180 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -mcpu=gfx942 -amdgpu-mfma-vgpr-form < %s | FileCheck %s3 4target triple = "amdgcn-amd-amdhsa"5 6define amdgpu_kernel void @test_rewrite_mfma_copy_to_agpr_phi(ptr addrspace(1) %arg0, ptr addrspace(1) %arg1, i1 %cond) #0 {7; CHECK-LABEL: test_rewrite_mfma_copy_to_agpr_phi:8; CHECK: ; %bb.0: ; %bb9; CHECK-NEXT: s_load_dword s6, s[4:5], 0x1010; CHECK-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x011; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v012; CHECK-NEXT: v_lshlrev_b32_e32 v32, 7, v013; CHECK-NEXT: s_waitcnt lgkmcnt(0)14; CHECK-NEXT: s_bitcmp0_b32 s6, 015; CHECK-NEXT: s_cbranch_scc0 .LBB0_216; CHECK-NEXT: ; %bb.1: ; %else17; CHECK-NEXT: global_load_dwordx4 a[28:31], v32, s[2:3] offset:11218; CHECK-NEXT: global_load_dwordx4 a[24:27], v32, s[2:3] offset:9619; CHECK-NEXT: global_load_dwordx4 a[20:23], v32, s[2:3] offset:8020; CHECK-NEXT: global_load_dwordx4 a[16:19], v32, s[2:3] offset:6421; CHECK-NEXT: global_load_dwordx4 a[12:15], v32, s[2:3] offset:4822; CHECK-NEXT: global_load_dwordx4 a[8:11], v32, s[2:3] offset:3223; CHECK-NEXT: global_load_dwordx4 a[4:7], v32, s[2:3] offset:1624; CHECK-NEXT: global_load_dwordx4 a[0:3], v32, s[2:3]25; CHECK-NEXT: v_mov_b32_e32 v33, 2.026; CHECK-NEXT: v_mov_b32_e32 v34, 4.027; CHECK-NEXT: s_waitcnt vmcnt(0)28; CHECK-NEXT: s_nop 029; CHECK-NEXT: v_mfma_f32_32x32x1_2b_f32 a[0:31], v33, v34, a[0:31]30; CHECK-NEXT: ; kill: def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $exec31; CHECK-NEXT: s_cbranch_execz .LBB0_332; CHECK-NEXT: s_branch .LBB0_433; CHECK-NEXT: .LBB0_2:34; CHECK-NEXT: ; implicit-def: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr3135; CHECK-NEXT: .LBB0_3: ; %if36; CHECK-NEXT: s_nop 1537; CHECK-NEXT: global_load_dwordx4 a[28:31], v32, s[0:1] offset:11238; CHECK-NEXT: global_load_dwordx4 a[24:27], v32, s[0:1] offset:9639; CHECK-NEXT: global_load_dwordx4 a[20:23], v32, s[0:1] offset:8040; CHECK-NEXT: global_load_dwordx4 a[16:19], v32, s[0:1] offset:6441; CHECK-NEXT: global_load_dwordx4 a[12:15], v32, s[0:1] offset:4842; CHECK-NEXT: global_load_dwordx4 a[8:11], v32, s[0:1] offset:3243; CHECK-NEXT: global_load_dwordx4 a[4:7], v32, s[0:1] offset:1644; CHECK-NEXT: global_load_dwordx4 a[0:3], v32, s[0:1]45; CHECK-NEXT: v_mov_b32_e32 v32, 2.046; CHECK-NEXT: v_mov_b32_e32 v33, 4.047; CHECK-NEXT: s_waitcnt vmcnt(0)48; CHECK-NEXT: s_nop 049; CHECK-NEXT: v_mfma_f32_32x32x1_2b_f32 a[0:31], v32, v33, a[0:31]50; CHECK-NEXT: ; kill: def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $exec51; CHECK-NEXT: .LBB0_4: ; %endif52; CHECK-NEXT: ;;#ASMSTART53; CHECK-NEXT: ; use a[0:31]54; CHECK-NEXT: ;;#ASMEND55; CHECK-NEXT: s_endpgm56bb:57 %id = call i32 @llvm.amdgcn.workitem.id.x()58 