Skip to content

Commit 03ec026

Browse files
LuoYuankeLuoYuanke
committed
[X86] Add test cases for sub with select.
1 parent b0a0df9 commit 03ec026

File tree

1 file changed

+227
-0
lines changed

1 file changed

+227
-0
lines changed

‎llvm/test/CodeGen/X86/vector-bo-select.ll

Lines changed: 227 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -891,3 +891,230 @@ define <8 x double> @fdiv_v8f64_cast_cond(i8 noundef zeroext %pb, <8 x double> n
891891
%r = fdiv <8 x double> %x, %s
892892
ret <8 x double> %r
893893
}
894+
895+
define <4 x i32> @sub_v4i32(<4 x i1> %b, <4 x i32> noundef %x, <4 x i32> noundef %y) {
896+
; AVX2-LABEL: sub_v4i32:
897+
; AVX2: # %bb.0:
898+
; AVX2-NEXT: vpslld $31, %xmm0, %xmm0
899+
; AVX2-NEXT: vpsrad $31, %xmm0, %xmm0
900+
; AVX2-NEXT: vpand %xmm2, %xmm0, %xmm0
901+
; AVX2-NEXT: vpsubd %xmm0, %xmm1, %xmm0
902+
; AVX2-NEXT: retq
903+
;
904+
; AVX512F-LABEL: sub_v4i32:
905+
; AVX512F: # %bb.0:
906+
; AVX512F-NEXT: # kill: def $xmm2 killed $xmm2 def $zmm2
907+
; AVX512F-NEXT: vpslld $31, %xmm0, %xmm0
908+
; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
909+
; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1} {z}
910+
; AVX512F-NEXT: vpsubd %xmm0, %xmm1, %xmm0
911+
; AVX512F-NEXT: vzeroupper
912+
; AVX512F-NEXT: retq
913+
;
914+
; AVX512VL-LABEL: sub_v4i32:
915+
; AVX512VL: # %bb.0:
916+
; AVX512VL-NEXT: vpslld $31, %xmm0, %xmm0
917+
; AVX512VL-NEXT: vptestmd %xmm0, %xmm0, %k1
918+
; AVX512VL-NEXT: vmovdqa32 %xmm2, %xmm0 {%k1} {z}
919+
; AVX512VL-NEXT: vpsubd %xmm0, %xmm1, %xmm0
920+
; AVX512VL-NEXT: retq
921+
%s = select <4 x i1> %b, <4 x i32> %y, <4 x i32> zeroinitializer
922+
%r = sub <4 x i32> %x, %s
923+
ret <4 x i32> %r
924+
}
925+
926+
; negative test - sub is not commutative; there is no identity constant for operand 0
927+
928+
define <8 x i32> @sub_v8i32_commute(<8 x i1> %b, <8 x i32> noundef %x, <8 x i32> noundef %y) {
929+
; AVX2-LABEL: sub_v8i32_commute:
930+
; AVX2: # %bb.0:
931+
; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
932+
; AVX2-NEXT: vpslld $31, %ymm0, %ymm0
933+
; AVX2-NEXT: vpsrad $31, %ymm0, %ymm0
934+
; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0
935+
; AVX2-NEXT: vpsubd %ymm1, %ymm0, %ymm0
936+
; AVX2-NEXT: retq
937+
;
938+
; AVX512F-LABEL: sub_v8i32_commute:
939+
; AVX512F: # %bb.0:
940+
; AVX512F-NEXT: # kill: def $ymm2 killed $ymm2 def $zmm2
941+
; AVX512F-NEXT: vpmovsxwq %xmm0, %zmm0
942+
; AVX512F-NEXT: vpsllq $63, %zmm0, %zmm0
