release/19.x: AMDGPU: Fix inst-selection of large scratch offsets with sgpr base (#110256)

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

@@ -1911,7 +1911,7 @@ bool AMDGPUDAGToDAGISel::SelectScratchSAddr(SDNode *Parent, SDValue Addr,
                     0);
   }
 
-  Offset = CurDAG->getTargetConstant(COffsetVal, DL, MVT::i16);
+  Offset = CurDAG->getTargetConstant(COffsetVal, DL, MVT::i32);
 
   return true;
 }
@@ -1967,7 +1967,7 @@ bool AMDGPUDAGToDAGISel::SelectScratchSVAddr(SDNode *N, SDValue Addr,
           return false;
         if (checkFlatScratchSVSSwizzleBug(VAddr, SAddr, SplitImmOffset))
           return false;
-        Offset = CurDAG->getTargetConstant(SplitImmOffset, SDLoc(), MVT::i16);
+        Offset = CurDAG->getTargetConstant(SplitImmOffset, SDLoc(), MVT::i32);
         return true;
       }
     }
@@ -2000,7 +2000,7 @@ bool AMDGPUDAGToDAGISel::SelectScratchSVAddr(SDNode *N, SDValue Addr,
   if (checkFlatScratchSVSSwizzleBug(VAddr, SAddr, ImmOffset))
     return false;
   SAddr = SelectSAddrFI(CurDAG, SAddr);
-  Offset = CurDAG->getTargetConstant(ImmOffset, SDLoc(), MVT::i16);
+  Offset = CurDAG->getTargetConstant(ImmOffset, SDLoc(), MVT::i32);
   return true;
 }
 

llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch.ll

@@ -1513,4 +1513,243 @@ bb:
   ret void
 }
 
+define amdgpu_gs void @sgpr_base_large_offset(ptr addrspace(1) %out, ptr addrspace(5) inreg %sgpr_base) {
+; GFX9-LABEL: sgpr_base_large_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_add_u32 flat_scratch_lo, s0, s5
+; GFX9-NEXT:    s_addc_u32 flat_scratch_hi, s1, 0
+; GFX9-NEXT:    s_add_u32 s0, s2, 0xffe8
+; GFX9-NEXT:    scratch_load_dword v2, off, s0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    global_store_dword v[0:1], v2, off
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sgpr_base_large_offset:
+; GFX10:       ; %bb.0: ; %entry
+; GFX10-NEXT:    s_add_u32 s0, s0, s5
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
+; GFX10-NEXT:    s_add_u32 s0, s2, 0xffe8
+; GFX10-NEXT:    scratch_load_dword v2, off, s0
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    global_store_dword v[0:1], v2, off
+; GFX10-NEXT:    s_endpgm
+;
+; GFX940-LABEL: sgpr_base_large_offset:
+; GFX940:       ; %bb.0: ; %entry
+; GFX940-NEXT:    s_add_u32 s0, s0, 0xffe8
+; GFX940-NEXT:    scratch_load_dword v2, off, s0
+; GFX940-NEXT:    s_waitcnt vmcnt(0)
+; GFX940-NEXT:    global_store_dword v[0:1], v2, off sc0 sc1
+; GFX940-NEXT:    s_endpgm
+;
+; GFX11-LABEL: sgpr_base_large_offset:
+; GFX11:       ; %bb.0: ; %entry
+; GFX11-NEXT:    s_add_u32 s0, s0, 0xffe8
+; GFX11-NEXT:    scratch_load_b32 v2, off, s0
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX11-NEXT:    s_nop 0
+; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT:    s_endpgm
+;
+; GFX12-LABEL: sgpr_base_large_offset:
+; GFX12:       ; %bb.0: ; %entry
+; GFX12-NEXT:    scratch_load_b32 v2, off, s0 offset:65512
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX12-NEXT:    s_nop 0
+; GFX12-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT:    s_endpgm
+entry:
+  %large_offset = getelementptr i8, ptr addrspace(5) %sgpr_base, i32 65512
+  %load = load i32, ptr addrspace(5) %large_offset, align 4
+  store i32 %load, ptr addrspace(1) %out
+  ret void
+}
+
+define amdgpu_gs void @sgpr_base_large_offset_split(ptr addrspace(1) %out, ptr addrspace(5) inreg %sgpr_base) {
+; GFX9-LABEL: sgpr_base_large_offset_split:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_add_u32 flat_scratch_lo, s0, s5
+; GFX9-NEXT:    s_addc_u32 flat_scratch_hi, s1, 0
+; GFX9-NEXT:    s_and_b32 s0, s2, -4
+; GFX9-NEXT:    s_add_u32 s0, s0, 0x100ffe8
+; GFX9-NEXT:    scratch_load_dword v2, off, s0 glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    global_store_dword v[0:1], v2, off
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sgpr_base_large_offset_split:
+; GFX10:       ; %bb.0: ; %entry
+; GFX10-NEXT:    s_add_u32 s0, s0, s5
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
+; GFX10-NEXT:    s_and_b32 s0, s2, -4
+; GFX10-NEXT:    s_add_u32 s0, s0, 0x100ffe8
+; GFX10-NEXT:    scratch_load_dword v2, off, s0 glc dlc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    global_store_dword v[0:1], v2, off
+; GFX10-NEXT:    s_endpgm
+;
+; GFX940-LABEL: sgpr_base_large_offset_split:
+; GFX940:       ; %bb.0: ; %entry
+; GFX940-NEXT:    s_and_b32 s0, s0, -4
+; GFX940-NEXT:    s_add_u32 s0, s0, 0x100ffe8
+; GFX940-NEXT:    scratch_load_dword v2, off, s0 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0)
+; GFX940-NEXT:    global_store_dword v[0:1], v2, off sc0 sc1
+; GFX940-NEXT:    s_endpgm
+;
+; GFX11-LABEL: sgpr_base_large_offset_split:
+; GFX11:       ; %bb.0: ; %entry
+; GFX11-NEXT:    s_and_b32 s0, s0, -4
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT:    s_add_u32 s0, s0, 0x100ffe8
+; GFX11-NEXT:    scratch_load_b32 v2, off, s0 glc dlc
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX11-NEXT:    s_nop 0
+; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT:    s_endpgm
+;
+; GFX12-LABEL: sgpr_base_large_offset_split:
+; GFX12:       ; %bb.0: ; %entry
+; GFX12-NEXT:    s_and_b32 s0, s0, -4
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_add_co_u32 s0, s0, 0x100ffe8
+; GFX12-NEXT:    scratch_load_b32 v2, off, s0 scope:SCOPE_SYS
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX12-NEXT:    s_nop 0
+; GFX12-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT:    s_endpgm
+entry:
+  ;%allignedBase = alloca [33554432 x i8], align 4, addrspace(5)
+  %sgpr_base_i32 = ptrtoint ptr addrspace(5) %sgpr_base to i32
+  %sgpr_base_i32_align4 = and i32 %sgpr_base_i32, 4294967292
+  %sgpr_base_align4 = inttoptr i32 %sgpr_base_i32_align4 to ptr addrspace(5)
+  %split_offset = getelementptr inbounds [33554432 x i8], ptr addrspace(5) %sgpr_base_align4, i32 0, i32 16842728
+  %load = load volatile i32, ptr addrspace(5) %split_offset, align 4
+  store i32 %load, ptr addrspace(1) %out
+  ret void
+}
+
+define amdgpu_gs void @sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset(ptr addrspace(5) inreg %sgpr_base, i32 inreg %sidx, i32 %vidx) {
+; GFX9-LABEL: sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset:
+; GFX9:       ; %bb.0: ; %bb
+; GFX9-NEXT:    s_add_u32 flat_scratch_lo, s0, s5
+; GFX9-NEXT:    v_add_u32_e32 v0, s3, v0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0xffe8
+; GFX9-NEXT:    s_addc_u32 flat_scratch_hi, s1, 0
+; GFX9-NEXT:    v_add3_u32 v0, s2, v0, v1
+; GFX9-NEXT:    v_mov_b32_e32 v1, 15
+; GFX9-NEXT:    scratch_store_dword v0, v1, off
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset:
+; GFX10:       ; %bb.0: ; %bb
+; GFX10-NEXT:    s_add_u32 s0, s0, s5
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
+; GFX10-NEXT:    v_add_nc_u32_e32 v0, s3, v0
+; GFX10-NEXT:    v_mov_b32_e32 v1, 15
+; GFX10-NEXT:    v_add3_u32 v0, s2, v0, 0xffe8
+; GFX10-NEXT:    scratch_store_dword v0, v1, off
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_endpgm
+;
+; GFX940-LABEL: sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset:
+; GFX940:       ; %bb.0: ; %bb
+; GFX940-NEXT:    v_add_u32_e32 v0, s1, v0
+; GFX940-NEXT:    v_mov_b32_e32 v1, 0xffe8
+; GFX940-NEXT:    v_add3_u32 v0, s0, v0, v1
+; GFX940-NEXT:    v_mov_b32_e32 v1, 15
+; GFX940-NEXT:    scratch_store_dword v0, v1, off sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0)
+; GFX940-NEXT:    s_endpgm
+;
+; GFX11-LABEL: sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset:
+; GFX11:       ; %bb.0: ; %bb
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_add_nc_u32 v0, s1, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_add3_u32 v0, s0, v0, 0xffe8
+; GFX11-NEXT:    scratch_store_b32 v0, v1, off dlc
+; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-NEXT:    s_endpgm
+;
+; GFX12-LABEL: sgpr_base_plus_sgpr_plus_vgpr_plus_large_imm_offset:
+; GFX12:       ; %bb.0: ; %bb
+; GFX12-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_add_nc_u32 v0, s1, v0
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX12-NEXT:    v_add_nc_u32_e32 v0, s0, v0
+; GFX12-NEXT:    scratch_store_b32 v0, v1, off offset:65512 scope:SCOPE_SYS
+; GFX12-NEXT:    s_wait_storecnt 0x0
+; GFX12-NEXT:    s_endpgm
+bb:
+  %add1 = add nsw i32 %sidx, %vidx
+  %add2 = add nsw i32 %add1, 65512
+  %gep = getelementptr inbounds [33554432 x i8], ptr addrspace(5) %sgpr_base, i32 0, i32 %add2
+  store volatile i32 15, ptr addrspace(5) %gep, align 4
+  ret void
+}
+
+define amdgpu_gs void @sgpr_base_negative_offset(ptr addrspace(1) %out, ptr addrspace(5) inreg %scevgep) {
+; GFX9-LABEL: sgpr_base_negative_offset:
+; GFX9:       ; %bb.0: ; %entry
+; GFX9-NEXT:    s_add_u32 flat_scratch_lo, s0, s5
+; GFX9-NEXT:    s_addc_u32 flat_scratch_hi, s1, 0
+; GFX9-NEXT:    s_add_u32 s0, s2, 0xffffffe8
+; GFX9-NEXT:    scratch_load_dword v2, off, s0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    global_store_dword v[0:1], v2, off
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: sgpr_base_negative_offset:
+; GFX10:       ; %bb.0: ; %entry
+; GFX10-NEXT:    s_add_u32 s0, s0, s5
+; GFX10-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
+; GFX10-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
+; GFX10-NEXT:    scratch_load_dword v2, off, s2 offset:-24
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    global_store_dword v[0:1], v2, off
+; GFX10-NEXT:    s_endpgm
+;
+; GFX940-LABEL: sgpr_base_negative_offset:
+; GFX940:       ; %bb.0: ; %entry
+; GFX940-NEXT:    s_add_u32 s0, s0, 0xffffffe8
+; GFX940-NEXT:    scratch_load_dword v2, off, s0
+; GFX940-NEXT:    s_waitcnt vmcnt(0)
+; GFX940-NEXT:    global_store_dword v[0:1], v2, off sc0 sc1
+; GFX940-NEXT:    s_endpgm
+;
+; GFX11-LABEL: sgpr_base_negative_offset:
+; GFX11:       ; %bb.0: ; %entry
+; GFX11-NEXT:    scratch_load_b32 v2, off, s0 offset:-24
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX11-NEXT:    s_nop 0
+; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-NEXT:    s_endpgm
+;
+; GFX12-LABEL: sgpr_base_negative_offset:
+; GFX12:       ; %bb.0: ; %entry
+; GFX12-NEXT:    scratch_load_b32 v2, off, s0 offset:-24
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    global_store_b32 v[0:1], v2, off
+; GFX12-NEXT:    s_nop 0
+; GFX12-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX12-NEXT:    s_endpgm
+entry:
+  %scevgep28 = getelementptr i8, ptr addrspace(5) %scevgep, i32 -24
+  %0 = load i32, ptr addrspace(5) %scevgep28, align 4
+  store i32 %0, ptr addrspace(1) %out
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x()