release/21.x: [TailDup] Delay aggressive computed-goto taildup to after RegAlloc. (#150911)

llvm/include/llvm/CodeGen/MachineBasicBlock.h

@@ -323,10 +323,11 @@ class MachineBasicBlock
   const MachineFunction *getParent() const { return xParent; }
   MachineFunction *getParent() { return xParent; }
 
-  /// Returns true if the original IR terminator is an `indirectbr`. This
-  /// typically corresponds to a `goto` in C, rather than jump tables.
-  bool terminatorIsComputedGoto() const {
-    return back().isIndirectBranch() &&
+  /// Returns true if the original IR terminator is an `indirectbr` with
+  /// successor blocks. This typically corresponds to a `goto` in C, rather than
+  /// jump tables.
+  bool terminatorIsComputedGotoWithSuccessors() const {
+    return back().isIndirectBranch() && !succ_empty() &&
            llvm::all_of(successors(), [](const MachineBasicBlock *Succ) {
              return Succ->isIRBlockAddressTaken();
            });

llvm/lib/CodeGen/TailDuplicator.cpp

@@ -604,12 +604,21 @@ bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
   bool HasComputedGoto = false;
   if (!TailBB.empty()) {
     HasIndirectbr = TailBB.back().isIndirectBranch();
-    HasComputedGoto = TailBB.terminatorIsComputedGoto();
+    HasComputedGoto = TailBB.terminatorIsComputedGotoWithSuccessors();
   }
 
   if (HasIndirectbr && PreRegAlloc)
     MaxDuplicateCount = TailDupIndirectBranchSize;
 
+  // Allow higher limits when the block has computed-gotos and running after
+  // register allocation. NB. This basically unfactors computed gotos that were
+  // factored early on in the compilation process to speed up edge based data
+  // flow. If we do not unfactor them again, it can seriously pessimize code
+  // with many computed jumps in the source code, such as interpreters.
+  // Therefore we do not restrict the computed gotos.
+  if (HasComputedGoto && !PreRegAlloc)
+    MaxDuplicateCount = std::max(MaxDuplicateCount, 10u);
+
   // Check the instructions in the block to determine whether tail-duplication
   // is invalid or unlikely to be profitable.
   unsigned InstrCount = 0;
@@ -663,12 +672,7 @@ bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
   // Duplicating a BB which has both multiple predecessors and successors will
   // may cause huge amount of PHI nodes. If we want to remove this limitation,
   // we have to address https://github.com/llvm/llvm-project/issues/78578.
-  // NB. This basically unfactors computed gotos that were factored early on in
-  // the compilation process to speed up edge based data flow. If we do not
-  // unfactor them again, it can seriously pessimize code with many computed
-  // jumps in the source code, such as interpreters. Therefore we do not
-  // restrict the computed gotos.
-  if (!HasComputedGoto && TailBB.pred_size() > TailDupPredSize &&
+  if (PreRegAlloc && TailBB.pred_size() > TailDupPredSize &&
       TailBB.succ_size() > TailDupSuccSize) {
     // If TailBB or any of its successors contains a phi, we may have to add a
     // large number of additional phis with additional incoming values.