943+
; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k1
944+
; AVX512F-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1} {z}
945+
; AVX512F-NEXT: vpsubd %ymm1, %ymm0, %ymm0
946+
; AVX512F-NEXT: retq
947+
;
948+
; AVX512VL-LABEL: sub_v8i32_commute:
949+
; AVX512VL: # %bb.0:
950+
; AVX512VL-NEXT: vpmovsxwd %xmm0, %ymm0
951+
; AVX512VL-NEXT: vpslld $31, %ymm0, %ymm0
952+
; AVX512VL-NEXT: vptestmd %ymm0, %ymm0, %k1
953+
; AVX512VL-NEXT: vmovdqa32 %ymm2, %ymm0 {%k1} {z}
954+
; AVX512VL-NEXT: vpsubd %ymm1, %ymm0, %ymm0
955+
; AVX512VL-NEXT: retq
956+
%s = select <8 x i1> %b, <8 x i32> %y, <8 x i32> zeroinitializer
957+
%r = sub <8 x i32> %s, %x
958+
ret <8 x i32> %r
959+
}
960+
961+
define <16 x i32> @sub_v16i32_swap(<16 x i1> %b, <16 x i32> noundef %x, <16 x i32> noundef %y) {
962+
; AVX2-LABEL: sub_v16i32_swap:
963+
; AVX2: # %bb.0:
964+
; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
965+
; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
966+
; AVX2-NEXT: vpslld $31, %ymm5, %ymm5
967+
; AVX2-NEXT: vpsrad $31, %ymm5, %ymm5
968+
; AVX2-NEXT: vpandn %ymm4, %ymm5, %ymm4
969+
; AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
970+
; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
971+
; AVX2-NEXT: vpslld $31, %ymm0, %ymm0
972+
; AVX2-NEXT: vpsrad $31, %ymm0, %ymm0
973+
; AVX2-NEXT: vpandn %ymm3, %ymm0, %ymm0
974+
; AVX2-NEXT: vpsubd %ymm0, %ymm1, %ymm0
975+
; AVX2-NEXT: vpsubd %ymm4, %ymm2, %ymm1
976+
; AVX2-NEXT: retq
977+
;
978+
; AVX512-LABEL: sub_v16i32_swap:
979+
; AVX512: # %bb.0:
980+
; AVX512-NEXT: vpmovsxbd %xmm0, %zmm0
981+
; AVX512-NEXT: vpslld $31, %zmm0, %zmm0
982+
; AVX512-NEXT: vptestnmd %zmm0, %zmm0, %k1
983+
; AVX512-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1} {z}
984+
; AVX512-NEXT: vpsubd %zmm0, %zmm1, %zmm0
985+
; AVX512-NEXT: retq
986+
%s = select <16 x i1> %b, <16 x i32> zeroinitializer, <16 x i32> %y
987+
%r = sub <16 x i32> %x, %s
988+
ret <16 x i32> %r
989+
}
990+
991+
; negative test - sub is not commutative; there is no identity constant for operand 0
992+
993+
define <16 x i32> @sub_v16i32_commute_swap(<16 x i1> %b, <16 x i32> noundef %x, <16 x i32> noundef %y) {
994+
; AVX2-LABEL: sub_v16i32_commute_swap:
995+
; AVX2: # %bb.0:
996+
; AVX2-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
997+
; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
998+
; AVX2-NEXT: vpslld $31, %ymm5, %ymm5
999+
