[AMDGPU][SILoadStoreOptimizer] Fix lds address operand offset (#176816)
The offset operand in GLOBAL_LOAD_ASYNC_TO_LDS_B128, for instance, is
added to both the lds and global address, but SILoadStoreOptimizer is
currently unaware of that. This PR inserts an add to counteract the
offset meant for the global address. This one add is better than not
doing the optimization at all, and having to insert 2 adds for each
global address calculation (with no offset).
```
; ENABLE-LABEL: name: promote_async_load_offset
; ENABLE: liveins: $ttmp7, $vgpr0, $sgpr0_sgpr1
; ENABLE-NEXT: {{ $}}
; ENABLE-NEXT: renamable $vgpr1 = V_LSHLREV_B32_e32 8, $vgpr0, implicit $exec
; ENABLE-NEXT: renamable $vgpr2, renamable $vcc_lo = V_ADD_CO_U32_e64 $vgpr0, 512, 0, implicit $exec
; ENABLE-NEXT: renamable $vgpr3, dead $sgpr_null = V_ADDC_U32_e64 0, killed $vgpr0, killed $vcc_lo, 0, implicit $exec
; ENABLE-NEXT: renamable $vgpr1 = disjoint V_OR_B32_e32 0, killed $vgpr1, implicit $exec
; ENABLE-NEXT: renamable $vgpr0 = V_ADD_U32_e32 256, $vgpr1, implicit $exec
; ENABLE-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B128 killed $vgpr0, $vgpr2_vgpr3, -256, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt :: (load store (s128), align 1, addrspace 3)
; ENABLE-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B128 killed $vgpr1, killed $vgpr2_vgpr3, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt :: (load store (s128), align 1, addrspace 3)
; DISABLE-LABEL: name: promote_async_load_offset
; DISABLE: liveins: $ttmp7, $vgpr0, $sgpr0_sgpr1
; DISABLE-NEXT: {{ $}}
; DISABLE-NEXT: renamable $vgpr1 = V_LSHLREV_B32_e32 8, $vgpr0, implicit $exec
; DISABLE-NEXT: renamable $vgpr2, renamable $vcc_lo = V_ADD_CO_U32_e64 256, $vgpr0, 0, implicit $exec
; DISABLE-NEXT: renamable $vgpr3, $sgpr_null = V_ADDC_U32_e64 0, $vgpr0, killed $vcc_lo, 0, implicit $exec
; DISABLE-NEXT: renamable $vgpr1 = disjoint V_OR_B32_e32 0, killed $vgpr1, implicit $exec
; DISABLE-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B128 $vgpr1, killed $vgpr2_vgpr3, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt :: (load store (s128), align 1, addrspace 3)
; DISABLE-NEXT: renamable $vgpr2, renamable $vcc_lo = V_ADD_CO_U32_e64 512, $vgpr0, 0, implicit $exec
; DISABLE-NEXT: renamable $vgpr3, $sgpr_null = V_ADDC_U32_e64 0, killed $vgpr0, killed $vcc_lo, 0, implicit $exec
; DISABLE-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B128 killed $vgpr1, killed $vgpr2_vgpr3, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt :: (load store (s128), align 1, addrspace 3)
```
This PR also promotes the global address to an offset when the offset is
calculated with V_ADD_U64 on applicable gfx versions, (and inversely
adds the LDS offset), whereas previously the optimization opportunity
was missed entirely.
Welcome to the LLVM project!
This repository contains the source code for LLVM, a toolkit for the construction of highly optimized compilers, optimizers, and run-time environments.
The LLVM project has multiple components. The core of the project is itself called “LLVM”. This contains all of the tools, libraries, and header files needed to process intermediate representations and convert them into object files. Tools include an assembler, disassembler, bitcode analyzer, and bitcode optimizer.
C-like languages use the Clang frontend. This component compiles C, C++, Objective-C, and Objective-C++ code into LLVM bitcode -- and from there into object files, using LLVM.
Other components include: the libc++ C++ standard library, the LLD linker, and more.
Consult the Getting Started with LLVM page for information on building and running LLVM.
For information on how to contribute to the LLVM project, please take a look at the Contributing to LLVM guide.
Join the LLVM Discourse forums, Discord chat, LLVM Office Hours or Regular sync-ups.
The LLVM project has adopted a code of conduct for participants to all modes of communication within the project.