174 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-- -mattr=sse | FileCheck %s --check-prefixes=CHECK,SSE3; RUN: llc < %s -mtriple=x86_64-- -mattr=avx | FileCheck %s --check-prefixes=CHECK,AVX4 5; With no-wrap:6; (X * Y) == 0 --> (X == 0) || (Y == 0)7; (X * Y) != 0 --> (X != 0) && (Y != 0)8 9define i1 @mul_nsw_eq0_i8(i8 %x, i8 %y) {10; CHECK-LABEL: mul_nsw_eq0_i8:11; CHECK: # %bb.0:12; CHECK-NEXT: testb %sil, %sil13; CHECK-NEXT: sete %cl14; CHECK-NEXT: testb %dil, %dil15; CHECK-NEXT: sete %al16; CHECK-NEXT: orb %cl, %al17; CHECK-NEXT: retq18 %m = mul nsw i8 %x, %y19 %r = icmp eq i8 %m, 020 ret i1 %r21}22 23; negative test - not valid if mul can overflow24 25define i1 @mul_eq0_i8(i8 %x, i8 %y) {26; CHECK-LABEL: mul_eq0_i8:27; CHECK: # %bb.0:28; CHECK-NEXT: movl %edi, %eax29; CHECK-NEXT: # kill: def $al killed $al killed $eax30; CHECK-NEXT: mulb %sil31; CHECK-NEXT: testb %al, %al32; CHECK-NEXT: sete %al33; CHECK-NEXT: retq34 %m = mul i8 %x, %y35 %r = icmp eq i8 %m, 036 ret i1 %r37}38 39; negative test - don't try with minsize40 41define i1 @mul_nsw_eq0_i8_size(i8 %x, i8 %y) minsize {42; CHECK-LABEL: mul_nsw_eq0_i8_size:43; CHECK: # %bb.0:44; CHECK-NEXT: movl %edi, %eax45; CHECK-NEXT: # kill: def $al killed $al killed $eax46; CHECK-NEXT: mulb %sil47; CHECK-NEXT: testb %al, %al48; CHECK-NEXT: sete %al49; CHECK-NEXT: retq50 %m = mul nsw i8 %x, %y51 %r = icmp eq i8 %m, 052 ret i1 %r53}54 55define i1 @mul_nsw_ne0_i16(i16 %x, i16 %y) {56; CHECK-LABEL: mul_nsw_ne0_i16:57; CHECK: # %bb.0:58; CHECK-NEXT: testw %si, %si59; CHECK-NEXT: setne %cl60; CHECK-NEXT: testw %di, %di61; CHECK-NEXT: setne %al62; CHECK-NEXT: andb %cl, %al63; CHECK-NEXT: retq64 %m = mul nsw i16 %x, %y65 %r = icmp ne i16 %m, 066 ret i1 %r67}68 69define i1 @mul_nuw_eq0_i32(i32 %x, i32 %y) {70; CHECK-LABEL: mul_nuw_eq0_i32:71; CHECK: # %bb.0:72; CHECK-NEXT: testl %esi, %esi73; CHECK-NEXT: sete %cl74; CHECK-NEXT: testl %edi, %edi75; CHECK-NEXT: sete %al76; CHECK-NEXT: orb %cl, %al77; CHECK-NEXT: retq78 %m = mul nuw i32 %x, %y79 %r = icmp eq i32 %m, 080 ret i1 %r81}82 83define i1 @mul_nsw_nuw_ne0_i64(i64 %x, i64 %y) {84; CHECK-LABEL: mul_nsw_nuw_ne0_i64:85; CHECK: # %bb.0:86; CHECK-NEXT: testq %rsi, %rsi87; CHECK-NEXT: setne %cl88; CHECK-NEXT: testq %rdi, %rdi89; CHECK-NEXT: setne %al90; CHECK-NEXT: andb %cl, %al91; CHECK-NEXT: retq92 %m = mul nsw nuw i64 %x, %y93 %r = icmp ne i64 %m, 094 ret i1 %r95}96 97define <16 x i1> @mul_nuw_eq0_v16i8(<16 x i8> %x, <16 x i8> %y) {98; SSE-LABEL: mul_nuw_eq0_v16i8:99; SSE: # %bb.0:100; SSE-NEXT: pxor %xmm2, %xmm2101; SSE-NEXT: pcmpeqb %xmm2, %xmm1102; SSE-NEXT: pcmpeqb %xmm2, %xmm0103; SSE-NEXT: por %xmm1, %xmm0104; SSE-NEXT: retq105;106; AVX-LABEL: mul_nuw_eq0_v16i8:107; AVX: # %bb.0:108; AVX-NEXT: vpxor %xmm2, %xmm2, %xmm2109; AVX-NEXT: vpcmpeqb %xmm2, %xmm1, %xmm1110; AVX-NEXT: vpcmpeqb %xmm2, %xmm0, %xmm0111; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0112; AVX-NEXT: retq113 %m = mul nuw <16 x i8> %x, %y114 %r = icmp eq <16 x i8> %m, zeroinitializer115 ret <16 x i1> %r116}117 118define <4 x i1> @mul_nsw_ne0_v4i32(<4 x i32> %x, <4 x i32> %y) {119; SSE-LABEL: mul_nsw_ne0_v4i32:120; SSE: # %bb.0:121; SSE-NEXT: pxor %xmm2, %xmm2122; SSE-NEXT: pcmpeqd %xmm2, %xmm1123; SSE-NEXT: pcmpeqd %xmm2, %xmm0124; SSE-NEXT: por %xmm1, %xmm0125; SSE-NEXT: pcmpeqd %xmm1, %xmm1126; SSE-NEXT: pxor %xmm1, %xmm0127; SSE-NEXT: retq128;129; AVX-LABEL: mul_nsw_ne0_v4i32:130; AVX: # %bb.0:131; AVX-NEXT: vpxor %xmm2, %xmm2, %xmm2132; AVX-NEXT: vpcmpeqd %xmm2, %xmm1, %xmm1133; AVX-NEXT: vpcmpeqd %xmm2, %xmm0, %xmm0134; AVX-NEXT: vpor %xmm1, %xmm0, %xmm0135; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1136; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0137; AVX-NEXT: retq138 %m = mul nsw <4 x i32> %x, %y139 %r = icmp ne <4 x i32> %m, zeroinitializer140 ret <4 x i1> %r141}142 143; negative test - don't try with minsize144; TODO: SSE would be much smaller if decomposed.145 146define <4 x i1> @mul_nsw_ne0_v4i32_size(<4 x i32> %x, <4 x i32> %y) minsize {147; SSE-LABEL: mul_nsw_ne0_v4i32_size:148; SSE: # %bb.0:149; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]150; SSE-NEXT: pmuludq %xmm1, %xmm0151; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]152; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,3,3]153; SSE-NEXT: pmuludq %xmm2, %xmm1154; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]155; SSE-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]156; SSE-NEXT: pxor %xmm1, %xmm1157; SSE-NEXT: pcmpeqd %xmm0, %xmm1158; SSE-NEXT: pcmpeqd %xmm0, %xmm0159; SSE-NEXT: pxor %xmm1, %xmm0160; SSE-NEXT: retq161;162; AVX-LABEL: mul_nsw_ne0_v4i32_size:163; AVX: # %bb.0:164; AVX-NEXT: vpmulld %xmm1, %xmm0, %xmm0165; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1166; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0167; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1168; AVX-NEXT: vpxor %xmm1, %xmm0, %xmm0169; AVX-NEXT: retq170 %m = mul nsw <4 x i32> %x, %y171 %r = icmp ne <4 x i32> %m, zeroinitializer172 ret <4 x i1> %r173}174