; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 %s -o - | FileCheck -check-prefixes=GFX942,GFX942-SDAG %s ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 %s -o - | FileCheck -check-prefixes=GFX942,GFX942-GISEL %s define void @memset_pattern_i128_len0_dynvalue(ptr addrspace(1) align 16 %a, i128 %value) { ; GFX942-LABEL: memset_pattern_i128_len0_dynvalue: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i128 %value, i64 0, i1 false) ret void } define void @memset_pattern_i128_len1_dynvalue(ptr addrspace(1) align 16 %a, i128 %value) { ; GFX942-LABEL: memset_pattern_i128_len1_dynvalue: ; GFX942: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-NEXT: s_waitcnt vmcnt(0) ; GFX942-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i128 %value, i64 1, i1 false) ret void } define void @memset_pattern_i128_len1(ptr addrspace(1) align 16 %a) { ; GFX942-SDAG-LABEL: memset_pattern_i128_len1: ; GFX942-SDAG: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 0xaaaaaaaa ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_len1: ; GFX942-GISEL: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[0:1] ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 1, i1 false) ret void } define void @memset_pattern_i128_constlen_mainloop_and_residual_taken(ptr addrspace(1) align 16 %a) { ; GFX942-SDAG-LABEL: memset_pattern_i128_constlen_mainloop_and_residual_taken: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: s_mov_b32 s5, 0xcccccccc ; GFX942-SDAG-NEXT: s_mov_b32 s4, 0xdddddddd ; GFX942-SDAG-NEXT: s_mov_b32 s7, 0xaaaaaaaa ; GFX942-SDAG-NEXT: s_mov_b32 s6, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[4:5] ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[6:7], s[6:7] ; GFX942-SDAG-NEXT: s_mov_b64 s[2:3], 0x1000 ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[2:3], v[0:1] ; GFX942-SDAG-NEXT: .LBB3_1: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s0, s0, 16 ; GFX942-SDAG-NEXT: s_addc_u32 s1, s1, 0 ; GFX942-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:240 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:224 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:208 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:192 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:176 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:160 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:144 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:128 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:112 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:96 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:80 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:64 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:48 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:32 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:16 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[2:3], v[4:7], off ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[2:3], v[2:3], 0, s[2:3] ; GFX942-SDAG-NEXT: s_and_b64 vcc, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_vccnz .LBB3_1 ; GFX942-SDAG-NEXT: ; %bb.2: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0x100 ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[4:5], v[0:1], 0, s[0:1] ; GFX942-SDAG-NEXT: s_mov_b32 s0, 1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v0, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v1, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xaaaaaaaa ; GFX942-SDAG-NEXT: .LBB3_3: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_i32 s1, s0, 1 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[4:5], v[0:3], off ; GFX942-SDAG-NEXT: s_cmp_lt_u32 s0, 3 ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[4:5], v[4:5], 0, 16 ; GFX942-SDAG-NEXT: s_mov_b32 s0, s1 ; GFX942-SDAG-NEXT: s_cbranch_scc1 .LBB3_3 ; GFX942-SDAG-NEXT: ; %bb.4: ; %memset.pattern-post-expansion ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_constlen_mainloop_and_residual_taken: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], v[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[4:5] ; GFX942-GISEL-NEXT: .LBB3_1: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 16, v8 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:16 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:32 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:48 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:64 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:80 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:96 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:112 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:128 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:144 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:160 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:176 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:192 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:208 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:224 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[6:7], v[2:5], off offset:240 ; GFX942-GISEL-NEXT: v_addc_co_u32_e64 v9, s[0:1], 