[llvm] 609a489 - [AArch64][GlobalISel] Reland SLT/SGT TBNZ optimization

Jessica Paquette via llvm-commits llvm-commits at lists.llvm.org
Fri Feb 7 11:16:20 PST 2020


Author: Jessica Paquette
Date: 2020-02-07T11:15:25-08:00
New Revision: 609a489e052e8ed2731aebbcc7828de1102e871d

URL: https://github.com/llvm/llvm-project/commit/609a489e052e8ed2731aebbcc7828de1102e871d
DIFF: https://github.com/llvm/llvm-project/commit/609a489e052e8ed2731aebbcc7828de1102e871d.diff

LOG: [AArch64][GlobalISel] Reland SLT/SGT TBNZ optimization

The issue in the previous commits was that we swap the LHS and RHS while
looking for the constant. In SLT/SGT, the constant must be on the RHS, or the
optimization is invalid.

Move the swapping logic after the check for the SLT/SGT case and update tests.

Original commits:

d78cefb1601070cb028b61bbc1bd6f25a9c1837c
a3738414072900ace9cbbe209d0195a3443d1d54

Added: 
    llvm/test/CodeGen/AArch64/GlobalISel/tbnz-slt.mir
    llvm/test/CodeGen/AArch64/GlobalISel/tbz-sgt.mir

Modified: 
    llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp b/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
index 3d6d11f35217..83f1004c258c 100644
--- a/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
+++ b/llvm/lib/Target/AArch64/AArch64InstructionSelector.cpp
@@ -1222,26 +1222,55 @@ bool AArch64InstructionSelector::selectCompareBranch(
   Register LHS = CCMI->getOperand(2).getReg();
   Register RHS = CCMI->getOperand(3).getReg();
   auto VRegAndVal = getConstantVRegValWithLookThrough(RHS, MRI);
-  if (!VRegAndVal)
+  MachineIRBuilder MIB(I);
+  const auto Pred = (CmpInst::Predicate)CCMI->getOperand(1).getPredicate();
+  MachineInstr *LHSMI = getDefIgnoringCopies(LHS, MRI);
+
+  // When we can emit a TB(N)Z, prefer that.
+  //
+  // Handle non-commutative condition codes first.
+  // 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 (VRegAndVal && LHSMI->getOpcode() != TargetOpcode::G_AND) {
+    int64_t C = VRegAndVal->Value;
+
+    // When we have a greater-than comparison, we can just test if the msb is
+    // zero.
+    if (C == -1 && Pred == CmpInst::ICMP_SGT) {
+      uint64_t Bit = MRI.getType(LHS).getSizeInBits() - 1;
+      emitTestBit(LHS, Bit, /*IsNegative = */ false, DestMBB, MIB);
+      I.eraseFromParent();
+      return true;
+    }
+
+    // When we have a less than comparison, we can just test if the msb is not
+    // zero.
+    if (C == 0 && Pred == CmpInst::ICMP_SLT) {
+      uint64_t Bit = MRI.getType(LHS).getSizeInBits() - 1;
+      emitTestBit(LHS, Bit, /*IsNegative = */ true, DestMBB, MIB);
+      I.eraseFromParent();
+      return true;
+    }
+  }
+
+  if (!VRegAndVal) {
     std::swap(RHS, LHS);
+    VRegAndVal = getConstantVRegValWithLookThrough(RHS, MRI);
+    LHSMI = getDefIgnoringCopies(LHS, MRI);
+  }
 
-  MachineIRBuilder MIB(I);
-  VRegAndVal = getConstantVRegValWithLookThrough(RHS, MRI);
   if (!VRegAndVal || VRegAndVal->Value != 0) {
     // If we can't select a CBZ then emit a cmp + Bcc.
     if (!emitIntegerCompare(CCMI->getOperand(2), CCMI->getOperand(3),
                             CCMI->getOperand(1), MIB))
       return false;
-    const AArch64CC::CondCode CC = changeICMPPredToAArch64CC(
-        (CmpInst::Predicate)CCMI->getOperand(1).getPredicate());
+    const AArch64CC::CondCode CC = changeICMPPredToAArch64CC(Pred);
     MIB.buildInstr(AArch64::Bcc, {}, {}).addImm(CC).addMBB(DestMBB);
     I.eraseFromParent();
     return true;
   }
 
