153 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefix=SDAG %s3; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefix=GISEL %s4 5; Check for verifier error due to trying to save and restore SCC6; around a waterfall looop when it was never defined. We have to get7; an accurate liveness at the use point and cannot rely on an8; imprecise maybe-live query.9 10define void @issue92561(ptr addrspace(1) %arg) {11; SDAG-LABEL: issue92561:12; SDAG: ; %bb.0: ; %bb13; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)14; SDAG-NEXT: s_clause 0x115; SDAG-NEXT: global_load_b128 v[4:7], v[0:1], off offset:1616; SDAG-NEXT: global_load_b128 v[0:3], v[0:1], off17; SDAG-NEXT: v_mov_b32_e32 v8, 018; SDAG-NEXT: s_mov_b32 s12, 019; SDAG-NEXT: s_mov_b32 s3, exec_lo20; SDAG-NEXT: s_mov_b32 s13, s1221; SDAG-NEXT: s_mov_b32 s14, s1222; SDAG-NEXT: s_mov_b32 s15, s1223; SDAG-NEXT: s_waitcnt vmcnt(0)24; SDAG-NEXT: .LBB0_1: ; =>This Inner Loop Header: Depth=125; SDAG-NEXT: v_readfirstlane_b32 s4, v026; SDAG-NEXT: v_readfirstlane_b32 s5, v127; SDAG-NEXT: v_readfirstlane_b32 s6, v228; SDAG-NEXT: v_readfirstlane_b32 s7, v329; SDAG-NEXT: v_readfirstlane_b32 s8, v430; SDAG-NEXT: v_readfirstlane_b32 s9, v531; SDAG-NEXT: v_readfirstlane_b32 s10, v632; SDAG-NEXT: v_readfirstlane_b32 s11, v733; SDAG-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]34; SDAG-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]35; SDAG-NEXT: v_cmp_eq_u64_e64 s1, s[8:9], v[4:5]36; SDAG-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)37; SDAG-NEXT: v_cmp_eq_u64_e64 s2, s[10:11], v[6:7]38; SDAG-NEXT: s_and_b32 s0, vcc_lo, s039; SDAG-NEXT: s_and_b32 s0, s0, s140; SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)41; SDAG-NEXT: s_and_b32 s0, s0, s242; SDAG-NEXT: s_and_saveexec_b32 s0, s043; SDAG-NEXT: image_sample_c_lz v9, [v8, v8, v8, v8], s[4:11], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY44; SDAG-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr745; SDAG-NEXT: s_xor_b32 exec_lo, exec_lo, s046; SDAG-NEXT: s_cbranch_execnz .LBB0_147; SDAG-NEXT: ; %bb.2:48; SDAG-NEXT: s_mov_b32 exec_lo, s349; SDAG-NEXT: v_dual_mov_b32 v0, 0x7fc00000 :: v_dual_mov_b32 v1, 1.050; SDAG-NEXT: s_mov_b32 s0, s1251; SDAG-NEXT: s_mov_b32 s1, s1252; SDAG-NEXT: s_mov_b32 s2, s1253; SDAG-NEXT: s_mov_b32 s3, s1254; SDAG-NEXT: s_mov_b32 s4, s1255; SDAG-NEXT: s_mov_b32 s5, s1256; SDAG-NEXT: s_mov_b32 s6, s1257; SDAG-NEXT: s_mov_b32 s7, s1258; SDAG-NEXT: s_clause 0x259; SDAG-NEXT: image_sample_c_lz v0, [v8, v8, v0, v8], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY60; SDAG-NEXT: image_sample_c_lz v2, [v8, v8, v8, v8], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY61; SDAG-NEXT: image_sample_c_lz v1, [v8, v1, v8, v8], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY62; SDAG-NEXT: s_waitcnt vmcnt(2)63; SDAG-NEXT: v_dual_add_f32 v0, v9, v0 :: v_dual_mov_b32 v9, v864; SDAG-NEXT: s_waitcnt vmcnt(0)65; SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)66; SDAG-NEXT: v_add_f32_e32 v0, v1, v067; SDAG-NEXT: v_add_f32_e32 v0, v2, v068; SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)69; SDAG-NEXT: v_dual_mul_f32 v7, 0x3e800000, v0 :: v_dual_mov_b32 v0, 070; SDAG-NEXT: image_store v[7:9], [v0, v0], s[0:7] dim:SQ_RSRC_IMG_2D unorm71; SDAG-NEXT: s_setpc_b64 s[30:31]72;73; GISEL-LABEL: issue92561:74; GISEL: ; %bb.0: ; %bb75; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)76; GISEL-NEXT: s_clause 0x177; GISEL-NEXT: global_load_b128 v[4:7], v[0:1], off78; GISEL-NEXT: global_load_b128 v[0:3], v[0:1], off offset:1679; GISEL-NEXT: v_mov_b32_e32 v8, 080; GISEL-NEXT: s_mov_b32 s20, 081; GISEL-NEXT: s_mov_b32 s3, exec_lo82; GISEL-NEXT: s_mov_b32 s21, s2083; GISEL-NEXT: s_mov_b32 