[llvm-bugs] [Bug 39323] New: [AMDGPU][MC][DISASSEMBLER] Incorrect MIMG VAddr size for many opcodes

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=39323

            Bug ID: 39323
           Summary: [AMDGPU][MC][DISASSEMBLER] Incorrect MIMG VAddr size
                    for many opcodes
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AMDGPU
          Assignee: unassignedbugs at nondot.org
          Reporter: dpreobrazhensky at luxoft.com
                CC: llvm-bugs at lists.llvm.org

Many MIMG instructions have incorrect VAddr size when disassembled.

For example, according with AMD doc, image_gather4 does not support 1D so at
least 2 address components are required (2D). However the following GFX8 code

    0x00,0x01,0x00,0xf1,0x01,0x05,0x62,0x00

is disassembled to 

    image_gather4 v[5:8], v1, s[8:15], s[12:15] dmask:0x1

The expected result is:

    image_gather4 v[5:8], v[1:2], s[8:15], s[12:15] dmask:0x1

Well, I assume this is because GFX8 reuses hardcoded tables from GFX9 which
supports a16.

However when looking at GFX9 code, the results are the same. Moreover, a16
modifier does not affect VAddr size. For example, the following GFX9 code

    0x00,0x01,0x04,0xf1,0x01,0x05,0x62,0x00
    0x00,0x81,0x04,0xf1,0x01,0x05,0x62,0x00

is disassembled to:

    image_gather4_cl v[5:8], v1, s[8:15], s[12:15] dmask:0x1
    image_gather4_cl v[5:8], v1, s[8:15], s[12:15] dmask:0x1 a16

Note that:
- a16 does not affect VAddr size.
- according with AMD doc, VAddr cannot have less than 3 components for
image_gather4_cl (so at least 2 with a16=1).

Expected results are:

    image_gather4_cl v[5:8], v[1:3], s[8:15], s[12:15] dmask:0x1
    image_gather4_cl v[5:8], v[1:2], s[8:15], s[12:15] dmask:0x1 a16

This issue results in disassembler test failures for all GATHER and most SAMPLE
instructions for both GFX8 and GFX9. Assembler tests mostly pass (there are no
issues with VAddr size).

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20181016/4ee2e1bb/attachment.html>