-  // Try to fold things into the branch.
-  const auto Pred = (CmpInst::Predicate)CCMI->getOperand(1).getPredicate();
-  MachineInstr *LHSMI = getDefIgnoringCopies(LHS, MRI);
+  // Try to emit a TB(N)Z for an eq or ne condition.
   if (tryOptAndIntoCompareBranch(LHSMI, VRegAndVal->Value, Pred, DestMBB,
                                  MIB)) {
     I.eraseFromParent();

diff  --git a/llvm/test/CodeGen/AArch64/GlobalISel/tbnz-slt.mir b/llvm/test/CodeGen/AArch64/GlobalISel/tbnz-slt.mir
new file mode 100644
index 000000000000..abdacf8b129f
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/tbnz-slt.mir
@@ -0,0 +1,151 @@
+# 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
+
+...
+---
+name:            dont_commute
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: dont_commute
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   %zero:gpr64 = COPY $xzr
+  ; CHECK:   $xzr = SUBSXrr %zero, %copy, 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
+    %zero:gpr(s64) = G_CONSTANT i64 0
+    %cmp:gpr(s32) = G_ICMP intpred(slt), %zero, %copy(s64)
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR

diff  --git a/llvm/test/CodeGen/AArch64/GlobalISel/tbz-sgt.mir b/llvm/test/CodeGen/AArch64/GlobalISel/tbz-sgt.mir
new file mode 100644
index 000000000000..fe1fd9385e61
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/tbz-sgt.mir
@@ -0,0 +1,151 @@
+# 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 tbz when we have a sgt compare against -1.
+#
+# The bit tested should be the size of the test register minus 1.
+#
+
+...
+---
+name:            tbzx_sgt
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: tbzx_sgt
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   TBZX %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
+    %negative_one:gpr(s64) = G_CONSTANT i64 -1
+    %cmp:gpr(s32) = G_ICMP intpred(sgt), %copy(s64), %negative_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:            tbzw_sgt
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: tbzw_sgt
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr32 = COPY $w0
+  ; CHECK:   TBZW %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
+    %negative_one:gpr(s32) = G_CONSTANT i32 -1
+    %cmp:gpr(s32) = G_ICMP intpred(sgt), %copy(s32), %negative_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:            no_tbz_not_negative_one
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: no_tbz_not_negative_one
+  ; 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 12, %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(sgt), %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:   %negative_one:gpr64 = MOVi64imm -1
+  ; CHECK:   %and:gpr64common = ANDXri %copy, 8000
+  ; CHECK:   $xzr = SUBSXrr %and, %negative_one, implicit-def $nzcv
+  ; CHECK:   Bcc 12, %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(s64) = COPY $x0
+    %bit:gpr(s64) = G_CONSTANT i64 8
+    %negative_one:gpr(s64) = G_CONSTANT i64 -1
+    %c:gpr(s64) = G_CONSTANT i64 8
+    %and:gpr(s64) = G_AND %copy, %bit
+    %cmp:gpr(s32) = G_ICMP intpred(sgt), %and(s64), %negative_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_commute
+alignment:       4
+legalized:       true
+regBankSelected: true
+body:             |
+  ; CHECK-LABEL: name: dont_commute
+  ; CHECK: bb.0:
+  ; CHECK:   successors: %bb.0(0x40000000), %bb.1(0x40000000)
+  ; CHECK:   %copy:gpr64 = COPY $x0
+  ; CHECK:   %negative_one:gpr64 = MOVi64imm -1
+  ; CHECK:   $xzr = SUBSXrr %negative_one, %copy, implicit-def $nzcv
+  ; CHECK:   Bcc 12, %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(s64) = COPY $x0
+    %negative_one:gpr(s64) = G_CONSTANT i64 -1
+    %cmp:gpr(s32) = G_ICMP intpred(sgt), %negative_one, %copy(s64)
+    %cmp_trunc:gpr(s1) = G_TRUNC %cmp(s32)
+    G_BRCOND %cmp_trunc(s1), %bb.1
+    G_BR %bb.0
+  bb.1:
+    RET_ReallyLR


        


More information about the llvm-commits mailing list