[libcxx-commits] [flang] [clang-tools-extra] [lld] [llvm] [compiler-rt] [lldb] [libc] [libcxx] [clang] [AMDGPU] Use alias info to relax waitcounts for LDS DMA (PR #74537)

[Jay Foad via libcxx-commits]

jayfoad wrote:

> > How does this work in a case like this?
> > ```
> > call void @llvm.amdgcn.raw.buffer.load.lds(<4 x i32> %rsrc, ptr addrspace(3) @lds.3, i32 4, i32 0, i32 0, i32 0, i32 0)
> > call void @llvm.amdgcn.raw.buffer.load.lds(<4 x i32> %rsrc, ptr addrspace(3) %ptr, i32 4, i32 0, i32 0, i32 0, i32 0)
> > %val.3 = load float, ptr addrspace(3) @lds.3, align 4
> > ```
> > 
> > 
> >     
> >       
> >     
> > 
> >       
> >     
> > 
> >     
> >   
> > i.e.
> > ```
> > * store to known lds address `@lds.3` (this will use slot 0 and another slot e.g. slot 3?)
> > 
> > * store to unknown lds address (this will use slot 0?)
> > 
> > * load from known lds address `@lds.3` (this will use slot 3?)
> > ```
> 
> It does not know the pointer, so it uses default slot 0 and waits till 0.

Test case:
```
@lds.0 = internal addrspace(3) global [64 x float] poison, align 16
@lds.1 = internal addrspace(3) global [64 x float] poison, align 16

declare void @llvm.amdgcn.raw.buffer.load.lds(<4 x i32> %rsrc, ptr addrspace(3) nocapture, i32 %size, i32 %voffset, i32 %soffset, i32 %offset, i32 %aux)

define amdgpu_kernel void @f(<4 x i32> %rsrc, i32 %i1, i32 %i2, ptr addrspace(1) %out, ptr addrspace(3) %ptr) {
main_body:
  call void @llvm.amdgcn.raw.buffer.load.lds(<4 x i32> %rsrc, ptr addrspace(3) @lds.0, i32 4, i32 0, i32 0, i32 0, i32 0)
  call void @llvm.amdgcn.raw.buffer.load.lds(<4 x i32> %rsrc, ptr addrspace(3) %ptr, i32 4, i32 0, i32 0, i32 0, i32 0)
  %gep.0 = getelementptr float, ptr addrspace(3) @lds.0, i32 %i1
  %gep.1 = getelementptr float, ptr addrspace(3) @lds.1, i32 %i2
  %val.0 = load volatile float, ptr addrspace(3) %gep.0, align 4
  %val.1 = load volatile float, ptr addrspace(3) %gep.1, align 4
  %out.gep.1 = getelementptr float, ptr addrspace(1) %out, i32 1
  store float %val.0, ptr addrspace(1) %out
  store float %val.1, ptr addrspace(1) %out.gep.1
  ret void
}
```
Generates:
```
	s_load_dwordx8 s[4:11], s[0:1], 0x24
	s_load_dword s2, s[0:1], 0x44
	s_mov_b32 m0, 0
	v_mov_b32_e32 v2, 0
	s_waitcnt lgkmcnt(0)
	buffer_load_dword off, s[4:7], 0 lds
	s_mov_b32 m0, s2
	s_lshl_b32 s0, s8, 2
	buffer_load_dword off, s[4:7], 0 lds
	s_lshl_b32 s1, s9, 2
	v_mov_b32_e32 v0, s0
	v_mov_b32_e32 v1, s1
	s_waitcnt vmcnt(1)
	ds_read_b32 v0, v0
	s_waitcnt vmcnt(0)
	ds_read_b32 v1, v1 offset:256
	s_waitcnt lgkmcnt(0)
	global_store_dwordx2 v2, v[0:1], s[10:11]
	s_endpgm
```
The `s_waitcnt vmcnt(1)` seems incorrect, because the second buffer-load-to-lds might clobber `@lds.0`.

> I have to tell anyone interested here: before I even wrote this code it didn't know of the dependency and did not wait for anything at all. Everyone was happy.

I am still happy, because buffer/flat/global-load-to-lds was removed in GFX11.

https://github.com/llvm/llvm-project/pull/74537