323 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-- | FileCheck %s3; RUN: llc < %s -mtriple=x86_64-- -O0 | FileCheck %s -check-prefixes=CHECK-X64,CHECK-X64-O04; RUN: llc < %s -mtriple=x86_64-- -O2 | FileCheck %s -check-prefixes=CHECK-X64,CHECK-X64-O25 6; CHECK-LABEL: shift17define void @shift1(i256 %x, i256 %a, ptr nocapture %r) nounwind readnone {8; CHECK-LABEL: shift1:9; CHECK: # %bb.0: # %entry10; CHECK-NEXT: pushl %ebp11; CHECK-NEXT: movl %esp, %ebp12; CHECK-NEXT: pushl %ebx13; CHECK-NEXT: pushl %edi14; CHECK-NEXT: pushl %esi15; CHECK-NEXT: andl $-16, %esp16; CHECK-NEXT: subl $112, %esp17; CHECK-NEXT: movl 40(%ebp), %ecx18; CHECK-NEXT: movl 8(%ebp), %eax19; CHECK-NEXT: movl 12(%ebp), %edx20; CHECK-NEXT: movl 16(%ebp), %esi21; CHECK-NEXT: movl 32(%ebp), %edi22; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)23; CHECK-NEXT: movl 28(%ebp), %edi24; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)25; CHECK-NEXT: movl 24(%ebp), %edi26; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)27; CHECK-NEXT: movl 20(%ebp), %edi28; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)29; CHECK-NEXT: movl 36(%ebp), %edi30; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)31; CHECK-NEXT: movl %esi, {{[0-9]+}}(%esp)32; CHECK-NEXT: movl %edx, {{[0-9]+}}(%esp)33; CHECK-NEXT: movl %eax, {{[0-9]+}}(%esp)34; CHECK-NEXT: sarl $31, %edi35; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)36; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)37; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)38; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)39; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)40; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)41; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)42; CHECK-NEXT: movl %edi, {{[0-9]+}}(%esp)43; CHECK-NEXT: movl %ecx, %eax44; CHECK-NEXT: shrb $5, %al45; CHECK-NEXT: movzbl %al, %eax46; CHECK-NEXT: movl 40(%esp,%eax,4), %edx47; CHECK-NEXT: movl 36(%esp,%eax,4), %esi48; CHECK-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill49; CHECK-NEXT: shrdl %cl, %edx, %esi50; CHECK-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill51; CHECK-NEXT: movl 44(%esp,%eax,4), %esi52; CHECK-NEXT: shrdl %cl, %esi, %edx53; CHECK-NEXT: movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill54; CHECK-NEXT: movl 48(%esp,%eax,4), %ebx55; CHECK-NEXT: shrdl %cl, %ebx, %esi56; CHECK-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill57; CHECK-NEXT: movl 52(%esp,%eax,4), %esi58; CHECK-NEXT: shrdl %cl, %esi, %ebx59; CHECK-NEXT: movl 56(%esp,%eax,4), %edx60; CHECK-NEXT: shrdl %cl, %edx, %esi61; CHECK-NEXT: movl 32(%esp,%eax,4), %edi62; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill63; CHECK-NEXT: movl 60(%esp,%eax,4), %eax64; CHECK-NEXT: shrdl %cl, %eax, %edx65; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload66; CHECK-NEXT: shrdl %cl, %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill67; CHECK-NEXT: # kill: def $cl killed $cl killed $ecx68; CHECK-NEXT: sarl %cl, %eax69; CHECK-NEXT: movl 72(%ebp), %ecx70; CHECK-NEXT: movl %eax, 28(%ecx)71; CHECK-NEXT: movl %edx, 24(%ecx)72; CHECK-NEXT: movl %esi, 20(%ecx)73; CHECK-NEXT: movl %ebx, 16(%ecx)74; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload75; CHECK-NEXT: movl %eax, 12(%ecx)76; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload77; CHECK-NEXT: movl %eax, 8(%ecx)78; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload79; CHECK-NEXT: movl %eax, 4(%ecx)80; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload81; CHECK-NEXT: movl %eax, (%ecx)82; CHECK-NEXT: leal -12(%ebp), %esp83; CHECK-NEXT: popl %esi84; CHECK-NEXT: popl %edi85; CHECK-NEXT: popl %ebx86; CHECK-NEXT: popl %ebp87; CHECK-NEXT: retl88;89; CHECK-X64-O0-LABEL: shift1:90; CHECK-X64-O0: # %bb.0: # %entry91; CHECK-X64-O0-NEXT: movq {{[0-9]+}}(%rsp), %rax92; CHECK-X64-O0-NEXT: movq {{[0-9]+}}(%rsp), %rax93; CHECK-X64-O0-NEXT: movq {{[0-9]+}}(%rsp), %rax94; CHECK-X64-O0-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)95; CHECK-X64-O0-NEXT: movq %rsi, -{{[0-9]+}}(%rsp)96; CHECK-X64-O0-NEXT: movq %rdx, -{{[0-9]+}}(%rsp)97; CHECK-X64-O0-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)98; CHECK-X64-O0-NEXT: sarq $63, %rcx99; CHECK-X64-O0-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)100; CHECK-X64-O0-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)101; CHECK-X64-O0-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)102; CHECK-X64-O0-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)103; CHECK-X64-O0-NEXT: movb %r8b, %cl104; CHECK-X64-O0-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill105; CHECK-X64-O0-NEXT: movb %cl, %dl106; CHECK-X64-O0-NEXT: shrb $6, %dl107; CHECK-X64-O0-NEXT: movzbl %dl, %edx108; CHECK-X64-O0-NEXT: movl %edx, %edi109; CHECK-X64-O0-NEXT: movq -56(%rsp,%rdi,8), %rsi110; CHECK-X64-O0-NEXT: movq -72(%rsp,%rdi,8), %r8111; CHECK-X64-O0-NEXT: movq -64(%rsp,%rdi,8), %r9112; CHECK-X64-O0-NEXT: movq %r9, %rdx113; CHECK-X64-O0-NEXT: shrdq %cl, %rsi, %rdx114; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload115; CHECK-X64-O0-NEXT: movq -48(%rsp,%rdi,8), %rdi116; CHECK-X64-O0-NEXT: shrdq %cl, %rdi, %rsi117; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload118; CHECK-X64-O0-NEXT: shrdq %cl, %r9, %r8119; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload120; CHECK-X64-O0-NEXT: movq %r8, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill121; CHECK-X64-O0-NEXT: sarq %cl, %rdi122; CHECK-X64-O0-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rcx # 8-byte Reload123; CHECK-X64-O0-NEXT: movq %rdi, 24(%rax)124; CHECK-X64-O0-NEXT: movq %rsi, 16(%rax)125; CHECK-X64-O0-NEXT: movq %rdx, 8(%rax)126; CHECK-X64-O0-NEXT: movq %rcx, (%rax)127; CHECK-X64-O0-NEXT: retq128;129; CHECK-X64-O2-LABEL: shift1:130; CHECK-X64-O2: # %bb.0: # %entry131; CHECK-X64-O2-NEXT: movq {{[0-9]+}}(%rsp), %rax132; CHECK-X64-O2-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)133; CHECK-X64-O2-NEXT: movq %rdx, -{{[0-9]+}}(%rsp)134; CHECK-X64-O2-NEXT: movq %rsi, -{{[0-9]+}}(%rsp)135; CHECK-X64-O2-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)136; CHECK-X64-O2-NEXT: sarq $63, %rcx137; CHECK-X64-O2-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)138; CHECK-X64-O2-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)139; CHECK-X64-O2-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)140; CHECK-X64-O2-NEXT: movq %rcx, -{{[0-9]+}}(%rsp)141; CHECK-X64-O2-NEXT: movl %r8d, %ecx142; CHECK-X64-O2-NEXT: shrb $6, %cl143; CHECK-X64-O2-NEXT: movzbl %cl, %edx144; CHECK-X64-O2-NEXT: movq -56(%rsp,%rdx,8), %rsi145; CHECK-X64-O2-NEXT: movq -72(%rsp,%rdx,8), %rdi146; CHECK-X64-O2-NEXT: movq -64(%rsp,%rdx,8), %r9147; CHECK-X64-O2-NEXT: movq %r9, %r10148; CHECK-X64-O2-NEXT: movl %r8d, %ecx149; CHECK-X64-O2-NEXT: shrdq %cl, %rsi, %r10150; CHECK-X64-O2-NEXT: movq -48(%rsp,%rdx,8), %rdx151; CHECK-X64-O2-NEXT: shrdq %cl, %rdx, %rsi152; CHECK-X64-O2-NEXT: shrdq %cl, %r9, %rdi153; CHECK-X64-O2-NEXT: sarq %cl, %rdx154; CHECK-X64-O2-NEXT: movq %rdx, 24(%rax)155; CHECK-X64-O2-NEXT: movq %rsi, 16(%rax)156; CHECK-X64-O2-NEXT: movq %r10, 8(%rax)157; CHECK-X64-O2-NEXT: movq %rdi, (%rax)158; CHECK-X64-O2-NEXT: retq159entry:160 %0 = ashr i256 %x, %a161 store i256 %0, ptr %r162 ret void163}164 165define i256 @shift2(i256 %c) nounwind166; CHECK-LABEL: shift2:167; CHECK: # %bb.0:168; CHECK-NEXT: pushl %ebp169; CHECK-NEXT: movl %esp, %ebp170; CHECK-NEXT: pushl %ebx171; CHECK-NEXT: pushl %edi172; CHECK-NEXT: pushl %esi173; CHECK-NEXT: andl $-16, %esp174; CHECK-NEXT: subl $112, %esp175; CHECK-NEXT: movl 12(%ebp), %ecx176; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)177; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)178; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)179; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)180; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)181; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)182; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)183; CHECK-NEXT: movl $1, {{[0-9]+}}(%esp)184; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)185; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)186; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)187; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)188; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)189; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)190; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)191; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)192; CHECK-NEXT: movl %ecx, %eax193; CHECK-NEXT: shrb $3, %al194; CHECK-NEXT: andb $28, %al195; CHECK-NEXT: negb %al196; CHECK-NEXT: movsbl %al, %eax197; CHECK-NEXT: movl 68(%esp,%eax), %esi198; CHECK-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill199; CHECK-NEXT: movl 72(%esp,%eax), %edx200; CHECK-NEXT: movl %edx, %edi201; CHECK-NEXT: shldl %cl, %esi, %edi202; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill203; CHECK-NEXT: movl 76(%esp,%eax), %esi204; CHECK-NEXT: movl %esi, %edi205; CHECK-NEXT: shldl %cl, %edx, %edi206; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill207; CHECK-NEXT: movl 80(%esp,%eax), %edx208; CHECK-NEXT: movl %edx, %edi209; CHECK-NEXT: shldl %cl, %esi, %edi210; CHECK-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill211; CHECK-NEXT: movl 84(%esp,%eax), %esi212; CHECK-NEXT: movl %esi, %ebx213; CHECK-NEXT: shldl %cl, %edx, %ebx214; CHECK-NEXT: movl 88(%esp,%eax), %edi215; CHECK-NEXT: movl %edi, %edx216; CHECK-NEXT: shldl %cl, %esi, %edx217; CHECK-NEXT: movl 64(%esp,%eax), %esi218; CHECK-NEXT: movl %esi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill219; CHECK-NEXT: movl 92(%esp,%eax), %esi220; CHECK-NEXT: shldl %cl, %edi, %esi221; CHECK-NEXT: movl 8(%ebp), %eax222; CHECK-NEXT: movl %esi, 28(%eax)223; CHECK-NEXT: movl %edx, 24(%eax)224; CHECK-NEXT: movl %ebx, 20(%eax)225; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload226; CHECK-NEXT: movl %edx, 16(%eax)227; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload228; CHECK-NEXT: movl %edx, 12(%eax)229; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edx # 4-byte Reload230; CHECK-NEXT: movl %edx, 8(%eax)231; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %edi # 4-byte Reload232; CHECK-NEXT: movl %edi, %edx233; CHECK-NEXT: shll %cl, %edx234; CHECK-NEXT: # kill: def $cl killed $cl killed $ecx235; CHECK-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %esi # 4-byte Reload236; CHECK-NEXT: shldl %cl, %edi, %esi237; CHECK-NEXT: movl %esi, 4(%eax)238; CHECK-NEXT: movl %edx, (%eax)239; CHECK-NEXT: leal -12(%ebp), %esp240; CHECK-NEXT: popl %esi241; CHECK-NEXT: popl %edi242; CHECK-NEXT: popl %ebx243; CHECK-NEXT: popl %ebp244; CHECK-NEXT: retl $4245;246; CHECK-X64-O0-LABEL: shift2:247; CHECK-X64-O0: # %bb.0:248; CHECK-X64-O0-NEXT: movq %rdi, %rax249; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)250; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)251; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)252; CHECK-X64-O0-NEXT: movq $1, -{{[0-9]+}}(%rsp)253; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)254; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)255; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)256; CHECK-X64-O0-NEXT: movq $0, -{{[0-9]+}}(%rsp)257; CHECK-X64-O0-NEXT: movb %sil, %cl258; CHECK-X64-O0-NEXT: movb %cl, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill259; CHECK-X64-O0-NEXT: movb %cl, %dl260; CHECK-X64-O0-NEXT: shrb $3, %dl261; CHECK-X64-O0-NEXT: andb $24, %dl262; CHECK-X64-O0-NEXT: negb %dl263; CHECK-X64-O0-NEXT: movsbq %dl, %r8264; CHECK-X64-O0-NEXT: movq -40(%rsp,%r8), %r9265; CHECK-X64-O0-NEXT: movq -32(%rsp,%r8), %rdx266; CHECK-X64-O0-NEXT: movq -24(%rsp,%r8), %r10267; CHECK-X64-O0-NEXT: movq %r10, %rsi268; CHECK-X64-O0-NEXT: shldq %cl, %rdx, %rsi269; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload270; CHECK-X64-O0-NEXT: movq -16(%rsp,%r8), %r8271; CHECK-X64-O0-NEXT: shldq %cl, %r10, %r8272; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload273; CHECK-X64-O0-NEXT: movq %r9, %r10274; CHECK-X64-O0-NEXT: shlq %cl, %r10275; CHECK-X64-O0-NEXT: movb {{[-0-9]+}}(%r{{[sb]}}p), %cl # 1-byte Reload276; CHECK-X64-O0-NEXT: movq %r10, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill277; CHECK-X64-O0-NEXT: shldq %cl, %r9, %rdx278; CHECK-X64-O0-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rcx # 8-byte Reload279; CHECK-X64-O0-NEXT: movq %r8, 24(%rdi)280; CHECK-X64-O0-NEXT: movq %rsi, 16(%rdi)281; CHECK-X64-O0-NEXT: movq %rdx, 8(%rdi)282; CHECK-X64-O0-NEXT: movq %rcx, (%rdi)283; CHECK-X64-O0-NEXT: retq284;285; CHECK-X64-O2-LABEL: shift2:286; CHECK-X64-O2: # %bb.0:287; CHECK-X64-O2-NEXT: movq %rsi, %rcx288; CHECK-X64-O2-NEXT: movq %rdi, %rax289; CHECK-X64-O2-NEXT: xorps %xmm0, %xmm0290; CHECK-X64-O2-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)291; CHECK-X64-O2-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)292; CHECK-X64-O2-NEXT: movaps %xmm0, -{{[0-9]+}}(%rsp)293; CHECK-X64-O2-NEXT: movq $0, -{{[0-9]+}}(%rsp)294; CHECK-X64-O2-NEXT: movq $1, -{{[0-9]+}}(%rsp)295; CHECK-X64-O2-NEXT: movl %ecx, %edx296; CHECK-X64-O2-NEXT: shrb $3, %dl297; CHECK-X64-O2-NEXT: andb $24, %dl298; CHECK-X64-O2-NEXT: negb %dl299; CHECK-X64-O2-NEXT: movsbq %dl, %rdx300; CHECK-X64-O2-NEXT: movq -40(%rsp,%rdx), %rsi301; CHECK-X64-O2-NEXT: movq -32(%rsp,%rdx), %rdi302; CHECK-X64-O2-NEXT: movq -24(%rsp,%rdx), %r8303; CHECK-X64-O2-NEXT: movq %r8, %r9304; CHECK-X64-O2-NEXT: shldq %cl, %rdi, %r9305; CHECK-X64-O2-NEXT: movq -16(%rsp,%rdx), %rdx306; CHECK-X64-O2-NEXT: shldq %cl, %r8, %rdx307; CHECK-X64-O2-NEXT: movq %rsi, %r8308; CHECK-X64-O2-NEXT: shlq %cl, %r8309; CHECK-X64-O2-NEXT: # kill: def $cl killed $cl killed $rcx310; CHECK-X64-O2-NEXT: shldq %cl, %rsi, %rdi311; CHECK-X64-O2-NEXT: movq %rdx, 24(%rax)312; CHECK-X64-O2-NEXT: movq %r9, 16(%rax)313; CHECK-X64-O2-NEXT: movq %rdi, 8(%rax)314; CHECK-X64-O2-NEXT: movq %r8, (%rax)315; CHECK-X64-O2-NEXT: retq316{317 %b = shl i256 1, %c ; %c must not be a constant318 ; Special case when %c is 0:319 ret i256 %b320}321;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:322; CHECK-X64: {{.*}}323