; AVX2-NEXT: vpsrad $31, %ymm5, %ymm5
1000+
; AVX2-NEXT: vpandn %ymm4, %ymm5, %ymm4
1001+
; AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
1002+
; AVX2-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
1003+
; AVX2-NEXT: vpslld $31, %ymm0, %ymm0
1004+
; AVX2-NEXT: vpsrad $31, %ymm0, %ymm0
1005+
; AVX2-NEXT: vpandn %ymm3, %ymm0, %ymm0
1006+
; AVX2-NEXT: vpsubd %ymm1, %ymm0, %ymm0
1007+
; AVX2-NEXT: vpsubd %ymm2, %ymm4, %ymm1
1008+
; AVX2-NEXT: retq
1009+
;
1010+
; AVX512-LABEL: sub_v16i32_commute_swap:
1011+
; AVX512: # %bb.0:
1012+
; AVX512-NEXT: vpmovsxbd %xmm0, %zmm0
1013+
; AVX512-NEXT: vpslld $31, %zmm0, %zmm0
1014+
; AVX512-NEXT: vptestnmd %zmm0, %zmm0, %k1
1015+
; AVX512-NEXT: vmovdqa32 %zmm2, %zmm0 {%k1} {z}
1016+
; AVX512-NEXT: vpsubd %zmm1, %zmm0, %zmm0
1017+
; AVX512-NEXT: retq
1018+
%s = select <16 x i1> %b, <16 x i32> zeroinitializer, <16 x i32> %y
1019+
%r = sub <16 x i32> %s, %x
1020+
ret <16 x i32> %r
1021+
}
1022+
1023+
define <8 x i32> @sub_v8i32_cast_cond(i8 noundef zeroext %pb, <8 x i32> noundef %x, <8 x i32> noundef %y) {
1024+
; AVX2-LABEL: sub_v8i32_cast_cond:
1025+
; AVX2: # %bb.0:
1026+
; AVX2-NEXT: vmovd %edi, %xmm2
1027+
; AVX2-NEXT: vpbroadcastb %xmm2, %ymm2
1028+
; AVX2-NEXT: vmovdqa {{.*#+}} ymm3 = [1,2,4,8,16,32,64,128]
1029+
; AVX2-NEXT: vpand %ymm3, %ymm2, %ymm2
1030+
; AVX2-NEXT: vpcmpeqd %ymm3, %ymm2, %ymm2
1031+
; AVX2-NEXT: vpxor %xmm3, %xmm3, %xmm3
1032+
; AVX2-NEXT: vblendvps %ymm2, %ymm1, %ymm3, %ymm1
1033+
; AVX2-NEXT: vpsubd %ymm1, %ymm0, %ymm0
1034+
; AVX2-NEXT: retq
1035+
;
1036+
; AVX512F-LABEL: sub_v8i32_cast_cond:
1037+
; AVX512F: # %bb.0:
1038+
; AVX512F-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
1039+
; AVX512F-NEXT: kmovw %edi, %k1
1040+
; AVX512F-NEXT: vmovdqa32 %zmm1, %zmm1 {%k1} {z}
1041+
; AVX512F-NEXT: vpsubd %ymm1, %ymm0, %ymm0
1042+
; AVX512F-NEXT: retq
1043+
;
1044+
; AVX512VL-LABEL: sub_v8i32_cast_cond:
1045+
; AVX512VL: # %bb.0:
1046+
; AVX512VL-NEXT: kmovw %edi, %k1
1047+
; AVX512VL-NEXT: vmovdqa32 %ymm1, %ymm1 {%k1} {z}
1048+
; AVX512VL-NEXT: vpsubd %ymm1, %ymm0, %ymm0
1049+
; AVX512VL-NEXT: retq
1050+
%b = bitcast i8 %pb to <8 x i1>
1051+
%s = select <8 x i1> %b, <8 x i32> %y, <8 x i32> zeroinitializer
1052+
%r = sub <8 x i32> %x, %s
1053+
ret <8 x i32> %r
1054+
}
1055+
1056+
define <8 x i64> @sub_v8i64_cast_cond(i8 noundef zeroext %pb, <8 x i64> noundef %x, <8 x i64> noundef %y) {
1057+
; AVX2-LABEL: sub_v8i64_cast_cond:
1058+
; AVX2: # %bb.0:
1059+
; AVX2-NEXT: movl %edi, %eax
1060+
; AVX2-NEXT: shrb %al
1061+
; AVX2-NEXT: andb $1, %al
1062+
; AVX2-NEXT: movzbl %al, %eax
1063+
; AVX2-NEXT: movl %edi, %ecx
1064+
; AVX2-NEXT: andb $1, %cl
1065+
; AVX2-NEXT: movzbl %cl, %ecx
1066+
; AVX2-NEXT: vmovd %ecx, %xmm4
1067+
; AVX2-NEXT: vpinsrb $2, %eax, %xmm4, %xmm4
1068+
; AVX2-NEXT: movl %edi, %eax
1069+
; AVX2-NEXT: shrb $2, %al
1070+
; AVX2-NEXT: andb $1, %al
1071+
; AVX2-NEXT: movzbl %al, %eax
1072+
; AVX2-NEXT: vpinsrb $4, %eax, %xmm4, %xmm4
1073+
; AVX2-NEXT: movl %edi, %eax
1074+
; AVX2-NEXT: shrb $3, %al
1075+
; AVX2-NEXT: andb $1, %al
1076+
; AVX2-NEXT: movzbl %al, %eax
1077+
; AVX2-NEXT: vpinsrb $6, %eax, %xmm4, %xmm4
1078+
; AVX2-NEXT: movl %edi, %eax
1079+
; AVX2-NEXT: shrb $4, %al
1080+
; AVX2-NEXT: andb $1, %al
1081+
; AVX2-NEXT: movzbl %al, %eax
1082+
; AVX2-NEXT: vpinsrb $8, %eax, %xmm4, %xmm5
1083+
; AVX2-NEXT: movl %edi, %eax
1084+
; AVX2-NEXT: shrb $5, %al
1085+
; AVX2-NEXT: andb $1, %al
1086+
; AVX2-NEXT: movzbl %al, %eax
1087+
; AVX2-NEXT: vpinsrb $10, %eax, %xmm5, %xmm5
1088+
; AVX2-NEXT: movl %edi, %eax
1089+
; AVX2-NEXT: shrb $6, %al
1090+
; AVX2-NEXT: andb $1, %al
1091+
; AVX2-NEXT: movzbl %al, %eax
1092+
; AVX2-NEXT: vpinsrb $12, %eax, %xmm5, %xmm5
1093+
; AVX2-NEXT: shrb $7, %dil
1094+
; AVX2-NEXT: movzbl %dil, %eax
1095+
; AVX2-NEXT: vpinsrb $14, %eax, %xmm5, %xmm5
1096+
; AVX2-NEXT: vpunpckhwd {{.*#+}} xmm5 = xmm5[4,4,5,5,6,6,7,7]
1097+
; AVX2-NEXT: vpslld $31, %xmm5, %xmm5
1098+
; AVX2-NEXT: vpsrad $31, %xmm5, %xmm5
1099+
; AVX2-NEXT: vpmovsxdq %xmm5, %ymm5
1100+
; AVX2-NEXT: vpand %ymm3, %ymm5, %ymm3
1101+
; AVX2-NEXT: vpmovzxwd {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero
1102+
; AVX2-NEXT: vpslld $31, %xmm4, %xmm4
1103+
; AVX2-NEXT: vpsrad $31, %xmm4, %xmm4
1104+
; AVX2-NEXT: vpmovsxdq %xmm4, %ymm4
1105+
; AVX2-NEXT: vpand %ymm2, %ymm4, %ymm2
1106+
; AVX2-NEXT: vpsubq %ymm2, %ymm0, %ymm0
1107+
; AVX2-NEXT: vpsubq %ymm3, %ymm1, %ymm1
1108+
; AVX2-NEXT: retq
1109+
;
1110+
; AVX512-LABEL: sub_v8i64_cast_cond:
1111+
; AVX512: # %bb.0:
1112+
; AVX512-NEXT: kmovw %edi, %k1
1113+
; AVX512-NEXT: vmovdqa64 %zmm1, %zmm1 {%k1} {z}
1114+
; AVX512-NEXT: vpsubq %zmm1, %zmm0, %zmm0
1115+
; AVX512-NEXT: retq
1116+
%b = bitcast i8 %pb to <8 x i1>
1117+
%s = select <8 x i1> %b, <8 x i64> %y, <8 x i64> zeroinitializer
1118+
%r = sub <8 x i64> %x, %s
1119+
ret <8 x i64> %r
1120+
}

0 commit comments

Comments
 (0)