58 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 32; RUN: llc < %s -mtriple=x86_64-unknown | FileCheck %s3 4; When optimize for size, the constant $858993459 is moved into a register,5; and use that register in following two andl instructions.6 7define i32 @cnt32_optsize(i32 %x) nounwind readnone optsize {8; CHECK-LABEL: cnt32_optsize:9; CHECK: # %bb.0:10; CHECK-NEXT: movl %edi, %eax11; CHECK-NEXT: shrl %eax12; CHECK-NEXT: andl $1431655765, %eax # imm = 0x5555555513; CHECK-NEXT: subl %eax, %edi14; CHECK-NEXT: movl $858993459, %eax # imm = 0x3333333315; CHECK-NEXT: movl %edi, %ecx16; CHECK-NEXT: andl %eax, %ecx17; CHECK-NEXT: shrl $2, %edi18; CHECK-NEXT: andl %eax, %edi19; CHECK-NEXT: addl %ecx, %edi20; CHECK-NEXT: movl %edi, %eax21; CHECK-NEXT: shrl $4, %eax22; CHECK-NEXT: addl %edi, %eax23; CHECK-NEXT: andl $252645135, %eax # imm = 0xF0F0F0F24; CHECK-NEXT: imull $16843009, %eax, %eax # imm = 0x101010125; CHECK-NEXT: shrl $24, %eax26; CHECK-NEXT: retq27 %cnt = tail call i32 @llvm.ctpop.i32(i32 %x)28 ret i32 %cnt29}30 31; When optimize for speed, the constant $858993459 can be directly folded into32; two andl instructions.33 34define i32 @cnt32_optspeed(i32 %x) nounwind readnone {35; CHECK-LABEL: cnt32_optspeed:36; CHECK: # %bb.0:37; CHECK-NEXT: movl %edi, %eax38; CHECK-NEXT: shrl %eax39; CHECK-NEXT: andl $1431655765, %eax # imm = 0x5555555540; CHECK-NEXT: subl %eax, %edi41; CHECK-NEXT: movl %edi, %eax42; CHECK-NEXT: andl $858993459, %eax # imm = 0x3333333343; CHECK-NEXT: shrl $2, %edi44; CHECK-NEXT: andl $858993459, %edi # imm = 0x3333333345; CHECK-NEXT: addl %eax, %edi46; CHECK-NEXT: movl %edi, %eax47; CHECK-NEXT: shrl $4, %eax48; CHECK-NEXT: addl %edi, %eax49; CHECK-NEXT: andl $252645135, %eax # imm = 0xF0F0F0F50; CHECK-NEXT: imull $16843009, %eax, %eax # imm = 0x101010151; CHECK-NEXT: shrl $24, %eax52; CHECK-NEXT: retq53 %cnt = tail call i32 @llvm.ctpop.i32(i32 %x)54 ret i32 %cnt55}56 57declare i32 @llvm.ctpop.i32(i32) nounwind readnone58