AArch64: Skip analyzeCompare for subregister uses (#141654)

The API is broken and doesn't provide a way to report the used
subregister, so it's unsafe to use. This will produce illegal
folds if the subregister is silently dropped.

Author: arsenm

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

@@ -1281,8 +1281,9 @@ bool AArch64InstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
                                       int64_t &CmpValue) const {
   // The first operand can be a frame index where we'd normally expect a
   // register.
+  // FIXME: Pass subregisters out of analyzeCompare
   assert(MI.getNumOperands() >= 2 && "All AArch64 cmps should have 2 operands");
-  if (!MI.getOperand(1).isReg())
+  if (!MI.getOperand(1).isReg() || MI.getOperand(1).getSubReg())
     return false;
 
   switch (MI.getOpcode()) {
@@ -1292,6 +1293,9 @@ bool AArch64InstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
   case AArch64::PTEST_PP_ANY:
     SrcReg = MI.getOperand(0).getReg();
     SrcReg2 = MI.getOperand(1).getReg();
+    if (MI.getOperand(2).getSubReg())
+      return false;
+
     // Not sure about the mask and value for now...
     CmpMask = ~0;
     CmpValue = 0;
@@ -1311,6 +1315,11 @@ bool AArch64InstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
     // Replace SUBSWrr with SUBWrr if NZCV is not used.
     SrcReg = MI.getOperand(1).getReg();
     SrcReg2 = MI.getOperand(2).getReg();
+
+    // FIXME: Pass subregisters out of analyzeCompare
+    if (MI.getOperand(2).getSubReg())
+      return false;
+
     CmpMask = ~0;
     CmpValue = 0;
     return true;

llvm/test/CodeGen/AArch64/peephole-opt-analyzeCompare-subreg-use.mir

@@ -0,0 +1,67 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=aarch64-- -run-pass=peephole-opt -o - %s | FileCheck %s
+
+# Make sure that analyzeCompare doesn't produce illegal folds due to
+# ignoring the subregister index on the use operands.
+
+---
+name:            analyze_compare_subreg_use_lhs
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: analyze_compare_subreg_use_lhs
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY $wzr
+    ; CHECK-NEXT: [[SUBSWrr:%[0-9]+]]:gpr32 = SUBSWrr [[COPY]].sub_32, [[COPY2]], implicit-def dead $nzcv
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr64all = IMPLICIT_DEF
+    ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:gpr64 = INSERT_SUBREG [[DEF]], killed [[SUBSWrr]], %subreg.sub_32
+    ; CHECK-NEXT: [[RORVXr:%[0-9]+]]:gpr64 = RORVXr [[COPY1]], killed [[INSERT_SUBREG]]
+    ; CHECK-NEXT: $x0 = COPY [[RORVXr]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %0:gpr64 = COPY $x1
+    %1:gpr64 = COPY $x0
+    %2:gpr32 = COPY $wzr
+    %3:gpr32 = SUBSWrr %0.sub_32, %2, implicit-def dead $nzcv
+    %4:gpr64all = IMPLICIT_DEF
+    %5:gpr64 = INSERT_SUBREG %4, killed %3, %subreg.sub_32
+    %6:gpr64 = RORVXr %1, killed %5
+    $x0 = COPY %6
+    RET_ReallyLR implicit $x0
+
+...
+
+---
+name:            analyze_compare_subreg_use_rhs
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: analyze_compare_subreg_use_rhs
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY $wzr
+    ; CHECK-NEXT: [[SUBSWrr:%[0-9]+]]:gpr32 = SUBSWrr [[COPY2]], [[COPY]].sub_32, implicit-def dead $nzcv
+    ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr64all = IMPLICIT_DEF
+    ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:gpr64 = INSERT_SUBREG [[DEF]], killed [[SUBSWrr]], %subreg.sub_32
+    ; CHECK-NEXT: [[RORVXr:%[0-9]+]]:gpr64 = RORVXr [[COPY1]], killed [[INSERT_SUBREG]]
+    ; CHECK-NEXT: $x0 = COPY [[RORVXr]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $x0
+    %0:gpr64 = COPY $x1
+    %1:gpr64 = COPY $x0
+    %2:gpr32 = COPY $wzr
+    %3:gpr32 = SUBSWrr %2, %0.sub_32, implicit-def dead $nzcv
+    %4:gpr64all = IMPLICIT_DEF
+    %5:gpr64 = INSERT_SUBREG %4, killed %3, %subreg.sub_32
+    %6:gpr64 = RORVXr %1, killed %5
+    $x0 = COPY %6
+    RET_ReallyLR implicit $x0
+
+...