llvm/test/CodeGen/AArch64/late-taildup-computed-goto.ll

@@ -0,0 +1,143 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -tail-dup-pred-size=2 -tail-dup-succ-size=2 -o - %s | FileCheck %s
+
+target triple = "arm64-apple-macosx13.0.0"
+
+@opcode.targets = local_unnamed_addr constant [6 x ptr] [ptr blockaddress(@test_interp, %op1.bb), ptr blockaddress(@test_interp, %op6.bb), ptr blockaddress(@test_interp, %loop.header), ptr blockaddress(@test_interp, %op2.bb), ptr blockaddress(@test_interp, %op4.bb), ptr blockaddress(@test_interp, %op5.bb)]
+
+define void @test_interp(ptr %frame, ptr %dst) {
+; CHECK-LABEL: test_interp:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    stp x24, x23, [sp, #-64]! ; 16-byte Folded Spill
+; CHECK-NEXT:    stp x22, x21, [sp, #16] ; 16-byte Folded Spill
+; CHECK-NEXT:    stp x20, x19, [sp, #32] ; 16-byte Folded Spill
+; CHECK-NEXT:    stp x29, x30, [sp, #48] ; 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 64
+; CHECK-NEXT:    .cfi_offset w30, -8
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    .cfi_offset w19, -24
+; CHECK-NEXT:    .cfi_offset w20, -32
+; CHECK-NEXT:    .cfi_offset w21, -40
+; CHECK-NEXT:    .cfi_offset w22, -48
+; CHECK-NEXT:    .cfi_offset w23, -56
+; CHECK-NEXT:    .cfi_offset w24, -64
+; CHECK-NEXT:  Lloh0:
+; CHECK-NEXT:    adrp x21, _opcode.targets@PAGE
+; CHECK-NEXT:  Lloh1:
+; CHECK-NEXT:    add x21, x21, _opcode.targets@PAGEOFF
+; CHECK-NEXT:    mov x24, xzr
+; CHECK-NEXT:    add x8, x21, xzr, lsl #3
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    mov x23, xzr
+; CHECK-NEXT:    mov w22, #1 ; =0x1
+; CHECK-NEXT:    add x24, x24, #1
+; CHECK-NEXT:    br x8
+; CHECK-NEXT:  Ltmp0: ; Block address taken
+; CHECK-NEXT:  LBB0_1: ; %loop.header
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    add x8, x21, x24, lsl #3
+; CHECK-NEXT:    mov x20, xzr
+; CHECK-NEXT:    mov x23, xzr
+; CHECK-NEXT:    add x24, x24, #1
+; CHECK-NEXT:    br x8
+; CHECK-NEXT:  Ltmp1: ; Block address taken
+; CHECK-NEXT:  LBB0_2: ; %op1.bb
+; CHECK-NEXT:    str xzr, [x19]
+; CHECK-NEXT:  Ltmp2: ; Block address taken
+; CHECK-NEXT:  LBB0_3: ; %op6.bb
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr x0, [x20, #-8]!
+; CHECK-NEXT:    ldr x8, [x0, #8]
+; CHECK-NEXT:    str x22, [x0]
+; CHECK-NEXT:    ldr x8, [x8, #48]
+; CHECK-NEXT:    blr x8
+; CHECK-NEXT:    add x8, x21, x24, lsl #3
+; CHECK-NEXT:    add x24, x24, #1
+; CHECK-NEXT:    br x8
+; CHECK-NEXT:  Ltmp3: ; Block address taken
+; CHECK-NEXT:  LBB0_4: ; %op2.bb
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    add x8, x21, x24, lsl #3
+; CHECK-NEXT:    mov x20, xzr
+; CHECK-NEXT:    str x23, [x19]
+; CHECK-NEXT:    mov x23, xzr
+; CHECK-NEXT:    add x24, x24, #1
+; CHECK-NEXT:    br x8
+; CHECK-NEXT:  Ltmp4: ; Block address taken
+; CHECK-NEXT:  LBB0_5: ; %op4.bb
+; CHECK-NEXT:  Ltmp5: ; Block address taken
+; CHECK-NEXT:  LBB0_6: ; %op5.bb
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    str x23, [x19]
+; CHECK-NEXT:    ldur x8, [x23, #12]
+; CHECK-NEXT:    ldur x9, [x20, #-8]