br i1 %cond, label %if, label %else59 60if:61 %gep.0 = getelementptr <32 x float>, ptr addrspace(1) %arg0, i32 %id62 %in.0 = load <32 x float>, ptr addrspace(1) %gep.0, align 12863 %mai.0 = call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 2.0, float 4.0, <32 x float> %in.0, i32 0, i32 0, i32 0)64 br label %endif65 66else:67 %gep.1 = getelementptr <32 x float>, ptr addrspace(1) %arg1, i32 %id68 %in.1 = load <32 x float>, ptr addrspace(1) %gep.1, align 12869 %mai.1 = call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 2.0, float 4.0, <32 x float> %in.1, i32 0, i32 0, i32 0)70 br label %endif71 72endif:73 %phi = phi <32 x float> [ %mai.0, %if ], [ %mai.1, %else ]74 call void asm sideeffect "; use $0", "a"(<32 x float> %phi)75 ret void76}77 78define amdgpu_kernel void @test_rewrite_mfma_copy_to_agpr_phi_loop(ptr addrspace(1) %arg0, ptr addrspace(1) %arg1, i32 %n) #0 {79; CHECK-LABEL: test_rewrite_mfma_copy_to_agpr_phi_loop:80; CHECK: ; %bb.0: ; %entry81; CHECK-NEXT: s_load_dwordx2 s[2:3], s[4:5], 0x082; CHECK-NEXT: s_load_dword s0, s[4:5], 0x1083; CHECK-NEXT: v_and_b32_e32 v0, 0x3ff, v084; CHECK-NEXT: v_lshlrev_b32_e32 v32, 7, v085; CHECK-NEXT: s_mov_b32 s1, 086; CHECK-NEXT: s_waitcnt lgkmcnt(0)87; CHECK-NEXT: global_load_dwordx4 v[28:31], v32, s[2:3] offset:11288; CHECK-NEXT: global_load_dwordx4 v[24:27], v32, s[2:3] offset:9689; CHECK-NEXT: global_load_dwordx4 v[20:23], v32, s[2:3] offset:8090; CHECK-NEXT: global_load_dwordx4 v[16:19], v32, s[2:3] offset:6491; CHECK-NEXT: global_load_dwordx4 v[12:15], v32, s[2:3] offset:4892; CHECK-NEXT: global_load_dwordx4 v[8:11], v32, s[2:3] offset:3293; CHECK-NEXT: global_load_dwordx4 v[4:7], v32, s[2:3] offset:1694; CHECK-NEXT: global_load_dwordx4 v[0:3], v32, s[2:3]95; CHECK-NEXT: v_mov_b32_e32 v64, 4.096; CHECK-NEXT: v_mov_b32_e32 v65, 2.097; CHECK-NEXT: .LBB1_1: ; %loop98; CHECK-NEXT: ; =>This Inner Loop Header: Depth=199; CHECK-NEXT: s_waitcnt vmcnt(0)100; CHECK-NEXT: s_nop 15101; CHECK-NEXT: v_mov_b64_e32 v[62:63], v[30:31]102; CHECK-NEXT: v_mov_b64_e32 v[60:61], v[28:29]103; CHECK-NEXT: v_mov_b64_e32 v[58:59], v[26:27]104; CHECK-NEXT: v_mov_b64_e32 v[56:57], v[24:25]105; CHECK-NEXT: v_mov_b64_e32 v[54:55], v[22:23]106; CHECK-NEXT: v_mov_b64_e32 v[52:53], v[20:21]107; CHECK-NEXT: v_mov_b64_e32 v[50:51], v[18:19]108; CHECK-NEXT: v_mov_b64_e32 v[48:49], v[16:17]109; CHECK-NEXT: v_mov_b64_e32 v[46:47], v[14:15]110; CHECK-NEXT: v_mov_b64_e32 v[44:45], v[12:13]111; CHECK-NEXT: v_mov_b64_e32 v[42:43], v[10:11]112; CHECK-NEXT: v_mov_b64_e32 v[40:41], v[8:9]113; CHECK-NEXT: v_mov_b64_e32 v[38:39], v[6:7]114; CHECK-NEXT: v_mov_b64_e32 v[36:37], v[4:5]115; CHECK-NEXT: v_mov_b64_e32 v[34:35], v[2:3]116; CHECK-NEXT: v_mov_b64_e32 