From b151df8f6db2eb2ef1053bae166071aa68777ccb Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Wed, 4 Oct 2017 22:59:12 +0000
Subject: [PATCH] AMDGPU: Fix not accounting for instruction size in bundles

These were counted as 0. Fixes branch limit exceeded errors
in some large programs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@314944 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AMDGPU/SIInstrInfo.cpp          | 15 +++++-
 lib/Target/AMDGPU/SIInstrInfo.h            |  1 +
 test/CodeGen/AMDGPU/branch-relax-bundle.ll | 53 ++++++++++++++++++++++
 3 files changed, 68 insertions(+), 1 deletion(-)
 create mode 100644 test/CodeGen/AMDGPU/branch-relax-bundle.ll

diff --git a/lib/Target/AMDGPU/SIInstrInfo.cpp b/lib/Target/AMDGPU/SIInstrInfo.cpp
index 72d6119a6ba..807e60632ae 100644
--- a/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -4371,6 +4371,18 @@ unsigned SIInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
   return AMDGPU::NoRegister;
 }
 
+unsigned SIInstrInfo::getInstBundleSize(const MachineInstr &MI) const {
+  unsigned Size = 0;
+  MachineBasicBlock::const_instr_iterator I = MI.getIterator();
+  MachineBasicBlock::const_instr_iterator E = MI.getParent()->instr_end();
+  while (++I != E && I->isInsideBundle()) {
+    assert(!I->isBundle() && "No nested bundle!");
+    Size += getInstSizeInBytes(*I);
+  }
+
+  return Size;
+}
+
 unsigned SIInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
   unsigned Opc = MI.getOpcode();
   const MCInstrDesc &Desc = getMCOpcodeFromPseudo(Opc);
@@ -4414,9 +4426,10 @@ unsigned SIInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
   case TargetOpcode::IMPLICIT_DEF:
   case TargetOpcode::KILL:
   case TargetOpcode::DBG_VALUE:
-  case TargetOpcode::BUNDLE:
   case TargetOpcode::EH_LABEL:
     return 0;
+  case TargetOpcode::BUNDLE:
+    return getInstBundleSize(MI);
   case TargetOpcode::INLINEASM: {
     const MachineFunction *MF = MI.getParent()->getParent();
     const char *AsmStr = MI.getOperand(0).getSymbolName();
diff --git a/lib/Target/AMDGPU/SIInstrInfo.h b/lib/Target/AMDGPU/SIInstrInfo.h
index 93513e2de15..59725aeb1dc 100644
--- a/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/lib/Target/AMDGPU/SIInstrInfo.h
@@ -818,6 +818,7 @@ public:
   unsigned isStoreToStackSlot(const MachineInstr &MI,
                               int &FrameIndex) const override;
 
+  unsigned getInstBundleSize(const MachineInstr &MI) const;
   unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
 
   bool mayAccessFlatAddressSpace(const MachineInstr &MI) const;
diff --git a/test/CodeGen/AMDGPU/branch-relax-bundle.ll b/test/CodeGen/AMDGPU/branch-relax-bundle.ll
new file mode 100644
index 00000000000..9f0b8d3553e
--- /dev/null
+++ b/test/CodeGen/AMDGPU/branch-relax-bundle.ll
@@ -0,0 +1,53 @@
+; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs -amdgpu-s-branch-bits=5 < %s | FileCheck -check-prefix=GCN %s
+
+; Restrict maximum branch to between +15 and -16 dwords
+
+; Instructions inside a bundle were collectively counted as
+; 0-bytes. Make sure this is accounted for when estimating branch
+; distances
+
+; Bundle used for address in call sequence: 20 bytes
+; s_getpc_b64
+; s_add_u32
+; s_addc_u32
+
+; plus additional overhead
+; s_setpc_b64
+; and some register copies
+
+declare void @func() #0
+
+; GCN-LABEL: {{^}}bundle_size:
+; GCN: s_cbranch_scc0 [[BB_EXPANSION:BB[0-9]+_[0-9]+]]
+; GCN: s_getpc_b64
+; GCN-NEXT: s_add_u32
+; GCN-NEXT: s_addc_u32
+; GCN-NEXT: s_setpc_b64
+
+; GCN: {{^}}[[BB_EXPANSION]]:
+; GCN: s_getpc_b64
+; GCN: s_add_u32 s{{[0-9]+}}, s{{[0-9]+}}, func@
+; GCN: s_addc_u32 s{{[0-9]+}}, s{{[0-9]+}}, func@
+; GCN: s_swappc_b64
+define amdgpu_kernel void @bundle_size(i32 addrspace(1)* %arg, i32 %cnd) #0 {
+bb:
+  %cmp = icmp eq i32 %cnd, 0
+  br i1 %cmp, label %bb3, label %bb2 ; +8 dword branch
+
+bb2:
+  call void @func()
+  call void asm sideeffect
+  "v_nop_e64
+   v_nop_e64
+   v_nop_e64
+   v_nop_e64
+   v_nop_e64", ""() #0
+  br label %bb3
+
+bb3:
+  store volatile i32 %cnd, i32 addrspace(1)* %arg
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
-- 
2.40.0