1,0] ; encoding: [0x00,0x50,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x50,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,0,0] ; encoding: [0x00,0x48,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x48,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1,0] ; encoding: [0x00,0x58,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x58,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x10,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x08,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x18,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[0,1,0] ; encoding: [0x00,0x50,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x50,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,0,0] ; encoding: [0x00,0x48,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x48,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,1,0] ; encoding: [0x00,0x58,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x58,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x10,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x08,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x18,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x40,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[0,1,0] ; encoding: [0x00,0x50,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x50,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,0,0] ; encoding: [0x00,0x48,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x48,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,1,0] ; encoding: [0x00,0x58,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x58,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[0,0,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[0,1,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[1,0,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel_hi:[1,1,0] ; encoding: [0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x00,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[0,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x10,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,0,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x08,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,0,0] ; encoding: [0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x04]
+0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x04
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[0,1,0] ; encoding: [0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x14]
+0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x14
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,0,0] ; encoding: [0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x0c]
+0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x0c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0] ; encoding: [0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0x18,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
 # CHECK: v_dot2_f32_f16 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xa3,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0xc0,0xa3,0xd3,0x01,0x05,0x0e,0x1c
 
@@ -783,6 +1071,18 @@
 # CHECK: v_dot2_u32_u16 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xa7,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0xc0,0xa7,0xd3,0x01,0x05,0x0e,0x1c
 
+# CHECK: v_dot4_i32_i8 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xa8,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0xc0,0xa8,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot4_u32_u8 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xa9,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0xc0,0xa9,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_i32_i4 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xaa,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0xc0,0xaa,0xd3,0x01,0x05,0x0e,0x1c
+
+# CHECK: v_dot8_u32_u4 v0, v1, v2, v3 clamp ; encoding: [0x00,0xc0,0xab,0xd3,0x01,0x05,0x0e,0x1c]
+0x00,0xc0,0xab,0xd3,0x01,0x05,0x0e,0x1c
+
 # CHECK: v_dot2_f32_f16 v0, v1, v2, v3 ; encoding: [0x00,0x40,0xa3,0xd3,0x01,0x05,0x0e,0x1c]
 0x00,0x40,0xa3,0xd3,0x01,0x05,0x0e,0x1c