0, v9, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 0x1000, v6 ; GFX942-GISEL-NEXT: s_xor_b64 s[0:1], s[0:1], -1 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc ; GFX942-GISEL-NEXT: s_xor_b64 s[0:1], s[0:1], -1 ; GFX942-GISEL-NEXT: s_and_b64 vcc, exec, s[0:1] ; GFX942-GISEL-NEXT: s_cbranch_vccnz .LBB3_1 ; GFX942-GISEL-NEXT: ; %bb.2: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v4, vcc, 0x100, v0 ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_mov_b32 s4, 1 ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b32_e32 v7, 0 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v6, s4 ; GFX942-GISEL-NEXT: .LBB3_3: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[4:5], v[0:3], off ; GFX942-GISEL-NEXT: v_add_co_u32_e64 v4, s[0:1], 16, v4 ; GFX942-GISEL-NEXT: v_cmp_gt_u64_e32 vcc, 3, v[6:7] ; GFX942-GISEL-NEXT: v_add_u32_e32 v6, 1, v6 ; GFX942-GISEL-NEXT: v_addc_co_u32_e64 v5, s[0:1], 0, v5, s[0:1] ; GFX942-GISEL-NEXT: s_cbranch_vccnz .LBB3_3 ; GFX942-GISEL-NEXT: ; %bb.4: ; %memset.pattern-post-expansion ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 19, i1 false) ret void } define void @memset_pattern_i128_len1_lds(ptr addrspace(3) align 16 %a) { ; GFX942-SDAG-LABEL: memset_pattern_i128_len1_lds: ; GFX942-SDAG: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 0xaaaaaaaa ; GFX942-SDAG-NEXT: ds_write_b128 v0, v[2:5] ; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_len1_lds: ; GFX942-GISEL: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[0:1] ; GFX942-GISEL-NEXT: ds_write_b128 v0, v[2:5] ; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(3) align 16 %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 1, i1 false) ret void } define void @memset_pattern_i128_len1_no_align(ptr addrspace(1) %a) { ; GFX942-SDAG-LABEL: memset_pattern_i128_len1_no_align: ; GFX942-SDAG: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 0xaaaaaaaa ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_len1_no_align: ; GFX942-GISEL: ; %bb.0: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[0:1] ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 1, i1 false) ret void } define void @memset_pattern_i128_len16(ptr addrspace(1) align 16 %a) { ; GFX942-SDAG-LABEL: memset_pattern_i128_len16: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: s_mov_b32 s5, 0xcccccccc ; GFX942-SDAG-NEXT: s_mov_b32 s4, 0xdddddddd ; GFX942-SDAG-NEXT: s_mov_b32 s7, 0xaaaaaaaa ; GFX942-SDAG-NEXT: s_mov_b32 s6, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[2:3], s[4:5] ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[4:5], s[6:7] ; GFX942-SDAG-NEXT: s_mov_b64 s[2:3], 0x1000 ; GFX942-SDAG-NEXT: .LBB6_1: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s0, s0, 16 ; GFX942-SDAG-NEXT: s_addc_u32 s1, s1, 0 ; GFX942-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:240 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:224 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:208 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:192 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:176 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:160 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:144 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:128 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:112 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:96 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:80 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:64 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:48 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:32 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:16 ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, s[2:3] ; GFX942-SDAG-NEXT: s_and_b64 vcc, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_vccnz .LBB6_1 ; GFX942-SDAG-NEXT: ; %bb.2: ; %memset.pattern-post-expansion ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_len16: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b32 s2, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s3, 0xaaaaaaaa ; GFX942-GISEL-NEXT: s_mov_b32 s0, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s1, 0xcccccccc ; GFX942-GISEL-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[4:5] ; GFX942-GISEL-NEXT: .LBB6_1: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 16, v6 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:16 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:32 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:48 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:64 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:80 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:96 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:112 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:128 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:144 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:160 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:176 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:192 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:208 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:224 