[AArch64][GlobalISel] Emit TBNZ with G_BRCOND where the condition is SLT

[paquette]

Summary

When we have a G_ICMP which checks SLT, and the comparison is against 0, we can emit a TBNZ instead of a CBZ.

This lets us fold in things into the branch, which can provide some code size savings.

This is similar to the case in AArch64TargetLowering::LowerBR_CC.

[aemerson]

LGTM.

[paquette]

I think I forgot to type the Differential Revision bit here.

https://reviews.llvm.org/rGa37384140729

llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

bool AArch64InstructionSelector::selectCompareBranch(
   const auto Pred = (CmpInst::Predicate)CCMI->getOperand(1).getPredicate();
   MachineInstr *LHSMI = getDefIgnoringCopies(LHS, MRI);
   if (tryOptAndIntoCompareBranch(LHSMI, VRegAndVal->Value, Pred, DestMBB,
                                  MIB)) {
     I.eraseFromParent();
     return true;
   }
 
+  // When we have a less than comparison, we can just test if the last bit
+  // is not zero.
+  //
+  // Note that we don't want to do this when we have a G_AND because it can
+  // become a tst. The tst will make the test bit in the TB(N)Z redundant.
+  if (Pred == CmpInst::ICMP_SLT && LHSMI->getOpcode() != TargetOpcode::G_AND) {
+    uint64_t Bit = MRI.getType(LHS).getSizeInBits() - 1;
+    emitTestBit(LHS, Bit, /*IsNegative = */ true, DestMBB, MIB);
+    I.eraseFromParent();
+    return true;
+  }
+
   const RegisterBank &RB = *RBI.getRegBank(LHS, MRI, TRI);
   if (RB.getID() != AArch64::GPRRegBankID)
     return false;
   if (Pred != CmpInst::ICMP_NE && Pred != CmpInst::ICMP_EQ)
     return false;
 
   const unsigned CmpWidth = MRI.getType(LHS).getSizeInBits();
   unsigned CBOpc = 0;

llvm/test/CodeGen/AArch64/GlobalISel/tbnz-slt.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple aarch64-unknown-unknown -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s
+#
+# Test that we can produce a TBNZ when we have a slt compare against 0.
+#
+# The bit tested should be the size of the test register minus 1.
+#
+
+...
+---
+name:            tbnzx_slt
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: tbnzx_slt
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   TBNZX %copy, 63, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s64) = COPY $x0
+    %zero:gpr(s64) = G_CONSTANT i64 0
+    %cmp:gpr(s32) = G_ICMP intpred(slt), %copy(s64), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR
+
+...
+---
+name:            tbnzw_slt
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: tbnzw_slt
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr32 = COPY $w0
+  ; CHECK:   TBNZW %copy, 31, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s32) = COPY $w0
+    %zero:gpr(s32) = G_CONSTANT i32 0
+    %cmp:gpr(s32) = G_ICMP intpred(slt), %copy(s32), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR
+
+...
+---
+name:            no_tbnz_not_zero
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: no_tbnz_not_zero
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr32sp = COPY $w0
+  ; CHECK:   $wzr = SUBSWri %copy, 1, 0, implicit-def $nzcv
+  ; CHECK:   Bcc 11, %bb.1, implicit $nzcv
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s32) = COPY $w0
+    %one:gpr(s32) = G_CONSTANT i32 1
+    %cmp:gpr(s32) = G_ICMP intpred(slt), %copy(s32), %one
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR
+
+...
+---
+name:            dont_fold_and
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: dont_fold_and
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   $xzr = ANDSXri %copy, 8000, implicit-def $nzcv
+  ; CHECK:   %cmp:gpr32 = CSINCWr $wzr, $wzr, 10, implicit $nzcv
+  ; CHECK:   TBNZW %cmp, 0, %bb.1
+  ; CHECK:   B %bb.0
+  ; CHECK: bb.1:
+  ; CHECK:   RET_ReallyLR
+  bb.0:
+    successors: %bb.0, %bb.1
+    liveins: $x0
+    %copy:gpr(s64) = COPY $x0
+    %bit:gpr(s64) = G_CONSTANT i64 8
+    %zero:gpr(s64) = G_CONSTANT i64 0
+    %c:gpr(s64) = G_CONSTANT i64 8
+    %and:gpr(s64) = G_AND %copy, %bit
+    %cmp:gpr(s32) = G_ICMP intpred(slt), %and(s64), %zero
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR