AArch64: Remove reversedInstructionsWithoutDebug helper

When using reversedInstructionsWithoutDebug to construct a range from a
pair of MachineInstrBundleIterators, the range unexpectedly leaves out an
element. This results in mis-optimization as @mstorsjo points out in
https://reviews.llvm.org/D78157.

The problem is that when we convert a MachineInstrBundleIterator to a
reverse iterator, the result gets incremented:

  MachineInstrBundleIterator(++I.getReverse())

The comment there explains that the "resulting iterator will dereference
... to the previous node, which is somewhat unexpected; but converting
the two endpoints in a range will give the same range in reverse". This
makes it hard to understand what reversedInstructionsWithoutDebug will
do: I've removed the helper to prevent similar mistakes in the future.
This commit is contained in:
Vedant Kumar 2020-04-24 11:16:43 -07:00
parent ef423a3ba5
commit c0fa447e02
4 changed files with 58 additions and 12 deletions

View File

@ -1082,14 +1082,6 @@ inline auto instructionsWithoutDebug(IterT It, IterT End) {
});
}
/// Construct a range iterator which begins at \p It and moves backwards until
/// \p Begin is reached, skipping any debug instructions.
template <typename IterT>
inline auto reversedInstructionsWithoutDebug(IterT It, IterT Begin) {
return instructionsWithoutDebug(make_reverse_iterator(It),
make_reverse_iterator(Begin));
}
} // end namespace llvm
#endif // LLVM_CODEGEN_MACHINEBASICBLOCK_H

View File

@ -158,9 +158,9 @@ MachineInstr *AArch64ConditionOptimizer::findSuitableCompare(
return nullptr;
// Now find the instruction controlling the terminator.
auto B = MBB->begin();
for (MachineInstr &I : reversedInstructionsWithoutDebug(
Term == B ? Term : std::prev(Term), B)) {
for (MachineBasicBlock::iterator B = MBB->begin(), It = Term; It != B;) {
It = prev_nodbg(It, B);
MachineInstr &I = *It;
assert(!I.isTerminator() && "Spurious terminator");
// Check if there is any use of NZCV between CMP and Bcc.
if (I.readsRegister(AArch64::NZCV))

View File

@ -1185,7 +1185,7 @@ static bool areCFlagsAccessedBetweenInstrs(
// We iterate backward starting at \p To until we hit \p From.
for (const MachineInstr &Instr :
reversedInstructionsWithoutDebug(std::prev(To), From)) {
instructionsWithoutDebug(++To.getReverse(), From.getReverse())) {
if (((AccessToCheck & AK_Write) &&
Instr.modifiesRegister(AArch64::NZCV, TRI)) ||
((AccessToCheck & AK_Read) && Instr.readsRegister(AArch64::NZCV, TRI)))

View File

@ -0,0 +1,54 @@
# RUN: llc %s -o - -run-pass=peephole-opt | FileCheck %s
# Check that we don't optimize out a subs due to areCFlagsAccessedBetweenInstrs
# returning the wrong result; it should check the cneg before the subs which does
# modify cflags.
# CHECK-LABEL: f
# CHECK: SUBSWrr
# CHECK-NEXT: SUBWrr
# CHECK-NEXT: CSNEGWr
# CHECK-NEXT: SUBSWri
# CHECK-NEXT: CSNEGWr
# CHECK-NEXT: Bcc
--- |
target datalayout = "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-w64-windows-gnu"
define dso_local void @f() {
ret void
}
...
---
name: f
registers:
- { id: 43, class: gpr32, preferred-register: '' }
- { id: 44, class: gpr32, preferred-register: '' }
- { id: 46, class: gpr32, preferred-register: '' }
- { id: 47, class: gpr32, preferred-register: '' }
- { id: 48, class: gpr32common, preferred-register: '' }
- { id: 49, class: gpr32common, preferred-register: '' }
- { id: 50, class: gpr32, preferred-register: '' }
- { id: 51, class: gpr32, preferred-register: '' }
- { id: 52, class: gpr32, preferred-register: '' }
- { id: 53, class: gpr32, preferred-register: '' }
body: |
bb.0:
successors: %bb.0
%43 = MOVi32imm 1
%44 = MOVi32imm 1
%46 = MOVi32imm 1
%47 = MOVi32imm 1
%48 = nsw SUBSWrr killed %43, killed %46, implicit-def dead $nzcv
%49 = nsw SUBSWrr killed %44, killed %47, implicit-def dead $nzcv
%50 = SUBSWri %48, 0, 0, implicit-def $nzcv
%51 = CSNEGWr %48, %48, 5, implicit $nzcv
%52 = SUBSWri %49, 0, 0, implicit-def $nzcv
%53 = CSNEGWr %49, %49, 5, implicit $nzcv
Bcc 1, %bb.0, implicit $nzcv
RET_ReallyLR
...