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[2:5], off offset:240 ; GFX942-GISEL-NEXT: v_addc_co_u32_e64 v7, s[0:1], 0, v7, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 0x1000, v0 ; GFX942-GISEL-NEXT: s_xor_b64 s[0:1], s[0:1], -1 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_xor_b64 s[0:1], s[0:1], -1 ; GFX942-GISEL-NEXT: s_and_b64 vcc, exec, s[0:1] ; GFX942-GISEL-NEXT: s_cbranch_vccnz .LBB6_1 ; GFX942-GISEL-NEXT: ; %bb.2: ; %memset.pattern-post-expansion ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 16, i1 false) ret void } define void @memset_pattern_i128_dynlen(ptr addrspace(1) align 16 %a, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i128_dynlen: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[2:3] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB7_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, 0xaaaaaaaa ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB7_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[2:3] ; GFX942-SDAG-NEXT: global_store_dwordx4 v[0:1], v[4:7], off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, 16 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB7_2 ; GFX942-SDAG-NEXT: .LBB7_3: ; %Flow1 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_dynlen: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[2:3] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB7_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: s_mov_b32 s8, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s9, 0xcccccccc ; GFX942-GISEL-NEXT: s_mov_b32 s10, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s11, 0xaaaaaaaa ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[8:9] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[10:11] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[2:3] ; GFX942-GISEL-NEXT: .LBB7_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v8 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[0:1], v[4:7], off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 16, v0 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[8:9], v[2:3] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB7_2 ; GFX942-GISEL-NEXT: .LBB7_3: ; %Flow1 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 %len, i1 false) ret void } define void @memset_pattern_i128_dynlen_lds(ptr addrspace(3) align 16 %a, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i128_dynlen_lds: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, v1 ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB8_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xdddddddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xcccccccc ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xbbbbbbbb ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, 0xaaaaaaaa ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB8_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[6:7] ; GFX942-SDAG-NEXT: ds_write_b128 v0, v[2:5] ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_add_u32_e32 v0, 16, v0 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB8_2 ; GFX942-SDAG-NEXT: .LBB8_3: ; %Flow1 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i128_dynlen_lds: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v6, v1 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v7, v2 ; GFX942-GISEL-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB8_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: s_mov_b32 s4, 0xdddddddd ; GFX942-GISEL-NEXT: s_mov_b32 s5, 0xcccccccc ; GFX942-GISEL-NEXT: s_mov_b32 s6, 0xbbbbbbbb ; GFX942-GISEL-NEXT: s_mov_b32 s7, 0xaaaaaaaa ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[4:5] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[6:7] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[0:1] ; GFX942-GISEL-NEXT: .LBB8_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 1, v8 ; GFX942-GISEL-NEXT: ds_write_b128 v0, v[2:5] ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[8:9], v[6:7] ; GFX942-GISEL-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-GISEL-NEXT: v_add_u32_e32 v0, 16, v0 ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB8_2 ; GFX942-GISEL-NEXT: .LBB8_3: ; %Flow1 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(3) align 16 %a, i128 u0xaaaaaaaabbbbbbbbccccccccdddddddd, i64 %len, i1 false) ret void } define void @memset_pattern_i32_dynlen(ptr addrspace(1) align 16 %a, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i32_dynlen: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_and_b32_e32 v10, -4, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v11, v3 ; GFX942-SDAG-NEXT: v_and_b32_e32 v8, 3, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v9, 0 ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB9_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, 0xaabbccdd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, v4 ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[12:13], v[0:1] ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB9_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 4 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[10:11] ; GFX942-SDAG-NEXT: global_store_dwordx4 v[12:13], v[4:7], off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[12:13], v[12:13], 0, 64 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB9_2 ; GFX942-SDAG-NEXT: .LBB9_3: ; %Flow6 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[8:9] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB9_6 ; GFX942-SDAG-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: v_lshlrev_b64 v[2:3], 2, v[2:3] ; GFX942-SDAG-NEXT: v_and_b32_e32 v2, -16, v2 ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3] ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xaabbccdd ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB9_5: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[8:9] ; GFX942-SDAG-NEXT: global_store_dword v[0:1], v2, off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, 4 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB9_5 ; GFX942-SDAG-NEXT: .LBB9_6: ; %Flow4 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i32_dynlen: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_and_b32_e32 v8, 3, v2 ; GFX942-GISEL-NEXT: v_sub_co_u32_e32 v10, vcc, v2, v8 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v9, 0 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_subbrev_co_u32_e32 v11, vcc, 0, v3, vcc ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB9_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: s_mov_b32 s8, 0xaabbccdd ; GFX942-GISEL-NEXT: s_mov_b32 s9, s8 ; GFX942-GISEL-NEXT: s_mov_b32 s10, s8 ; GFX942-GISEL-NEXT: s_mov_b32 s11, s8 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[8:9] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[10:11] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[12:13], v[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[2:3] ; GFX942-GISEL-NEXT: .LBB9_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 4, v14 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[12:13], v[4:7], off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 64, v12 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[14:15], v[10:11] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB9_2 ; GFX942-GISEL-NEXT: .LBB9_3: ; %Flow6 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[8:9] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB9_6 ; GFX942-GISEL-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: v_lshrrev_b64 v[2:3], 2, v[2:3] ; GFX942-GISEL-NEXT: v_lshl_add_u64 v[0:1], v[2:3], 4, v[0:1] ; GFX942-GISEL-NEXT: v_mov_b32_e32 v4, 0xaabbccdd ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[2:3] ; GFX942-GISEL-NEXT: .LBB9_5: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v2, vcc, 1, v2 ; GFX942-GISEL-NEXT: global_store_dword v[0:1], v4, off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 4, v0 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[2:3], v[8:9] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB9_5 ; GFX942-GISEL-NEXT: .LBB9_6: ; %Flow4 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i32 u0xaabbccdd, i64 %len, i1 false) ret void } define void @memset_pattern_i32_dynval_dynlen(ptr addrspace(1) align 16 %a, i32 %val, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i32_dynval_dynlen: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v8, v3 ; GFX942-SDAG-NEXT: v_and_b32_e32 v10, -4, v8 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v11, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v9, v4 ; GFX942-SDAG-NEXT: v_and_b32_e32 v6, 3, v8 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, 0 ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB10_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, v2 ; GFX942-SDAG-NEXT: v_mov_b64_e32 v[12:13], v[0:1] ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB10_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 4 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[10:11] ; GFX942-SDAG-NEXT: global_store_dwordx4 v[12:13], v[2:5], off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[12:13], v[12:13], 0, 64 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB10_2 ; GFX942-SDAG-NEXT: .LBB10_3: ; %Flow6 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB10_6 ; GFX942-SDAG-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: v_lshlrev_b64 v[4:5], 2, v[8:9] ; GFX942-SDAG-NEXT: v_and_b32_e32 v4, -16, v4 ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, v[4:5] ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB10_5: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[6:7] ; GFX942-SDAG-NEXT: global_store_dword v[0:1], v2, off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, 4 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB10_5 ; GFX942-SDAG-NEXT: .LBB10_6: ; %Flow4 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i32_dynval_dynlen: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v8, v3 ; GFX942-GISEL-NEXT: v_and_b32_e32 v6, 3, v8 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v9, v4 ; GFX942-GISEL-NEXT: v_sub_co_u32_e32 v10, vcc, v8, v6 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v7, 0 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_subbrev_co_u32_e32 v11, vcc, 0, v9, vcc ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB10_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v4, v2 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v5, v2 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[12:13], v[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[14:15], s[2:3] ; GFX942-GISEL-NEXT: .LBB10_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v14, vcc, 4, v14 ; GFX942-GISEL-NEXT: global_store_dwordx4 v[12:13], v[2:5], off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v15, vcc, 0, v15, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 64, v12 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[14:15], v[10:11] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB10_2 ; GFX942-GISEL-NEXT: .LBB10_3: ; %Flow6 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB10_6 ; GFX942-GISEL-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: v_lshrrev_b64 v[4:5], 2, v[8:9] ; GFX942-GISEL-NEXT: v_lshl_add_u64 v[0:1], v[4:5], 4, v[0:1] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[2:3] ; GFX942-GISEL-NEXT: .LBB10_5: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v4, vcc, 1, v4 ; GFX942-GISEL-NEXT: global_store_dword v[0:1], v2, off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v5, vcc, 0, v5, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 4, v0 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[4:5], v[6:7] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB10_5 ; GFX942-GISEL-NEXT: .LBB10_6: ; %Flow4 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i32 %val, i64 %len, i1 false) ret void } ; For i96, the store size and the alloc size differ on amdgpu, this case is not ; optimized. define void @memset_pattern_i96_dynval_dynlen(ptr addrspace(1) align 16 %a, i96 %val, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i96_dynval_dynlen: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, v6 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, v5 ; GFX942-SDAG-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[2:3], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB11_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: .LBB11_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[4:5], v[6:7] ; GFX942-SDAG-NEXT: global_store_dwordx3 v[0:1], v[2:4], off ; GFX942-SDAG-NEXT: s_or_b64 s[0:1], vcc, s[0:1] ; GFX942-SDAG-NEXT: v_lshl_add_u64 v[0:1], v[0:1], 0, 16 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB11_2 ; GFX942-SDAG-NEXT: .LBB11_3: ; %Flow1 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[2:3] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i96_dynval_dynlen: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v8, v5 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v9, v6 ; GFX942-GISEL-NEXT: s_mov_b64 s[2:3], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[8:9] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB11_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], s[2:3] ; GFX942-GISEL-NEXT: .LBB11_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v6, vcc, 1, v6 ; GFX942-GISEL-NEXT: global_store_dwordx3 v[0:1], v[2:4], off ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v7, vcc, 0, v7, vcc ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 16, v0 ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e64 s[0:1], v[6:7], v[8:9] ; GFX942-GISEL-NEXT: s_or_b64 s[2:3], s[0:1], s[2:3] ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB11_2 ; GFX942-GISEL-NEXT: .LBB11_3: ; %Flow1 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] call void @llvm.experimental.memset.pattern(ptr addrspace(1) align 16 %a, i96 %val, i64 %len, i1 false) ret void } define void @memset_pattern_i64_as7_dynlen(ptr addrspace(7) inreg align 16 %a, i32 %offset, i64 %len) { ; GFX942-SDAG-LABEL: memset_pattern_i64_as7_dynlen: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v8, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v11, v2 ; GFX942-SDAG-NEXT: v_and_b32_e32 v10, -2, v8 ; GFX942-SDAG-NEXT: v_and_b32_e32 v6, 1, v8 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, 0 ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB12_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, 0xccccdddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0xaaaabbbb ; GFX942-SDAG-NEXT: v_add_u32_e32 v1, s16, v0 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, v2 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, v3 ; GFX942-SDAG-NEXT: s_mov_b64 s[8:9], 0 ; GFX942-SDAG-NEXT: .LBB12_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s8, s8, 2 ; GFX942-SDAG-NEXT: s_addc_u32 s9, s9, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[8:9], v[10:11] ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[2:5], v1, s[0:3], 0 offen ; GFX942-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-SDAG-NEXT: v_add_u32_e32 v1, 32, v1 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB12_2 ; GFX942-SDAG-NEXT: .LBB12_3: ; %Flow3 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB12_6 ; GFX942-SDAG-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: v_lshlrev_b32_e32 v1, 3, v8 ; GFX942-SDAG-NEXT: v_and_b32_e32 v1, -16, v1 ; GFX942-SDAG-NEXT: v_add3_u32 v2, v0, v1, s16 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v0, 0xccccdddd ; GFX942-SDAG-NEXT: v_mov_b32_e32 v1, 0xaaaabbbb ; GFX942-SDAG-NEXT: s_mov_b64 s[8:9], 0 ; GFX942-SDAG-NEXT: .LBB12_5: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s8, s8, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s9, s9, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[8:9], v[6:7] ; GFX942-SDAG-NEXT: buffer_store_dwordx2 v[0:1], v2, s[0:3], 0 offen ; GFX942-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-SDAG-NEXT: v_add_u32_e32 v2, 8, v2 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB12_5 ; GFX942-SDAG-NEXT: .LBB12_6: ; %Flow1 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i64_as7_dynlen: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v8, v1 ; GFX942-GISEL-NEXT: v_and_b32_e32 v6, 1, v8 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v9, v2 ; GFX942-GISEL-NEXT: v_sub_co_u32_e32 v10, vcc, v8, v6 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v7, 0 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_subbrev_co_u32_e32 v11, vcc, 0, v9, vcc ; GFX942-GISEL-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[10:11] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB12_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: s_mov_b32 s8, 0xccccdddd ; GFX942-GISEL-NEXT: s_mov_b32 s9, 0xaaaabbbb ; GFX942-GISEL-NEXT: s_mov_b64 s[10:11], s[8:9] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[8:9] ; GFX942-GISEL-NEXT: v_add_u32_e32 v1, s16, v0 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[4:5], s[10:11] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[12:13], s[4:5] ; GFX942-GISEL-NEXT: .LBB12_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v12, vcc, 2, v12 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[2:5], v1, s[0:3], 0 offen ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v13, vcc, 0, v13, vcc ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[12:13], v[10:11] ; GFX942-GISEL-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-GISEL-NEXT: v_add_u32_e32 v1, 32, v1 ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB12_2 ; GFX942-GISEL-NEXT: .LBB12_3: ; %Flow3 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-GISEL-NEXT: v_and_b32_e32 v1, 1, v6 ; GFX942-GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB12_6 ; GFX942-GISEL-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: v_lshrrev_b64 v[2:3], 1, v[8:9] ; GFX942-GISEL-NEXT: v_lshlrev_b32_e32 v1, 4, v2 ; GFX942-GISEL-NEXT: s_mov_b64 s[6:7], 0 ; GFX942-GISEL-NEXT: v_add3_u32 v4, v0, v1, s16 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v0, 0xccccdddd ; GFX942-GISEL-NEXT: v_mov_b32_e32 v1, 0xaaaabbbb ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[2:3], s[6:7] ; GFX942-GISEL-NEXT: .LBB12_5: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v2, vcc, 1, v2 ; GFX942-GISEL-NEXT: buffer_store_dwordx2 v[0:1], v4, s[0:3], 0 offen ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v3, vcc, 0, v3, vcc ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[2:3], v[6:7] ; GFX942-GISEL-NEXT: s_or_b64 s[6:7], vcc, s[6:7] ; GFX942-GISEL-NEXT: v_add_u32_e32 v4, 8, v4 ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB12_5 ; GFX942-GISEL-NEXT: .LBB12_6: ; %Flow1 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] %p = getelementptr inbounds i8, ptr addrspace(7) %a, i32 %offset call void @llvm.experimental.memset.pattern(ptr addrspace(7) %p, i64 u0xaaaabbbbccccdddd, i64 %len, i1 false) ret void } define void @memset_pattern_i64_as7_dynlen_dynval(ptr addrspace(7) inreg align 16 %a, i32 %offset, i64 %len, i64 %val) { ; GFX942-SDAG-LABEL: memset_pattern_i64_as7_dynlen_dynval: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v9, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, v2 ; GFX942-SDAG-NEXT: v_and_b32_e32 v6, -2, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v8, v3 ; GFX942-SDAG-NEXT: v_and_b32_e32 v2, 1, v4 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, 0 ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[6:7] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB13_3 ; GFX942-SDAG-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-SDAG-NEXT: v_add_u32_e32 v1, s16, v0 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v10, v8 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v11, v9 ; GFX942-SDAG-NEXT: s_mov_b64 s[8:9], 0 ; GFX942-SDAG-NEXT: .LBB13_2: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s8, s8, 2 ; GFX942-SDAG-NEXT: s_addc_u32 s9, s9, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[8:9], v[6:7] ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[8:11], v1, s[0:3], 0 offen ; GFX942-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-SDAG-NEXT: v_add_u32_e32 v1, 32, v1 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB13_2 ; GFX942-SDAG-NEXT: .LBB13_3: ; %Flow3 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[2:3] ; GFX942-SDAG-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-SDAG-NEXT: s_cbranch_execz .LBB13_6 ; GFX942-SDAG-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: v_lshlrev_b32_e32 v1, 3, v4 ; GFX942-SDAG-NEXT: v_and_b32_e32 v1, -16, v1 ; GFX942-SDAG-NEXT: v_add3_u32 v0, v0, v1, s16 ; GFX942-SDAG-NEXT: s_mov_b64 s[8:9], 0 ; GFX942-SDAG-NEXT: .LBB13_5: ; %memset.pattern-expansion-residual-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s8, s8, 1 ; GFX942-SDAG-NEXT: s_addc_u32 s9, s9, 0 ; GFX942-SDAG-NEXT: v_cmp_ge_u64_e32 vcc, s[8:9], v[2:3] ; GFX942-SDAG-NEXT: buffer_store_dwordx2 v[8:9], v0, s[0:3], 0 offen ; GFX942-SDAG-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-SDAG-NEXT: v_add_u32_e32 v0, 8, v0 ; GFX942-SDAG-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-SDAG-NEXT: s_cbranch_execnz .LBB13_5 ; GFX942-SDAG-NEXT: .LBB13_6: ; %Flow1 ; GFX942-SDAG-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i64_as7_dynlen_dynval: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v6, v1 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v7, v2 ; GFX942-GISEL-NEXT: v_and_b32_e32 v2, 1, v6 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v11, v4 ; GFX942-GISEL-NEXT: v_sub_co_u32_e32 v4, vcc, v6, v2 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v10, v3 ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_subbrev_co_u32_e32 v5, vcc, 0, v7, vcc ; GFX942-GISEL-NEXT: v_mov_b32_e32 v3, 0 ; GFX942-GISEL-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-GISEL-NEXT: v_cmp_ne_u64_e32 vcc, 0, v[4:5] ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[6:7], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB13_3 ; GFX942-GISEL-NEXT: ; %bb.1: ; %memset.pattern-expansion-main-body.preheader ; GFX942-GISEL-NEXT: v_add_u32_e32 v1, s16, v0 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[12:13], v[10:11] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[8:9], s[4:5] ; GFX942-GISEL-NEXT: .LBB13_2: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v8, vcc, 2, v8 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[10:13], v1, s[0:3], 0 offen ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v9, vcc, 0, v9, vcc ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[8:9], v[4:5] ; GFX942-GISEL-NEXT: s_or_b64 s[4:5], vcc, s[4:5] ; GFX942-GISEL-NEXT: v_add_u32_e32 v1, 32, v1 ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB13_2 ; GFX942-GISEL-NEXT: .LBB13_3: ; %Flow3 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[6:7] ; GFX942-GISEL-NEXT: v_and_b32_e32 v1, 1, v2 ; GFX942-GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 ; GFX942-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc ; GFX942-GISEL-NEXT: s_cbranch_execz .LBB13_6 ; GFX942-GISEL-NEXT: ; %bb.4: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: v_lshrrev_b64 v[4:5], 1, v[6:7] ; GFX942-GISEL-NEXT: v_lshlrev_b32_e32 v1, 4, v4 ; GFX942-GISEL-NEXT: s_mov_b64 s[6:7], 0 ; GFX942-GISEL-NEXT: v_add3_u32 v4, v0, v1, s16 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[6:7] ; GFX942-GISEL-NEXT: .LBB13_5: ; %memset.pattern-expansion-residual-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 1, v0 ; GFX942-GISEL-NEXT: buffer_store_dwordx2 v[10:11], v4, s[0:3], 0 offen ; GFX942-GISEL-NEXT: s_nop 0 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: v_cmp_ge_u64_e32 vcc, v[0:1], v[2:3] ; GFX942-GISEL-NEXT: s_or_b64 s[6:7], vcc, s[6:7] ; GFX942-GISEL-NEXT: v_add_u32_e32 v4, 8, v4 ; GFX942-GISEL-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX942-GISEL-NEXT: s_cbranch_execnz .LBB13_5 ; GFX942-GISEL-NEXT: .LBB13_6: ; %Flow1 ; GFX942-GISEL-NEXT: s_or_b64 exec, exec, s[4:5] ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] %p = getelementptr inbounds i8, ptr addrspace(7) %a, i32 %offset call void @llvm.experimental.memset.pattern(ptr addrspace(7) %p, i64 %val, i64 %len, i1 false) ret void } define void @memset_pattern_i64_as7_len33_dynval(ptr addrspace(7) inreg align 16 %a, i32 %offset, i64 %val) { ; GFX942-SDAG-LABEL: memset_pattern_i64_as7_len33_dynval: ; GFX942-SDAG: ; %bb.0: ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-SDAG-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-SDAG-NEXT: v_add_u32_e32 v0, s16, v0 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v4, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v5, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v6, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v7, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v8, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v9, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v10, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v11, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v12, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v13, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v14, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v15, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v16, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v17, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v18, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v19, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v20, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v21, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v22, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v23, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v24, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v25, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v26, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v27, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v28, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v29, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v30, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v31, v3 