[llvm] [AMDGPU] Support FLAT_SCRATCH in SILoadStoreOptimizer const offset promotion (PR #212507)

[Valery Pykhtin via llvm-commits]

================
@@ -21,18 +21,16 @@ define amdgpu_kernel void @soff1_voff1(i32 %soff) {
 ; GFX942-SDAG-NEXT:    s_load_dword s0, s[4:5], 0x24
 ; GFX942-SDAG-NEXT:    v_and_b32_e32 v0, 0x3ff, v0
 ; GFX942-SDAG-NEXT:    v_mov_b32_e32 v1, 1
+; GFX942-SDAG-NEXT:    v_mov_b32_e32 v2, 2
 ; GFX942-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX942-SDAG-NEXT:    v_add_u32_e32 v0, s0, v0
-; GFX942-SDAG-NEXT:    v_add_u32_e32 v2, 1, v0
-; GFX942-SDAG-NEXT:    v_add_u32_e32 v3, 2, v0
-; GFX942-SDAG-NEXT:    scratch_store_byte v2, v1, off sc0 sc1
+; GFX942-SDAG-NEXT:    v_add_u32_e32 v0, 1, v0
+; GFX942-SDAG-NEXT:    scratch_store_byte v0, v1, off sc0 sc1
----------------
vpykhtin wrote:

Not sure if this belongs to this patch but it choosed to increment base by 1 instead of using 1 as offset (but then v0 is reused with 1 and 3 offsets)

https://github.com/llvm/llvm-project/pull/212507