[llvm-bugs] [Bug 51720] New: [AMDGPU][GFX10] MIMG opcodes with bias and a16 have incorrect address size

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51720

            Bug ID: 51720
           Summary: [AMDGPU][GFX10] MIMG opcodes with bias and a16 have
                    incorrect address size
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: AMDGPU
          Assignee: unassignedbugs at nondot.org
          Reporter: d-pre at mail.ru
                CC: llvm-bugs at lists.llvm.org

According with AMD documentation, bias size is affected by a16. However AMDGPU
assembler does not correct bias size which results in test failures for
GATHER_B* and SAMPLER_B* opcodes.

For example, the following instruction may be assembled using sp3 but not llvm
AMDGPU assembler:

  image_gather4_b v[5:8], v[1:2], s[8:15], s[12:15] dmask:0x8
dim:SQ_RSRC_IMG_CUBE a16

AMDGPU assembler expects 96-bit address size:

  first dword:    bias
  second dword:   t, s
  third dword:    0, face

Actual address in this case should be as follows:

  first dword:    s, bias
  second dword:   face, t

Current implementation computes address size as shown here (see
https://github.com/llvm/llvm-project/blob/e5438f386854136d848989315f53788808afa37a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp#L158):

  unsigned AddrWords = BaseOpcode->NumExtraArgs;
  unsigned AddrComponents = ...
  if (IsA16)
    AddrWords += divideCeil(AddrComponents, 2);
  else
    AddrWords += AddrComponents;

In other words, NumExtraArgs accounts for offset, bias and z-compare (but not
derivatives which are handled elsewhere). So currently bias size is not
affected by a16.

Note that bias size may affect address size under the following conditions:
  - a16=1.
  - opcode has no z-compare (it follows bias and always 32-bit aligned).
  - opcode has no derivatives (it follows z-comp and also 32-bit aligned).

It looks like NumExtraArgs should be split into 3 separate fields (offset,
bias, zcomp) to account for bias size variations.

-- 
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/20210902/3e60bd73/attachment.html>