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v32, v1 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v33, v3 ; GFX942-SDAG-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-SDAG-NEXT: v_mov_b32_e32 v1, v0 ; GFX942-SDAG-NEXT: .LBB14_1: ; %memset.pattern-expansion-main-body ; GFX942-SDAG-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-SDAG-NEXT: s_add_u32 s4, s4, 32 ; GFX942-SDAG-NEXT: s_addc_u32 s5, s5, 0 ; GFX942-SDAG-NEXT: s_cselect_b64 s[6:7], -1, 0 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[2:5], v1, s[0:3], 0 offen ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[6:9], v1, s[0:3], 0 offen offset:16 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[10:13], v1, s[0:3], 0 offen offset:32 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[14:17], v1, s[0:3], 0 offen offset:48 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[18:21], v1, s[0:3], 0 offen offset:64 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[22:25], v1, s[0:3], 0 offen offset:80 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[26:29], v1, s[0:3], 0 offen offset:96 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[30:33], v1, s[0:3], 0 offen offset:112 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[2:5], v1, s[0:3], 0 offen offset:128 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[6:9], v1, s[0:3], 0 offen offset:144 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[10:13], v1, s[0:3], 0 offen offset:160 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[14:17], v1, s[0:3], 0 offen offset:176 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[18:21], v1, s[0:3], 0 offen offset:192 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[22:25], v1, s[0:3], 0 offen offset:208 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[26:29], v1, s[0:3], 0 offen offset:224 ; GFX942-SDAG-NEXT: buffer_store_dwordx4 v[30:33], v1, s[0:3], 0 offen offset:240 ; GFX942-SDAG-NEXT: v_add_u32_e32 v1, 0x2000, v1 ; GFX942-SDAG-NEXT: s_and_b64 vcc, exec, s[6:7] ; GFX942-SDAG-NEXT: s_cbranch_vccnz .LBB14_1 ; GFX942-SDAG-NEXT: ; %bb.2: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-SDAG-NEXT: buffer_store_dwordx2 v[2:3], v0, s[0:3], 0 offen offset:256 ; GFX942-SDAG-NEXT: s_waitcnt vmcnt(0) ; GFX942-SDAG-NEXT: s_setpc_b64 s[30:31] ; ; GFX942-GISEL-LABEL: memset_pattern_i64_as7_len33_dynval: ; GFX942-GISEL: ; %bb.0: ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-GISEL-NEXT: v_mov_b32_e32 v4, v1 ; GFX942-GISEL-NEXT: v_mov_b32_e32 v5, v2 ; GFX942-GISEL-NEXT: v_add_u32_e32 v2, s16, v0 ; GFX942-GISEL-NEXT: s_mov_b64 s[4:5], 0 ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[6:7], v[4:5] ; GFX942-GISEL-NEXT: v_mov_b64_e32 v[0:1], s[4:5] ; GFX942-GISEL-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-GISEL-NEXT: .LBB14_1: ; %memset.pattern-expansion-main-body ; GFX942-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-GISEL-NEXT: v_add_co_u32_e32 v0, vcc, 32, v0 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:16 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:32 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:48 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:64 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:80 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:96 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:112 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:128 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:144 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:160 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:176 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:192 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:208 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:224 ; GFX942-GISEL-NEXT: buffer_store_dwordx4 v[4:7], v3, s[0:3], 0 offen offset:240 ; GFX942-GISEL-NEXT: v_addc_co_u32_e32 v1, vcc, 0, v1, vcc ; GFX942-GISEL-NEXT: s_xor_b64 s[4:5], vcc, -1 ; GFX942-GISEL-NEXT: s_xor_b64 s[4:5], s[4:5], -1 ; GFX942-GISEL-NEXT: v_add_u32_e32 v3, 0x2000, v3 ; GFX942-GISEL-NEXT: s_and_b64 vcc, exec, s[4:5] ; GFX942-GISEL-NEXT: s_cbranch_vccnz .LBB14_1 ; GFX942-GISEL-NEXT: ; %bb.2: ; %memset.pattern-expansion-residual-body.preheader ; GFX942-GISEL-NEXT: buffer_store_dwordx2 v[4:5], v2, s[0:3], 0 offen offset:256 ; GFX942-GISEL-NEXT: s_waitcnt vmcnt(0) ; GFX942-GISEL-NEXT: s_setpc_b64 s[30:31] %p = getelementptr inbounds i8, ptr addrspace(7) %a, i32 %offset call void @llvm.experimental.memset.pattern(ptr addrspace(7) %p, i64 %val, i64 33, i1 false) ret void }