302 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=CHECK,AVX13; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=CHECK,AVX24 5;6; testz(~X,Y) -> testc(X,Y)7;8 9define i32 @ptestz_256_invert0(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {10; CHECK-LABEL: ptestz_256_invert0:11; CHECK: # %bb.0:12; CHECK-NEXT: movl %edi, %eax13; CHECK-NEXT: vptest %ymm1, %ymm014; CHECK-NEXT: cmovael %esi, %eax15; CHECK-NEXT: vzeroupper16; CHECK-NEXT: retq17 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>18 %t2 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %t1, <4 x i64> %d)19 %t3 = icmp ne i32 %t2, 020 %t4 = select i1 %t3, i32 %a, i32 %b21 ret i32 %t422}23 24;25; testz(X,~Y) -> testc(Y,X)26;27 28define i32 @ptestz_256_invert1(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {29; CHECK-LABEL: ptestz_256_invert1:30; CHECK: # %bb.0:31; CHECK-NEXT: movl %edi, %eax32; CHECK-NEXT: vptest %ymm0, %ymm133; CHECK-NEXT: cmovael %esi, %eax34; CHECK-NEXT: vzeroupper35; CHECK-NEXT: retq36 %t1 = xor <4 x i64> %d, <i64 -1, i64 -1, i64 -1, i64 -1>37 %t2 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %c, <4 x i64> %t1)38 %t3 = icmp ne i32 %t2, 039 %t4 = select i1 %t3, i32 %a, i32 %b40 ret i32 %t441}42 43;44; testc(~X,Y) -> testz(X,Y)45;46 47define i32 @ptestc_256_invert0(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {48; CHECK-LABEL: ptestc_256_invert0:49; CHECK: # %bb.0:50; CHECK-NEXT: movl %edi, %eax51; CHECK-NEXT: vptest %ymm1, %ymm052; CHECK-NEXT: cmovnel %esi, %eax53; CHECK-NEXT: vzeroupper54; CHECK-NEXT: retq55 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>56 %t2 = call i32 @llvm.x86.avx.ptestc.256(<4 x i64> %t1, <4 x i64> %d)57 %t3 = icmp ne i32 %t2, 058 %t4 = select i1 %t3, i32 %a, i32 %b59 ret i32 %t460}61 62;63; testnzc(~X,Y) -> testnzc(X,Y)64;65 66define i32 @ptestnzc_256_invert0(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {67; CHECK-LABEL: ptestnzc_256_invert0:68; CHECK: # %bb.0:69; CHECK-NEXT: movl %edi, %eax70; CHECK-NEXT: vptest %ymm1, %ymm071; CHECK-NEXT: cmovbel %esi, %eax72; CHECK-NEXT: vzeroupper73; CHECK-NEXT: retq74 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>75 %t2 = call i32 @llvm.x86.avx.ptestnzc.256(<4 x i64> %t1, <4 x i64> %d)76 %t3 = icmp ne i32 %t2, 077 %t4 = select i1 %t3, i32 %a, i32 %b78 ret i32 %t479}80 81define i32 @ptestnzc_256_invert0_commute(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {82; CHECK-LABEL: ptestnzc_256_invert0_commute:83; CHECK: # %bb.0:84; CHECK-NEXT: movl %edi, %eax85; CHECK-NEXT: vptest %ymm1, %ymm086; CHECK-NEXT: cmoval %esi, %eax87; CHECK-NEXT: vzeroupper88; CHECK-NEXT: retq89 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>90 %t2 = call i32 @llvm.x86.avx.ptestnzc.256(<4 x i64> %t1, <4 x i64> %d)91 %t3 = icmp eq i32 %t2, 092 %t4 = select i1 %t3, i32 %a, i32 %b93 ret i32 %t494}95 96;97; testc(X,~X) -> testc(X,-1)98;99 100define i32 @ptestc_256_not(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {101; AVX1-LABEL: ptestc_256_not:102; AVX1: # %bb.0:103; AVX1-NEXT: movl %edi, %eax104; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1105; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1106; AVX1-NEXT: vptest %ymm1, %ymm0107; AVX1-NEXT: cmovael %esi, %eax108; AVX1-NEXT: vzeroupper109; AVX1-NEXT: retq110;111; AVX2-LABEL: ptestc_256_not:112; AVX2: # %bb.0:113; AVX2-NEXT: movl %edi, %eax114; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1115; AVX2-NEXT: vptest %ymm1, %ymm0116; AVX2-NEXT: cmovael %esi, %eax117; AVX2-NEXT: vzeroupper118; AVX2-NEXT: retq119 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>120 %t2 = call i32 @llvm.x86.avx.ptestc.256(<4 x i64> %c, <4 x i64> %t1)121 %t3 = icmp ne i32 %t2, 0122 %t4 = select i1 %t3, i32 %a, i32 %b123 ret i32 %t4124}125 126;127; testz(AND(X,Y),AND(X,Y)) -> testz(X,Y)128;129 130define i32 @ptestz_256_and(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {131; CHECK-LABEL: ptestz_256_and:132; CHECK: # %bb.0:133; CHECK-NEXT: movl %edi, %eax134; CHECK-NEXT: vptest %ymm1, %ymm0135; CHECK-NEXT: cmovel %esi, %eax136; CHECK-NEXT: vzeroupper137; CHECK-NEXT: retq138 %t1 = and <4 x i64> %c, %d139 %t2 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %t1, <4 x i64> %t1)140 %t3 = icmp eq i32 %t2, 0141 %t4 = select i1 %t3, i32 %a, i32 %b142 ret i32 %t4143}144 145;146; testz(AND(~X,Y),AND(~X,Y)) -> testc(X,Y)147;148 149define i32 @ptestz_256_andc(<4 x i64> %c, <4 x i64> %d, i32 %a, i32 %b) {150; CHECK-LABEL: ptestz_256_andc:151; CHECK: # %bb.0:152; CHECK-NEXT: movl %edi, %eax153; CHECK-NEXT: vptest %ymm1, %ymm0154; CHECK-NEXT: cmovbl %esi, %eax155; CHECK-NEXT: vzeroupper156; CHECK-NEXT: retq157 %t1 = xor <4 x i64> %c, <i64 -1, i64 -1, i64 -1, i64 -1>158 %t2 = and <4 x i64> %t1, %d159 %t3 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %t2, <4 x i64> %t2)160 %t4 = icmp eq i32 %t3, 0161 %t5 = select i1 %t4, i32 %a, i32 %b162 ret i32 %t5163}164 165;166; testz(-1,X) -> testz(X,X)167;168 169define i32 @ptestz_256_allones0(<4 x i64> %c, i32 %a, i32 %b) {170; CHECK-LABEL: ptestz_256_allones0:171; CHECK: # %bb.0:172; CHECK-NEXT: movl %edi, %eax173; CHECK-NEXT: vptest %ymm0, %ymm0174; CHECK-NEXT: cmovnel %esi, %eax175; CHECK-NEXT: vzeroupper176; CHECK-NEXT: retq177 %t1 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>, <4 x i64> %c)178 %t2 = icmp ne i32 %t1, 0179 %t3 = select i1 %t2, i32 %a, i32 %b180 ret i32 %t3181}182 183;184; testz(X,-1) -> testz(X,X)185;186 187define i32 @ptestz_256_allones1(<4 x i64> %c, i32 %a, i32 %b) {188; CHECK-LABEL: ptestz_256_allones1:189; CHECK: # %bb.0:190; CHECK-NEXT: movl %edi, %eax191; CHECK-NEXT: vptest %ymm0, %ymm0192; CHECK-NEXT: cmovnel %esi, %eax193; CHECK-NEXT: vzeroupper194; CHECK-NEXT: retq195 %t1 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %c, <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>)196 %t2 = icmp ne i32 %t1, 0197 %t3 = select i1 %t2, i32 %a, i32 %b198 ret i32 %t3199}200 201;202; testz(ashr(X,bw-1),-1) -> testpd/testps/pmovmskb(X)203;204 205define i32 @ptestz_v8i32_signbits(<8 x i32> %c, i32 %a, i32 %b) {206; AVX1-LABEL: ptestz_v8i32_signbits:207; AVX1: # %bb.0:208; AVX1-NEXT: movl %edi, %eax209; AVX1-NEXT: vpsrad $31, %xmm0, %xmm1210; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0211; AVX1-NEXT: vpsrad $31, %xmm0, %xmm0212; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0213; AVX1-NEXT: vptest %ymm0, %ymm0214; AVX1-NEXT: cmovnel %esi, %eax215; AVX1-NEXT: vzeroupper216; AVX1-NEXT: retq217;218; AVX2-LABEL: ptestz_v8i32_signbits:219; AVX2: # %bb.0:220; AVX2-NEXT: movl %edi, %eax221; AVX2-NEXT: vtestps %ymm0, %ymm0222; AVX2-NEXT: cmovnel %esi, %eax223; AVX2-NEXT: vzeroupper224; AVX2-NEXT: retq225 %t1 = ashr <8 x i32> %c, <i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31>226 %t2 = bitcast <8 x i32> %t1 to <4 x i64>227 %t3 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %t2, <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>)228 %t4 = icmp ne i32 %t3, 0229 %t5 = select i1 %t4, i32 %a, i32 %b230 ret i32 %t5231}232 233define i32 @ptestz_v32i8_signbits(<32 x i8> %c, i32 %a, i32 %b) {234; AVX1-LABEL: ptestz_v32i8_signbits:235; AVX1: # %bb.0:236; AVX1-NEXT: movl %edi, %eax237; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1238; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2239; AVX1-NEXT: vpcmpgtb %xmm1, %xmm2, %xmm1240; AVX1-NEXT: vpcmpgtb %xmm0, %xmm2, %xmm0241; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0242; AVX1-NEXT: vptest %ymm0, %ymm0243; AVX1-NEXT: cmovnel %esi, %eax244; AVX1-NEXT: vzeroupper245; AVX1-NEXT: retq246;247; AVX2-LABEL: ptestz_v32i8_signbits:248; AVX2: # %bb.0:249; AVX2-NEXT: movl %edi, %eax250; AVX2-NEXT: vpmovmskb %ymm0, %ecx251; AVX2-NEXT: testl %ecx, %ecx252; AVX2-NEXT: cmovnel %esi, %eax253; AVX2-NEXT: vzeroupper254; AVX2-NEXT: retq255 %t1 = ashr <32 x i8> %c, <i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7>256 %t2 = bitcast <32 x i8> %t1 to <4 x i64>257 %t3 = call i32 @llvm.x86.avx.ptestz.256(<4 x i64> %t2, <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>)258 %t4 = icmp ne i32 %t3, 0259 %t5 = select i1 %t4, i32 %a, i32 %b260 ret i32 %t5261}262 263;264; testz(or(extract_lo(X),extract_hi(X),or(extract_lo(Y),extract_hi(Y)) -> testz(X,Y)265;266 267; FIXME: Foldable to ptest(xor(%0,%1),xor(%0,%1))268define i1 @PR38788(<16 x i16> %0, <16 x i16> %1) {269; AVX1-LABEL: PR38788:270; AVX1: # %bb.0:271; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2272; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3273; AVX1-NEXT: vpcmpeqw %xmm2, %xmm3, %xmm2274; AVX1-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0275; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0276; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1277; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1278; AVX1-NEXT: vptest %ymm1, %ymm0279; AVX1-NEXT: setae %al280; AVX1-NEXT: vzeroupper281; AVX1-NEXT: retq282;283; AVX2-LABEL: PR38788:284; AVX2: # %bb.0:285; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0286; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1287; AVX2-NEXT: vptest %ymm1, %ymm0288; AVX2-NEXT: setae %al289; AVX2-NEXT: vzeroupper290; AVX2-NEXT: retq291 %3 = icmp eq <16 x i16> %0, %1292 %4 = sext <16 x i1> %3 to <16 x i16>293 %5 = bitcast <16 x i16> %4 to <4 x i64>294 %6 = tail call i32 @llvm.x86.avx.ptestc.256(<4 x i64> %5, <4 x i64> <i64 -1, i64 -1, i64 -1, i64 -1>)295 %7 = icmp eq i32 %6, 0296 ret i1 %7297}298 299declare i32 @llvm.x86.avx.ptestz.256(<4 x i64>, <4 x i64>)300declare i32 @llvm.x86.avx.ptestc.256(<4 x i64>, <4 x i64>)301declare i32 @llvm.x86.avx.ptestnzc.256(<4 x i64>, <4 x i64>)302