v[32:33], v[0:1]117; CHECK-NEXT: s_add_i32 s1, s1, 1118; CHECK-NEXT: s_cmp_lt_u32 s1, s0119; CHECK-NEXT: v_mfma_f32_32x32x1_2b_f32 v[0:31], v65, v64, v[32:63]120; CHECK-NEXT: v_mfma_f32_32x32x1_2b_f32 v[0:31], v65, v64, v[0:31]121; CHECK-NEXT: s_cbranch_scc1 .LBB1_1122; CHECK-NEXT: ; %bb.2: ; %endif123; CHECK-NEXT: v_accvgpr_write_b32 a0, v32124; CHECK-NEXT: v_accvgpr_write_b32 a1, v33125; CHECK-NEXT: v_accvgpr_write_b32 a2, v34126; CHECK-NEXT: v_accvgpr_write_b32 a3, v35127; CHECK-NEXT: v_accvgpr_write_b32 a4, v36128; CHECK-NEXT: v_accvgpr_write_b32 a5, v37129; CHECK-NEXT: v_accvgpr_write_b32 a6, v38130; CHECK-NEXT: v_accvgpr_write_b32 a7, v39131; CHECK-NEXT: v_accvgpr_write_b32 a8, v40132; CHECK-NEXT: v_accvgpr_write_b32 a9, v41133; CHECK-NEXT: v_accvgpr_write_b32 a10, v42134; CHECK-NEXT: v_accvgpr_write_b32 a11, v43135; CHECK-NEXT: v_accvgpr_write_b32 a12, v44136; CHECK-NEXT: v_accvgpr_write_b32 a13, v45137; CHECK-NEXT: v_accvgpr_write_b32 a14, v46138; CHECK-NEXT: v_accvgpr_write_b32 a15, v47139; CHECK-NEXT: v_accvgpr_write_b32 a16, v48140; CHECK-NEXT: v_accvgpr_write_b32 a17, v49141; CHECK-NEXT: v_accvgpr_write_b32 a18, v50142; CHECK-NEXT: v_accvgpr_write_b32 a19, v51143; CHECK-NEXT: v_accvgpr_write_b32 a20, v52144; CHECK-NEXT: v_accvgpr_write_b32 a21, v53145; CHECK-NEXT: v_accvgpr_write_b32 a22, v54146; CHECK-NEXT: v_accvgpr_write_b32 a23, v55147; CHECK-NEXT: v_accvgpr_write_b32 a24, v56148; CHECK-NEXT: v_accvgpr_write_b32 a25, v57149; CHECK-NEXT: v_accvgpr_write_b32 a26, v58150; CHECK-NEXT: v_accvgpr_write_b32 a27, v59151; CHECK-NEXT: v_accvgpr_write_b32 a28, v60152; CHECK-NEXT: v_accvgpr_write_b32 a29, v61153; CHECK-NEXT: v_accvgpr_write_b32 a30, v62154; CHECK-NEXT: v_accvgpr_write_b32 a31, v63155; CHECK-NEXT: ;;#ASMSTART156; CHECK-NEXT: ; use a[0:31]157; CHECK-NEXT: ;;#ASMEND158; CHECK-NEXT: s_endpgm159entry:160 %id = call i32 @llvm.amdgcn.workitem.id.x()161 %gep.0 = getelementptr <32 x float>, ptr addrspace(1) %arg0, i32 %id162 %in.0 = load <32 x float>, ptr addrspace(1) %gep.0, align 128163 br label %loop164 165loop:166 %i.phi = phi i32 [ 0, %entry ], [ %i.inc, %loop ]167 %phi = phi <32 x float> [ %in.0, %entry ], [ %mai.1, %loop ]168 %mai.0 = call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 2.0, float 4.0, <32 x float> %phi, i32 0, i32 0, i32 0)169 %mai.1 = call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 2.0, float 4.0, <32 x float> %mai.0, i32 0, i32 0, i32 0)170 %i.inc = add i32 %i.phi, 1171 %loop.cond = icmp ult i32 %i.inc, %n172 br i1 %loop.cond, label %loop, label %endif173 174endif:175 call void asm sideeffect "; use $0", "a"(<32 x float> %phi)176 ret void177}178 179attributes #0 = { nounwind "amdgpu-flat-work-group-size"="1,256" "amdgpu-waves-per-eu"="1,1" }180