+; CHECK-NEXT:    add x23, x23, #20
+; CHECK-NEXT:    stp x8, x9, [x20, #-8]
+; CHECK-NEXT:    add x8, x21, x24, lsl #3
+; CHECK-NEXT:    add x20, x20, #8
+; CHECK-NEXT:    add x24, x24, #1
+; CHECK-NEXT:    br x8
+; CHECK-NEXT:    .loh AdrpAdd Lloh0, Lloh1
+entry:
+  br label %loop.header
+
+loop.header:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %op1.bb ], [ %iv.next, %op2.bb ], [ %iv.next, %op4.bb ], [ %iv.next, %op5.bb ], [ %iv.next, %op6.bb ], [ %iv.next, %loop.header ]
+  %stack.pointer = phi ptr [ %frame, %entry ], [ %stack.8, %op1.bb ], [ null, %op2.bb ], [ %stack.next, %op4.bb ], [ %stack.next.2, %op5.bb ], [ %stack.4, %op6.bb ], [ null, %loop.header ]
+  %next.instr = phi ptr [ null, %entry ], [ %next.instr, %op1.bb ], [ null, %op2.bb ], [ %next.instr.20, %op4.bb ], [ %next.instr.21, %op5.bb ], [ %next.instr, %op6.bb ], [ null, %loop.header ]
+  %iv.next = add i64 %iv, 1
+  %next_op = getelementptr [6 x ptr], ptr @opcode.targets, i64 0, i64 %iv
+  indirectbr ptr %next_op, [label %op1.bb, label %op6.bb, label %loop.header, label %op2.bb, label %op4.bb, label %op5.bb]
+
+op1.bb:
+  store ptr null, ptr %dst, align 8
+  %stack.8 = getelementptr i8, ptr %stack.pointer, i64 -8
+  %l.0 = load ptr, ptr %stack.8, align 8
+  store i64 1, ptr %l.0, align 8
+  %gep.0 = getelementptr i8, ptr %l.0, i64 8
+  %l.1 = load ptr, ptr %gep.0, align 8
+  %gep.1 = getelementptr i8, ptr %l.1, i64 48
+  %l.2 = load ptr, ptr %gep.1, align 8
+  tail call void %l.2(ptr nonnull %l.0)
+  br label %loop.header
+
+op2.bb:
+  store ptr %next.instr, ptr %dst, align 8
+  br label %loop.header
+
+op4.bb:
+  store ptr %next.instr, ptr %dst, align 8
+  %next.instr.20 = getelementptr i8, ptr %next.instr, i64 20
+  %stack.2 = getelementptr i8, ptr %stack.pointer, i64 -8
+  %l.3 = load ptr, ptr %stack.2, align 8
+  %next.instr.12 = getelementptr i8, ptr %next.instr, i64 12
+  %next.instr.12.val = load ptr, ptr %next.instr.12, align 2
+  store ptr %next.instr.12.val, ptr %stack.2, align 8
+  store ptr %l.3, ptr %stack.pointer, align 8
+  %stack.next = getelementptr i8, ptr %stack.pointer, i64 8
+  br label %loop.header
+
+op5.bb:
+  store ptr %next.instr, ptr %dst, align 8
+  %next.instr.21 = getelementptr i8, ptr %next.instr, i64 20
+  %stack.3 = getelementptr i8, ptr %stack.pointer, i64 -8
+  %l.4 = load ptr, ptr %stack.3, align 8
+  %next.instr.2 = getelementptr i8, ptr %next.instr, i64 12
+  %next.instr.2.val = load ptr, ptr %next.instr.2, align 2
+  store ptr %next.instr.2.val, ptr %stack.3, align 8
+  store ptr %l.4, ptr %stack.pointer, align 8
+  %stack.next.2 = getelementptr i8, ptr %stack.pointer, i64 8
+  br label %loop.header
+
+op6.bb:
+  %stack.4 = getelementptr i8, ptr %stack.pointer, i64 -8
+  %l.5 = load ptr, ptr %stack.4, align 8
+  store i64 1, ptr %l.5, align 8
+  %gep.5 = getelementptr i8, ptr %l.5, i64 8
+  %l.6 = load ptr, ptr %gep.5, align 8
+  %gep.6 = getelementptr i8, ptr %l.6, i64 48
+  %l.7 = load ptr, ptr %gep.6, align 8
+  tail call void %l.7(ptr nonnull %l.5)
+  br label %loop.header
+}

