[X86] make sure POP has implicit def/use of stack pointer when materializing 8-bit immediates for minsize

[ychen]

Summary

Otherwise PostRA list scheduler may reorder instruction, such as

schedule this
'''
pushq $0x8
pop %rbx
lea 0x2a0(%rsp),%r15
'''
to
'''
pushq $0x8
lea 0x2a0(%rsp),%r15
pop %rbx
'''
by mistake. The patch is to prevent this to happen by making sure POP has
implicit use of SP.

[craig.topper]

LGTM

llvm/lib/Target/X86/X86InstrInfo.cpp

if (Subtarget.is64Bit()) {
     MIB->getOperand(0)
         .setReg(getX86SubSuperRegister(MIB.getReg(0), 64));
   } else {
     assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
     StackAdjustment = 4;
     BuildMI(MBB, I, DL, TII.get(X86::PUSH32i8)).addImm(Imm);
     MIB->setDesc(TII.get(X86::POP32r));
   }
+  MIB->RemoveOperand(1);
+  MIB->addImplicitDefUseOperands(*MBB.getParent());
 
   // Build CFI if necessary.
   MachineFunction &MF = *MBB.getParent();
   const X86FrameLowering *TFL = Subtarget.getFrameLowering();
   bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
   bool NeedsDwarfCFI = !IsWin64Prologue && MF.needsFrameMoves();
   bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
   if (EmitCFI) {

llvm/test/CodeGen/X86/materialize.ll

 ; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK32
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK64
 ; RUN: llc -mtriple=x86_64-pc-win32 -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECKWIN64
 
+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+cmov %s -o /dev/null \
+; RUN:     -print-after postrapseudos -filter-print-funcs pr26023 2>&1 \
+; RUN:    | FileCheck %s --check-prefix=OPERAND32
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+cmov %s -o /dev/null \
+; RUN:     -print-after postrapseudos -filter-print-funcs one64_minsize 2>&1 \
+; RUN:    | FileCheck %s --check-prefix=OPERAND64
+
 define i32 @one32_nooptsize() {
 entry:
   ret i32 1
 
 ; When not optimizing for size, use mov.
 ; CHECK32-LABEL: one32_nooptsize:
 ; CHECK32:       movl $1, %eax
 ; CHECK32-NEXT:  retl

 ; CHECK64:       movl $5, %ecx
 ; CHECK64:       retq
 
 ; 32-bit X86 doesn't have a redzone.
 ; CHECK32-LABEL: pr26023:
 ; CHECK32:       pushl $5
 ; CHECK32:       popl %ecx
 ; CHECK32:       retl
+
+; Check push/pop have implicit def/use of $esp
+; OPERAND32:      PUSH32i8 5, implicit-def $esp, implicit $esp
+; OPERAND32-NEXT: CFI_INSTRUCTION adjust_cfa_offset 4
+; OPERAND32-NEXT: renamable $ecx = POP32r implicit-def $esp, implicit $esp
+; OPERAND32-NEXT: CFI_INSTRUCTION adjust_cfa_offset -4
 }
 
 
 define i64 @one64_minsize() minsize {
 entry:
   ret i64 1
 ; On 64-bit we don't do xor-inc yet, so push-pop it is.
 ; CHECK64-LABEL: one64_minsize:
 ; CHECK64:       pushq $1
 ; CHECK64:       .cfi_adjust_cfa_offset 8
 ; CHECK64:       popq %rax
 ; CHECK64:       .cfi_adjust_cfa_offset -8
 ; CHECK64-NEXT:  retq
 
 ; On Win64 we can't adjust the stack unless there's a frame pointer.
 ; CHECKWIN64-LABEL: one64_minsize:
 ; CHECKWIN64:       movl $1, %eax
 ; CHECKWIN64-NEXT:  retq
+
+; Check push/pop have implicit def/use of $rsp
+; OPERAND64:      PUSH64i8 1, implicit-def $rsp, implicit $rsp
+; OPERAND64-NEXT: CFI_INSTRUCTION adjust_cfa_offset 8
+; OPERAND64-NEXT: $rax = POP64r implicit-def $rsp, implicit $rsp
+; OPERAND64-NEXT: CFI_INSTRUCTION adjust_cfa_offset -8
+; OPERAND64-NEXT: RET 0, $rax
 }
 
 define i32 @minus_one32() optsize {
 entry:
   ret i32 -1
 
 ; CHECK32-LABEL: minus_one32:
 ; CHECK32:       xorl %eax, %eax