187 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=hexagon -hexagon-bit=false < %s | FileCheck %s3 4define bfloat @load_scalar_bf(ptr %addr) {5; CHECK-LABEL: load_scalar_bf:6; CHECK: .cfi_startproc7; CHECK-NEXT: // %bb.0: // %entry8; CHECK-NEXT: {9; CHECK-NEXT: [[REG:r[0-9]+]] = memuh(r0+#0)10; CHECK-NEXT: }11; CHECK-NEXT: {12; CHECK-NEXT: r0 = aslh([[REG]])13; CHECK-NEXT: jumpr r3114; CHECK-NEXT: }15 16entry:17 %0 = load bfloat, ptr %addr, align 218 ret bfloat %019}20 21define void @store_scalar_bf(bfloat %v, ptr %addr) {22; CHECK-LABEL: store_scalar_bf:23; CHECK: .cfi_startproc24; CHECK-NEXT: // %bb.0: // %entry25; CHECK-NEXT: {26; CHECK-NEXT: [[R_A:r[0-9]+]] = ##13107127; CHECK-NEXT: [[R_B:r[0-9]+]] = ##3276828; CHECK-NEXT: }29; CHECK-NEXT: {30; CHECK-NEXT: [[R_C:r[0-9]+]] = ##6553531; CHECK-NEXT: [[R_D:r[0-9]+]] = asrh(r0)32; CHECK-NEXT: [[R_A]] = and(r0,[[R_A]])33; CHECK-NEXT: }34; CHECK-NEXT: {35; CHECK-NEXT: [[P0:p[0-9]+]] = sfclass(r0,#16)36; CHECK-NEXT: [[R_E:r[0-9]+]] = and(r0,[[R_B]])37; CHECK-NEXT: [[P1:p[0-9]+]] = cmp.eq([[R_A]],[[R_B]])38; CHECK-NEXT: }39; CHECK-NEXT: {40; CHECK-NEXT: r0 = add(r0,[[R_E]])41; CHECK-NEXT: }42; CHECK-NEXT: {43; CHECK-NEXT: r0 = asrh(r0)44; CHECK-NEXT: if ([[P1]]) [[R_B]] = and([[R_D]],[[R_C]])45; CHECK-NEXT: }46; CHECK-NEXT: {47; CHECK-NEXT: if (![[P1]]) [[R_B]] = and(r0,[[R_C]])48; CHECK-NEXT: }49; CHECK-NEXT: {50; CHECK-NEXT: if ([[P0]]) [[R_B]] = ##3276751; CHECK-NEXT: }52; CHECK-NEXT: {53; CHECK-NEXT: jumpr r3154; CHECK-NEXT: memh(r1+#0) = [[R_B]]55; CHECK-NEXT: }56 57entry:58 store bfloat %v, ptr %addr, align 259 ret void60}61 62define bfloat @sum(bfloat %a, bfloat %b) #0 {63; CHECK-LABEL: sum:64; CHECK: .cfi_startproc65; CHECK-NEXT: // %bb.0: // %entry66; CHECK-NEXT: {67; CHECK-NEXT: [[R3:r[0-9]+]] = ##3276868; CHECK-NEXT: [[R4:r[0-9]+]] = ##13107169; CHECK-NEXT: }70; CHECK-NEXT: {71; CHECK-NEXT: [[R2:r[0-9]+]] = ##6553572; CHECK-NEXT: [[R6:r[0-9]+]] = and(r0,[[R3]])73; CHECK-NEXT: [[R5:r[0-9]+]] = and(r0,[[R4]])74; CHECK-NEXT: }75; CHECK-NEXT: {76; CHECK-NEXT: [[R8:r[0-9]+]] = and(r1,[[R3]])77; CHECK-NEXT: [[R7:r[0-9]+]] = and(r1,[[R4]])78; CHECK-NEXT: [[R6]] = add(r0,[[R6]])79; CHECK-NEXT: [[P0:p[0-9]+]] = cmp.eq([[R5]],[[R3]])80; CHECK-NEXT: }81; CHECK-NEXT: {82; CHECK-NEXT: [[R5]] = asrh(r0)83; CHECK-NEXT: [[P1:p[0-9]+]] = cmp.eq([[R7]],[[R3]])84; CHECK-NEXT: [[R7]] = asrh(r1)85; CHECK-NEXT: [[R8]] = add(r1,[[R8]])86; CHECK-NEXT: }87; CHECK-NEXT: {88; CHECK-NEXT: [[R6]] = asrh([[R6]])89; CHECK-NEXT: if ([[P0]]) [[R5]] = and([[R5]],[[R2]])90; CHECK-NEXT: }91; CHECK-NEXT: {92; CHECK-NEXT: [[P0]] = sfclass(r0,#16)93; CHECK-NEXT: if (![[P0]]) [[R5]] = and([[R6]],[[R2]])94; CHECK-NEXT: [[R6]] = asrh([[R8]])95; CHECK-NEXT: if ([[P1]]) [[R7]] = and([[R7]],[[R2]])96; CHECK-NEXT: }97; CHECK-NEXT: {98; CHECK-NEXT: [[P1]] = sfclass(r1,#16)99; CHECK-NEXT: [[R0:r[0-9]+]] = #32767100; CHECK-NEXT: if (![[P1]]) [[R7]] = and([[R6]],[[R2]])101; CHECK-NEXT: }102; CHECK-NEXT: {103; CHECK-NEXT: if ([[P0]]) [[R5]] = add([[R0]],#0)104; CHECK-NEXT: if ([[P1]]) [[R7]] = add([[R0]],#0)105; CHECK-NEXT: }106; CHECK-NEXT: {107; CHECK-NEXT: [[R6]] = aslh([[R7]])108; CHECK-NEXT: [[R5]] = aslh([[R5]])109; CHECK-NEXT: }110; CHECK-NEXT: {111; CHECK-NEXT: [[R1:r[0-9]+]] = sfadd([[R5]],[[R6]])112; CHECK-NEXT: }113; CHECK-NEXT: {114; CHECK-NEXT: [[R5]] = and([[R1]],[[R3]])115; CHECK-NEXT: [[R4]] = and([[R1]],[[R4]])116; CHECK-NEXT: [[R6]] = asrh([[R1]])117; CHECK-NEXT: }118; CHECK-NEXT: {119; CHECK-NEXT: [[P0]] = cmp.eq([[R4]],[[R3]])120; CHECK-NEXT: [[R5]] = add([[R1]],[[R5]])121; CHECK-NEXT: }122; CHECK-NEXT: {123; CHECK-NEXT: [[R3]] = asrh([[R5]])124; CHECK-NEXT: if ([[P0]]) [[R4]] = and([[R6]],[[R2]])125; CHECK-NEXT: }126; CHECK-NEXT: {127; CHECK-NEXT: [[P0]] = sfclass([[R1]],#16)128; CHECK-NEXT: if (![[P0]]) [[R4]] = and([[R3]],[[R2]])129; CHECK-NEXT: }130; CHECK-NEXT: {131; CHECK-NEXT: if ([[P0]]) [[R4]] = add([[R0]],#0)132; CHECK-NEXT: }133; CHECK-NEXT: {134; CHECK-NEXT: r0 = aslh([[R4]])135; CHECK-NEXT: jumpr r31136; CHECK-NEXT: }137 138entry:139 %add = fadd bfloat %a, %b140 ret bfloat %add141}142 143; Function Attrs: noinline nounwind optnone144define dso_local i32 @double_bf16(bfloat %a, bfloat %b) #0 {145; CHECK-LABEL: double_bf16:146; CHECK: .cfi_startproc147; CHECK-NEXT: // %bb.0: // %entry148; CHECK-NEXT: {149; CHECK-NEXT: [[PAIR_A:r[0-9]+:[0-9]+]] = convert_sf2df(r0)150; CHECK-NEXT: [[PAIR_B:r[0-9]+:[0-9]+]] = convert_sf2df(r1)151; CHECK-NEXT: [[SP:r[0-9]+]] = add([[SP]],#-16)152; CHECK-NEXT: }153; CHECK-NEXT: {154; CHECK-NEXT: [[RESULT:r[0-9]+:[0-9]+]] = dfadd([[PAIR_A]],[[PAIR_B]])155; CHECK-NEXT: }156; CHECK-NEXT: {157; CHECK-NEXT: memd([[SP]]+#8) = [[RESULT]]158; CHECK-NEXT: }159; CHECK-NEXT: {160; CHECK-NEXT: [[TMP:r[0-9]+]] = convert_df2sf([[RESULT]])161; CHECK-NEXT: }162; CHECK-NEXT: {163; CHECK-NEXT: [[TMP1:r[0-9]+]] = lsr([[TMP]],#16)164; CHECK-NEXT: [[R0:r[0-9]+]] = and([[TMP]],##-65536)165; CHECK-NEXT: memh([[SP]]+#6) = [[TMP1]].new166; CHECK-NEXT: }167; CHECK-NEXT: {168; CHECK-NEXT: [[R0]] = convert_sf2w([[R0]]):chop169; CHECK-NEXT: jumpr r31170; CHECK-NEXT: [[SP]] = add([[SP]],#16)171; CHECK-NEXT: }172 173entry:174 %c = alloca double, align 8175 %d = alloca bfloat, align 2176 %conv = fpext bfloat %a to double177 %conv1 = fpext bfloat %b to double178 %add = fadd double %conv, %conv1179 store double %add, ptr %c, align 8180 %2 = load double, ptr %c, align 8181 %conv2 = fptrunc double %2 to bfloat182 store bfloat %conv2, ptr %d, align 2183 %3 = load bfloat, ptr %d, align 2184 %conv3 = fptosi bfloat %3 to i32185 ret i32 %conv3186}187