llvm/test/CodeGen/X86/tail-dup-computed-goto.mir → llvm/test/CodeGen/X86/early-tail-dup-computed-goto.mir

@@ -1,6 +1,8 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
 # RUN: llc -mtriple=x86_64-unknown-linux-gnu -run-pass=early-tailduplication -tail-dup-pred-size=1 -tail-dup-succ-size=1 %s -o - | FileCheck %s
-# Check that only the computed goto is not be restrict by tail-dup-pred-size and tail-dup-succ-size.
+#
+# Check that only the computed goto and others are restricted by tail-dup-pred-size and tail-dup-succ-size.
+#
 --- |
   @computed_goto.dispatch = constant [5 x ptr] [ptr null, ptr blockaddress(@computed_goto, %bb1), ptr blockaddress(@computed_goto, %bb2), ptr blockaddress(@computed_goto, %bb3), ptr blockaddress(@computed_goto, %bb4)]
   declare i64 @f0()
@@ -30,54 +32,54 @@ tracksRegLiveness: true
 body:             |
   ; CHECK-LABEL: name: computed_goto
   ; CHECK: bb.0:
-  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
   ; CHECK-NEXT:   CALL64pcrel32 target-flags(x86-plt) @f0, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
   ; CHECK-NEXT:   ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
-  ; CHECK-NEXT:   [[COPY:%[0-9]+]]:gr64_nosp = COPY $rax
-  ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:gr64_nosp = COPY [[COPY]]
-  ; CHECK-NEXT:   JMP64m $noreg, 8, [[COPY]], @computed_goto.dispatch, $noreg
+  ; CHECK-NEXT:   [[COPY:%[0-9]+]]:gr64 = COPY $rax
+  ; CHECK-NEXT:   JMP_1 %bb.5
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.1.bb1 (ir-block-address-taken %ir-block.bb1):
-  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
   ; CHECK-NEXT:   CALL64pcrel32 target-flags(x86-plt) @f1, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
   ; CHECK-NEXT:   ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
-  ; CHECK-NEXT:   [[COPY2:%[0-9]+]]:gr64_nosp = COPY $rax
-  ; CHECK-NEXT:   [[COPY3:%[0-9]+]]:gr64_nosp = COPY [[COPY2]]
-  ; CHECK-NEXT:   JMP64m $noreg, 8, [[COPY2]], @computed_goto.dispatch, $noreg
+  ; CHECK-NEXT:   [[COPY1:%[0-9]+]]:gr64 = COPY $rax
+  ; CHECK-NEXT:   JMP_1 %bb.5
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.2.bb2 (ir-block-address-taken %ir-block.bb2):
-  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
   ; CHECK-NEXT:   CALL64pcrel32 target-flags(x86-plt) @f2, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
   ; CHECK-NEXT:   ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
-  ; CHECK-NEXT:   [[COPY4:%[0-9]+]]:gr64_nosp = COPY $rax
-  ; CHECK-NEXT:   [[COPY5:%[0-9]+]]:gr64_nosp = COPY [[COPY4]]
-  ; CHECK-NEXT:   JMP64m $noreg, 8, [[COPY4]], @computed_goto.dispatch, $noreg
+  ; CHECK-NEXT:   [[COPY2:%[0-9]+]]:gr64 = COPY $rax
+  ; CHECK-NEXT:   JMP_1 %bb.5
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.3.bb3 (ir-block-address-taken %ir-block.bb3):
-  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
   ; CHECK-NEXT:   CALL64pcrel32 target-flags(x86-plt) @f3, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
   ; CHECK-NEXT:   ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
-  ; CHECK-NEXT:   [[COPY6:%[0-9]+]]:gr64_nosp = COPY $rax
-  ; CHECK-NEXT:   [[COPY7:%[0-9]+]]:gr64_nosp = COPY [[COPY6]]
-  ; CHECK-NEXT:   JMP64m $noreg, 8, [[COPY6]], @computed_goto.dispatch, $noreg
+  ; CHECK-NEXT:   [[COPY3:%[0-9]+]]:gr64 = COPY $rax
+  ; CHECK-NEXT:   JMP_1 %bb.5
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT: bb.4.bb4 (ir-block-address-taken %ir-block.bb4):
-  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT:   successors: %bb.5(0x80000000)
   ; CHECK-NEXT: {{  $}}
   ; CHECK-NEXT:   ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
   ; CHECK-NEXT:   CALL64pcrel32 target-flags(x86-plt) @f4, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax
   ; CHECK-NEXT:   ADJCALLSTACKUP64 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
-  ; CHECK-NEXT:   [[COPY8:%[0-9]+]]:gr64_nosp = COPY $rax
-  ; CHECK-NEXT:   [[COPY9:%[0-9]+]]:gr64_nosp = COPY [[COPY8]]
-  ; CHECK-NEXT:   JMP64m $noreg, 8, [[COPY8]], @computed_goto.dispatch, $noreg
+  ; CHECK-NEXT:   [[COPY4:%[0-9]+]]:gr64 = COPY $rax
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT: bb.5:
+  ; CHECK-NEXT:   successors: %bb.1(0x20000000), %bb.2(0x20000000), %bb.3(0x20000000), %bb.4(0x20000000)
+  ; CHECK-NEXT: {{  $}}
+  ; CHECK-NEXT:   [[PHI:%[0-9]+]]:gr64_nosp = PHI [[COPY]], %bb.0, [[COPY4]], %bb.4, [[COPY3]], %bb.3, [[COPY2]], %bb.2, [[COPY1]], %bb.1
+  ; CHECK-NEXT:   JMP64m $noreg, 8, [[PHI]], @computed_goto.dispatch, $noreg
   bb.0:
     ADJCALLSTACKDOWN64 0, 0, 0, implicit-def dead $rsp, implicit-def dead $eflags, implicit-def dead $ssp, implicit $rsp, implicit $ssp
     CALL64pcrel32 target-flags(x86-plt) @f0, csr_64, implicit $rsp, implicit $ssp, implicit-def $rsp, implicit-def $ssp, implicit-def $rax