[AMDGPU] invalid PHI node generated after AMDGPULowerModuleLDS Pass

[hzwangchaochen]

The simplified LLVM IR is as follows:

@lds.size.2.align.2 = internal unnamed_addr addrspace(3) global [6144 x half] poison, align 2
define dso_local amdgpu_kernel void @flash_attentionV1(ptr addrspace(1) %out) {
bb3:
  switch i32 0, label %bb10 [
    i32 18, label %bb7
    i32 1, label %bb7
    i32 0, label %bb10
  ]
bb5:
  switch i32 0, label %bb10 [
    i32 18, label %bb7
    i32 1, label %bb7
    i32 0, label %bb10
  ]

bb7:                                              ; preds = %bb5, %bb5, %bb3, %bb3
  %phi8 = phi i64 [ ptrtoint (ptr addrspacecast (ptr addrspace(3) @lds.size.2.align.2 to ptr) to i64), %bb5 ], [ ptrtoint (ptr addrspacecast (ptr addrspace(3) @lds.size.2.align.2 to ptr) to i64), %bb5 ], [10, %bb3], [10, %bb3]
  store i64 %phi8, ptr addrspace(1) %out, align 8
  br label %bb10

bb10:                                             ; preds = %bb3, %bb3, %bb2
  ret void
}

编译命令如下：

./build/bin/llc -mtriple=amdgcn xxx.ll or ./build/bin/opt -S -mtriple=amdgcn-- -amdgpu-lower-module-ld xxx.ll

After AMDGPULowerModuleLDS Pass
bb5 and bb7 are transformed as follows:

bb5:                                              ; No predecessors!
  %0 = addrspacecast ptr addrspace(3) @llvm.amdgcn.kernel.flash_attentionV1.lds to ptr
  %1 = ptrtoint ptr %0 to i64
  %2 = addrspacecast ptr addrspace(3) @llvm.amdgcn.kernel.flash_attentionV1.lds to ptr
  %3 = ptrtoint ptr %2 to i64
  switch i32 0, label %bb10 [
    i32 18, label %bb7
    i32 1, label %bb7
    i32 0, label %bb10
  ]
bb7:                                              ; preds = %bb5, %bb5, %bb3, %bb3
  %phi8 = phi i64 [ %3, %bb5 ], [ %1, %bb5 ], [ 10, %bb3 ], [ 10, %bb3 ]
  store i64 %phi8, ptr addrspace(1) %out, align 8
  br label %bb10

Then we encountered the following error message:

PHI node has multiple entries for the same basic block with different incoming values!
  %phi8 = phi i64 [ %3, %bb5 ], [ %1, %bb5 ], [ 10, %bb3 ], [ 10, %bb3 ]
label %bb5
  %3 = ptrtoint ptr %2 to i64
  %1 = ptrtoint ptr %0 to i64
in function flash_attentionV1

Any ideas for this issues, thanks a lot~

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: None (hzwangchaochen)

The simplified LLVM IR is as follows: ``` @lds.size.2.align.2 = internal unnamed_addr addrspace(3) global [6144 x half] poison, align 2 define dso_local amdgpu_kernel void @flash_attentionV1(ptr addrspace(1) %out) { bb3: switch i32 0, label %bb10 [ i32 18, label %bb7 i32 1, label %bb7 i32 0, label %bb10 ] bb5: switch i32 0, label %bb10 [ i32 18, label %bb7 i32 1, label %bb7 i32 0, label %bb10 ]

bb7: ; preds = %bb5, %bb5, %bb3, %bb3
%phi8 = phi i64 [ ptrtoint (ptr addrspacecast (ptr addrspace(3) @lds.size.2.align.2 to ptr) to i64), %bb5 ], [ ptrtoint (ptr addrspacecast (ptr addrspace(3) @lds.size.2.align.2 to ptr) to i64), %bb5 ], [10, %bb3], [10, %bb3]
store i64 %phi8, ptr addrspace(1) %out, align 8
br label %bb10

bb10: ; preds = %bb3, %bb3, %bb2
ret void
}

编译命令如下：

./build/bin/llc -mtriple=amdgcn xxx.ll or ./build/bin/opt -S -mtriple=amdgcn-- -amdgpu-lower-module-ld xxx.ll

After AMDGPULowerModuleLDS Pass
bb5 and bb7 are transformed as follows:

bb5: ; No predecessors!
%0 = addrspacecast ptr addrspace(3) @llvm.amdgcn.kernel.flash_attentionV1.lds to ptr
%1 = ptrtoint ptr %0 to i64
%2 = addrspacecast ptr addrspace(3) @llvm.amdgcn.kernel.flash_attentionV1.lds to ptr
%3 = ptrtoint ptr %2 to i64
switch i32 0, label %bb10 [
i32 18, label %bb7
i32 1, label %bb7
i32 0, label %bb10
]
bb7: ; preds = %bb5, %bb5, %bb3, %bb3
%phi8 = phi i64 [ %3, %bb5 ], [ %1, %bb5 ], [ 10, %bb3 ], [ 10, %bb3 ]
store i64 %phi8, ptr addrspace(1) %out, align 8
br label %bb10

Then we encountered the following error message:

PHI node has multiple entries for the same basic block with different incoming values!
%phi8 = phi i64 [ %3, %bb5 ], [ %1, %bb5 ], [ 10, %bb3 ], [ 10, %bb3 ]
label %bb5
%3 = ptrtoint ptr %2 to i64
%1 = ptrtoint ptr %0 to i64
in function flash_attentionV1

Any ideas for this issues, thanks a lot~
</details>

[jayfoad]

I would say it's a problem with convertUsersOfConstantsToInstructions. When there are multiple uses of the same ConstantExpr in the same Function, it should only convert it to instructions once. In this case %2 and %3 are identical to %0 and %1. So it should have reused %1 instead of doing the conversion a second time.

[hzwangchaochen]

I would say it's a problem with convertUsersOfConstantsToInstructions. When there are multiple uses of the same ConstantExpr in the same Function, it should only convert it to instructions once. In this case %2 and %3 are identical to %0 and %1. So it should have reused %1 instead of doing the conversion a second time.

Thanks a lot, I'm wondering if anyone will fix this issue soon, or if we should look into modifying it ourselves.

[JonChesterfield]

Thanks for the bug report! If the amd crowd don't get to this I'll pick it up, or happy to review your PR if preferred

[hzwangchaochen]

Thanks for the bug report! If the amd crowd don't get to this I'll pick it up, or happy to review your PR if preferred

If you have a suitable solution, feel free to pick it up.