From e0b950eae9b2739c40e155339c7d0f9350399980 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Tue, 28 May 2019 16:46:02 +0000
Subject: [PATCH] AMDGPU: Don't enable all lanes with non-CSR VGPR spills

If the only VGPRs used for SGPR spilling were not CSRs, this was
enabling all laness and immediately restoring exec. This is the usual
situation in leaf functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@361848 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AMDGPU/SIFrameLowering.cpp     | 88 +++++++++++++----------
 test/CodeGen/AMDGPU/callee-frame-setup.ll | 16 +++++
 2 files changed, 65 insertions(+), 39 deletions(-)

diff --git a/lib/Target/AMDGPU/SIFrameLowering.cpp b/lib/Target/AMDGPU/SIFrameLowering.cpp
index 1eea77be620..e333154f83b 100644
--- a/lib/Target/AMDGPU/SIFrameLowering.cpp
+++ b/lib/Target/AMDGPU/SIFrameLowering.cpp
@@ -613,30 +613,36 @@ void SIFrameLowering::emitPrologue(MachineFunction &MF,
       .setMIFlag(MachineInstr::FrameSetup);
   }
 
-  if (!FuncInfo->getSGPRSpillVGPRs().empty()) {
-    if (LiveRegs.empty()) {
-      LiveRegs.init(TRI);
-      LiveRegs.addLiveIns(MBB);
-    }
+  // To avoid clobbering VGPRs in lanes that weren't active on function entry,
+  // turn on all lanes before doing the spill to memory.
+  unsigned ScratchExecCopy = AMDGPU::NoRegister;
 
-    // To avoid clobbering VGPRs in lanes that weren't active on function entry,
-    // turn on all lanes before doing the spill to memory.
-    unsigned ScratchExecCopy
-      = findScratchNonCalleeSaveRegister(MF, LiveRegs,
-                                         AMDGPU::SReg_64_XEXECRegClass);
-
-    BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), ScratchExecCopy)
-      .addImm(-1);
-
-    for (const SIMachineFunctionInfo::SGPRSpillVGPRCSR &Reg
-           : FuncInfo->getSGPRSpillVGPRs()) {
-      if (!Reg.FI.hasValue())
-        continue;
-      TII->storeRegToStackSlot(MBB, MBBI, Reg.VGPR, true,
-                               Reg.FI.getValue(), &AMDGPU::VGPR_32RegClass,
-                               &TII->getRegisterInfo());
+  for (const SIMachineFunctionInfo::SGPRSpillVGPRCSR &Reg
+         : FuncInfo->getSGPRSpillVGPRs()) {
+    if (!Reg.FI.hasValue())
+      continue;
+
+    if (ScratchExecCopy == AMDGPU::NoRegister) {
+      if (LiveRegs.empty()) {
+        LiveRegs.init(TRI);
+        LiveRegs.addLiveIns(MBB);
+      }
+
+      ScratchExecCopy
+        = findScratchNonCalleeSaveRegister(MF, LiveRegs,
+                                           AMDGPU::SReg_64_XEXECRegClass);
+
+      BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64),
+              ScratchExecCopy)
+        .addImm(-1);
     }
 
+    TII->storeRegToStackSlot(MBB, MBBI, Reg.VGPR, true,
+                             Reg.FI.getValue(), &AMDGPU::VGPR_32RegClass,
+                             &TII->getRegisterInfo());
+  }
+
+  if (ScratchExecCopy != AMDGPU::NoRegister) {
     // FIXME: Split block and make terminator.
     BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
       .addReg(ScratchExecCopy);
@@ -654,27 +660,31 @@ void SIFrameLowering::emitEpilogue(MachineFunction &MF,
   MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator();
   DebugLoc DL;
 
-  if (!FuncInfo->getSGPRSpillVGPRs().empty()) {
-    // See emitPrologue
-    LivePhysRegs LiveRegs(*ST.getRegisterInfo());
-    LiveRegs.addLiveIns(MBB);
+  unsigned ScratchExecCopy = AMDGPU::NoRegister;
+  for (const SIMachineFunctionInfo::SGPRSpillVGPRCSR &Reg
+         : FuncInfo->getSGPRSpillVGPRs()) {
+    if (!Reg.FI.hasValue())
+      continue;
 
-    unsigned ScratchExecCopy
-      = findScratchNonCalleeSaveRegister(MF, LiveRegs,
-                                         AMDGPU::SReg_64_XEXECRegClass);
-
-    BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), ScratchExecCopy)
-      .addImm(-1);
-
-    for (const SIMachineFunctionInfo::SGPRSpillVGPRCSR &Reg
-           : FuncInfo->getSGPRSpillVGPRs()) {
-      if (!Reg.FI.hasValue())
-        continue;
-      TII->loadRegFromStackSlot(MBB, MBBI, Reg.VGPR,
-                                Reg.FI.getValue(), &AMDGPU::VGPR_32RegClass,
-                                &TII->getRegisterInfo());
+    if (ScratchExecCopy == AMDGPU::NoRegister) {
+      // See emitPrologue
+      LivePhysRegs LiveRegs(*ST.getRegisterInfo());
+      LiveRegs.addLiveIns(MBB);
+
+      ScratchExecCopy
+        = findScratchNonCalleeSaveRegister(MF, LiveRegs,
+                                           AMDGPU::SReg_64_XEXECRegClass);
+
+      BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), ScratchExecCopy)
+        .addImm(-1);
     }
 
+    TII->loadRegFromStackSlot(MBB, MBBI, Reg.VGPR,
+                              Reg.FI.getValue(), &AMDGPU::VGPR_32RegClass,
+                              &TII->getRegisterInfo());
+  }
+
+  if (ScratchExecCopy != AMDGPU::NoRegister) {
     // FIXME: Split block and make terminator.
     BuildMI(MBB, MBBI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
       .addReg(ScratchExecCopy);
diff --git a/test/CodeGen/AMDGPU/callee-frame-setup.ll b/test/CodeGen/AMDGPU/callee-frame-setup.ll
index ebd6f96a5b8..bc9160772e2 100644
--- a/test/CodeGen/AMDGPU/callee-frame-setup.ll
+++ b/test/CodeGen/AMDGPU/callee-frame-setup.ll
@@ -135,5 +135,21 @@ define void @callee_func_sgpr_spill_no_calls(i32 %in) #0 {
   ret void
 }
 
+; Has no spilled CSR VGPRs used for SGPR spilling, so no need to
+; enable all lanes and restore.
+
+; GCN-LABEL: {{^}}spill_only_csr_sgpr:
+; GCN: s_waitcnt
+; GCN-NEXT: v_writelane_b32 v0, s42, 0
+; GCN-NEXT: ;;#ASMSTART
+; GCN-NEXT: ; clobber s42
+; GCN-NEXT: ;;#ASMEND
+; GCN-NEXT: v_readlane_b32 s42, v0, 0
+; GCN-NEXT: s_setpc_b64
+define void @spill_only_csr_sgpr() {
+  call void asm sideeffect "; clobber s42", "~{s42}"()
+  ret void
+}
+
 attributes #0 = { nounwind }
 attributes #1 = { nounwind "no-frame-pointer-elim"="true" }
-- 
2.40.0