diff options
Diffstat (limited to 'llvm/test/CodeGen/X86/known-bits.ll')
| -rw-r--r-- | llvm/test/CodeGen/X86/known-bits.ll | 76 |
1 files changed, 66 insertions, 10 deletions
diff --git a/llvm/test/CodeGen/X86/known-bits.ll b/llvm/test/CodeGen/X86/known-bits.ll index d9a7921131f..8f3b983251d 100644 --- a/llvm/test/CodeGen/X86/known-bits.ll +++ b/llvm/test/CodeGen/X86/known-bits.ll @@ -5,44 +5,100 @@ define void @knownbits_zext_in_reg(i8*) nounwind { ; X32-LABEL: knownbits_zext_in_reg: ; X32: # %bb.0: # %BB +; X32-NEXT: pushl %ebp ; X32-NEXT: pushl %ebx +; X32-NEXT: pushl %edi +; X32-NEXT: pushl %esi +; X32-NEXT: subl $16, %esp ; X32-NEXT: movl {{[0-9]+}}(%esp), %eax ; X32-NEXT: movzbl (%eax), %ecx ; X32-NEXT: imull $101, %ecx, %eax ; X32-NEXT: shrl $14, %eax -; X32-NEXT: imull $177, %ecx, %edx -; X32-NEXT: shrl $14, %edx -; X32-NEXT: movzbl %al, %ecx -; X32-NEXT: xorl %ebx, %ebx +; X32-NEXT: imull $177, %ecx, %ecx +; X32-NEXT: shrl $14, %ecx +; X32-NEXT: movzbl %al, %eax +; X32-NEXT: vpxor %xmm0, %xmm0, %xmm0 +; X32-NEXT: vpinsrd $1, %eax, %xmm0, %xmm1 +; X32-NEXT: vbroadcastss {{.*#+}} xmm2 = [3.57331108E-43,3.57331108E-43,3.57331108E-43,3.57331108E-43] +; X32-NEXT: vpand %xmm2, %xmm1, %xmm1 +; X32-NEXT: movzbl %cl, %eax +; X32-NEXT: vpinsrd $1, %eax, %xmm0, %xmm0 +; X32-NEXT: vpand %xmm2, %xmm0, %xmm0 +; X32-NEXT: vpextrd $1, %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill +; X32-NEXT: vpextrd $1, %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill +; X32-NEXT: xorl %ecx, %ecx +; X32-NEXT: vmovd %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill +; X32-NEXT: vmovd %xmm0, (%esp) # 4-byte Folded Spill +; X32-NEXT: vpextrd $2, %xmm1, %edi +; X32-NEXT: vpextrd $2, %xmm0, %esi +; X32-NEXT: vpextrd $3, %xmm1, %ebx +; X32-NEXT: vpextrd $3, %xmm0, %ebp ; X32-NEXT: .p2align 4, 0x90 ; X32-NEXT: .LBB0_1: # %CF ; X32-NEXT: # =>This Loop Header: Depth=1 ; X32-NEXT: # Child Loop BB0_2 Depth 2 -; X32-NEXT: movl %ecx, %eax -; X32-NEXT: divb %dl +; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload +; X32-NEXT: xorl %edx, %edx +; X32-NEXT: divl {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload +; X32-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload +; X32-NEXT: xorl %edx, %edx +; X32-NEXT: divl (%esp) # 4-byte Folded Reload +; X32-NEXT: movl %edi, %eax +; X32-NEXT: xorl %edx, %edx +; X32-NEXT: divl %esi +; X32-NEXT: movl %ebx, %eax +; X32-NEXT: xorl %edx, %edx +; X32-NEXT: divl %ebp ; X32-NEXT: .p2align 4, 0x90 ; X32-NEXT: .LBB0_2: # %CF237 ; X32-NEXT: # Parent Loop BB0_1 Depth=1 ; X32-NEXT: # => This Inner Loop Header: Depth=2 -; X32-NEXT: testb %bl, %bl +; X32-NEXT: testb %cl, %cl ; X32-NEXT: jne .LBB0_2 ; X32-NEXT: jmp .LBB0_1 ; ; X64-LABEL: knownbits_zext_in_reg: ; X64: # %bb.0: # %BB +; X64-NEXT: pushq %rbp +; X64-NEXT: pushq %rbx ; X64-NEXT: movzbl (%rdi), %eax ; X64-NEXT: imull $101, %eax, %ecx ; X64-NEXT: shrl $14, %ecx -; X64-NEXT: imull $177, %eax, %edx -; X64-NEXT: shrl $14, %edx +; X64-NEXT: imull $177, %eax, %eax +; X64-NEXT: shrl $14, %eax ; X64-NEXT: movzbl %cl, %ecx +; X64-NEXT: vpxor %xmm0, %xmm0, %xmm0 +; X64-NEXT: vpinsrd $1, %ecx, %xmm0, %xmm1 +; X64-NEXT: vbroadcastss {{.*#+}} xmm2 = [3.57331108E-43,3.57331108E-43,3.57331108E-43,3.57331108E-43] +; X64-NEXT: vpand %xmm2, %xmm1, %xmm1 +; X64-NEXT: movzbl %al, %eax +; X64-NEXT: vpinsrd $1, %eax, %xmm0, %xmm0 +; X64-NEXT: vpand %xmm2, %xmm0, %xmm0 +; X64-NEXT: vpextrd $1, %xmm1, %r8d +; X64-NEXT: vpextrd $1, %xmm0, %r9d ; X64-NEXT: xorl %esi, %esi +; X64-NEXT: vmovd %xmm1, %r10d +; X64-NEXT: vmovd %xmm0, %r11d +; X64-NEXT: vpextrd $2, %xmm1, %edi +; X64-NEXT: vpextrd $2, %xmm0, %ebx +; X64-NEXT: vpextrd $3, %xmm1, %ecx +; X64-NEXT: vpextrd $3, %xmm0, %ebp ; X64-NEXT: .p2align 4, 0x90 ; X64-NEXT: .LBB0_1: # %CF ; X64-NEXT: # =>This Loop Header: Depth=1 ; X64-NEXT: # Child Loop BB0_2 Depth 2 +; X64-NEXT: movl %r8d, %eax +; X64-NEXT: xorl %edx, %edx +; X64-NEXT: divl %r9d +; X64-NEXT: movl %r10d, %eax +; X64-NEXT: xorl %edx, %edx +; X64-NEXT: divl %r11d +; X64-NEXT: movl %edi, %eax +; X64-NEXT: xorl %edx, %edx +; X64-NEXT: divl %ebx ; X64-NEXT: movl %ecx, %eax -; X64-NEXT: divb %dl +; X64-NEXT: xorl %edx, %edx +; X64-NEXT: divl %ebp ; X64-NEXT: .p2align 4, 0x90 ; X64-NEXT: .LBB0_2: # %CF237 ; X64-NEXT: # Parent Loop BB0_1 Depth=1 |