s22, s2084; GISEL-NEXT: s_mov_b32 s23, s2085; GISEL-NEXT: s_mov_b32 s4, s2086; GISEL-NEXT: s_mov_b32 s5, s2087; GISEL-NEXT: s_mov_b32 s6, s2088; GISEL-NEXT: s_mov_b32 s7, s2089; GISEL-NEXT: s_mov_b32 s8, s2090; GISEL-NEXT: s_mov_b32 s9, s2091; GISEL-NEXT: s_mov_b32 s10, s2092; GISEL-NEXT: s_mov_b32 s11, s2093; GISEL-NEXT: s_waitcnt vmcnt(0)94; GISEL-NEXT: .LBB0_1: ; =>This Inner Loop Header: Depth=195; GISEL-NEXT: v_readfirstlane_b32 s12, v496; GISEL-NEXT: v_readfirstlane_b32 s13, v597; GISEL-NEXT: v_readfirstlane_b32 s14, v698; GISEL-NEXT: v_readfirstlane_b32 s15, v799; GISEL-NEXT: v_readfirstlane_b32 s16, v0100; GISEL-NEXT: v_readfirstlane_b32 s17, v1101; GISEL-NEXT: v_readfirstlane_b32 s18, v2102; GISEL-NEXT: v_readfirstlane_b32 s19, v3103; GISEL-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[12:13], v[4:5]104; GISEL-NEXT: v_cmp_eq_u64_e64 s0, s[14:15], v[6:7]105; GISEL-NEXT: v_cmp_eq_u64_e64 s1, s[16:17], v[0:1]106; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)107; GISEL-NEXT: v_cmp_eq_u64_e64 s2, s[18:19], v[2:3]108; GISEL-NEXT: s_and_b32 s0, vcc_lo, s0109; GISEL-NEXT: s_and_b32 s0, s0, s1110; GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)111; GISEL-NEXT: s_and_b32 s0, s0, s2112; GISEL-NEXT: s_and_saveexec_b32 s0, s0113; GISEL-NEXT: image_sample_c_lz v9, [v8, v8, v8, v8], s[12:19], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY114; GISEL-NEXT: ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7115; GISEL-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3116; GISEL-NEXT: ; implicit-def: $vgpr8117; GISEL-NEXT: s_xor_b32 exec_lo, exec_lo, s0118; GISEL-NEXT: s_cbranch_execnz .LBB0_1119; GISEL-NEXT: ; %bb.2:120; GISEL-NEXT: s_mov_b32 exec_lo, s3121; GISEL-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, 0x7fc00000122; GISEL-NEXT: v_mov_b32_e32 v2, 1.0123; GISEL-NEXT: s_clause 0x2124; GISEL-NEXT: image_sample_c_lz v0, [v1, v1, v0, v1], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY125; GISEL-NEXT: image_sample_c_lz v2, [v1, v2, v1, v1], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY126; GISEL-NEXT: image_sample_c_lz v3, [v1, v1, v1, v1], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY127; GISEL-NEXT: s_waitcnt vmcnt(2)128; GISEL-NEXT: v_add_f32_e32 v0, v9, v0129; GISEL-NEXT: s_waitcnt vmcnt(1)130; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)131; GISEL-NEXT: v_add_f32_e32 v0, v2, v0132; GISEL-NEXT: v_mov_b32_e32 v2, v1133; GISEL-NEXT: s_waitcnt vmcnt(0)134; GISEL-NEXT: v_add_f32_e32 v0, v3, v0135; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)136; GISEL-NEXT: v_mul_f32_e32 v0, 0x3e800000, v0137; GISEL-NEXT: image_store v[0:2], [v1, v1], s[4:11] dim:SQ_RSRC_IMG_2D unorm138; GISEL-NEXT: s_setpc_b64 s[30:31]139bb:140 %descriptor = load <8 x i32>, ptr addrspace(1) %arg, align 32141 %needs.waterfall = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0.0, float 0.0, <8 x i32> %descriptor, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)142 %i2 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0x7FF8000000000000, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)143 %i3 = fadd float %needs.waterfall, %i2144 %i4 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 1.000000e+00, float 0.0, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)145 %i5 = fadd float %i4, %i3146 %i6 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0.0, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)147 %i7 = fadd float %i6, %i5148 %i8 = fmul float %i7, 2.500000e-01149 %i9 = insertelement <3 x float> zeroinitializer, float %i8, i64 0150 call void @llvm.amdgcn.image.store.2d.v3f32.i32(<3 x float> %i9, i32 0, i32 0, i32 0, <8 x i32> zeroinitializer, i32 0, i32 0)151 ret void152}153