[llvm] [AMDGPU] Sink uniform buffer address offsets into soffset (PR #169230)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,457 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -global-isel=0 < %s | FileCheck -check-prefixes=CHECK %s
+
+; Test comprehensive patterns for ADD(divergent, uniform) optimization in buffer loads
+
+; Basic workitem.id.x + uniform
+define amdgpu_kernel void @test_basic_workitem_uniform(ptr addrspace(1) %output, i32 %soffset) {
+; CHECK-LABEL: test_basic_workitem_uniform:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_dword s6, s[4:5], 0x2c
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    buffer_load_dword v0, v0, s[0:3], s6 offen
+; CHECK-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
+; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    global_store_dword v1, v0, s[0:1]
+; CHECK-NEXT:    s_endpgm
+  %desc = call <4 x i32> asm "", "=s"()
----------------
arsenm wrote:

I'd prefer to test in non-kernel functions and rely on inreg arguments for SGPRs. That avoids a lot of prolog noise 

https://github.com/llvm/llvm-project/pull/169230