AMDGPU: add +xnack feature

Summary:
Enabling this feature will account for the two SGPRs used by the hardware
to store the XNACK_MASK physically.

The hardware only requires this reservation when the XNACK feature is
explicitly enabled. At some point, HSA will probably want to do that, but
it does increase SGPR register pressure, so leave it disabled by default
for now (but do add a small test).

Reviewers: arsenm, tstellarAMD

Subscribers: arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D15869

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@256794 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index d4af8d2e48d103ed5b48d6e8ae71e8938e3c79af..db869cf7dd8b43173e20f9e30f65871cf70cfa0a 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -118,6 +118,11 @@ def FeatureFlatAddressSpace : SubtargetFeature<"flat-address-space",
         "true",
         "Support flat address space">;
 
         "true",
         "Support flat address space">;
 

+def FeatureXNACK : SubtargetFeature<"xnack",
+        "EnableXNACK",
+        "true",
+        "Enable XNACK support">;
+

 def FeatureVGPRSpilling : SubtargetFeature<"vgpr-spilling",
         "EnableVGPRSpilling",
         "true",
 def FeatureVGPRSpilling : SubtargetFeature<"vgpr-spilling",
         "EnableVGPRSpilling",
         "true",

diff --git a/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index ba71dc05a8fcb3bf1d252cde56a5f54d43ff45db..9c37902643771167546c33546cc5aaab59997894 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -417,13 +417,13 @@ void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
     }
   }
 
     }
   }
 

-  if (VCCUsed || FlatUsed)
+  if (VCCUsed || FlatUsed || STM.isXNACKEnabled()) {

     MaxSGPR += 2;
 
     MaxSGPR += 2;
 

-  if (FlatUsed) {
-    MaxSGPR += 2;
-    // 2 additional for VI+.
-    if (STM.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
+    if (FlatUsed)
+      MaxSGPR += 2;
+
+    if (STM.isXNACKEnabled())

       MaxSGPR += 2;
   }
 
       MaxSGPR += 2;
   }
 


@@ -620,6 +620,9 @@ void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF,
   if (MFI->hasDispatchPtr())
     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
 
   if (MFI->hasDispatchPtr())
     header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR;
 

+  if (STM.isXNACKEnabled())
+    header.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED;
+

   header.kernarg_segment_byte_size = MFI->ABIArgOffset;
   header.wavefront_sgpr_count = KernelInfo.NumSGPR;
   header.workitem_vgpr_count = KernelInfo.NumVGPR;
   header.kernarg_segment_byte_size = MFI->ABIArgOffset;
   header.wavefront_sgpr_count = KernelInfo.NumSGPR;
   header.workitem_vgpr_count = KernelInfo.NumVGPR;

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 44e0c47877a9ff2a071f1a459cde56487fc42cb6..c6af5b93d257e381aef6048581fda1ef4b990f14 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -73,6 +73,7 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
       CaymanISA(false), FlatAddressSpace(false), FlatForGlobal(false),
       EnableIRStructurizer(true), EnablePromoteAlloca(false), EnableIfCvt(true),
       EnableLoadStoreOpt(false), EnableUnsafeDSOffsetFolding(false),
       CaymanISA(false), FlatAddressSpace(false), FlatForGlobal(false),
       EnableIRStructurizer(true), EnablePromoteAlloca(false), EnableIfCvt(true),
       EnableLoadStoreOpt(false), EnableUnsafeDSOffsetFolding(false),

+      EnableXNACK(false),

       WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
       EnableVGPRSpilling(false), SGPRInitBug(false), IsGCN(false),
       GCN1Encoding(false), GCN3Encoding(false), CIInsts(false), LDSBankCount(0),
       WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
       EnableVGPRSpilling(false), SGPRInitBug(false), IsGCN(false),
       GCN1Encoding(false), GCN3Encoding(false), CIInsts(false), LDSBankCount(0),

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.h b/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 9c7bb88f8f4a60858f10450221190db33fdb0db5..d3712276d5e7337c98b39d88e4203fcd00cc9eef 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -76,6 +76,7 @@ private:
   bool EnableIfCvt;
   bool EnableLoadStoreOpt;
   bool EnableUnsafeDSOffsetFolding;
   bool EnableIfCvt;
   bool EnableLoadStoreOpt;
   bool EnableUnsafeDSOffsetFolding;

+  bool EnableXNACK;

   unsigned WavefrontSize;
   bool CFALUBug;
   int LocalMemorySize;
   unsigned WavefrontSize;
   bool CFALUBug;
   int LocalMemorySize;


@@ -290,6 +291,10 @@ public:
   }
   bool isVGPRSpillingEnabled(const SIMachineFunctionInfo *MFI) const;
 
   }
   bool isVGPRSpillingEnabled(const SIMachineFunctionInfo *MFI) const;
 

+  bool isXNACKEnabled() const {
+    return EnableXNACK;
+  }
+

   unsigned getMaxWavesPerCU() const {
     if (getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)
       return 10;
   unsigned getMaxWavesPerCU() const {
     if (getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)
       return 10;

diff --git a/lib/Target/AMDGPU/SIRegisterInfo.cpp b/lib/Target/AMDGPU/SIRegisterInfo.cpp
index 353c2b930637645f9dce0aaf04435913258d1453..2afa00996609fb10d89986de9e539a88d3f2fffe 100644 (file)
--- a/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -37,13 +37,17 @@ unsigned SIRegisterInfo::reservedPrivateSegmentBufferReg(
   const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
   if (ST.hasSGPRInitBug()) {
     unsigned BaseIdx = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4 - 4;
   const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
   if (ST.hasSGPRInitBug()) {
     unsigned BaseIdx = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4 - 4;

+    if (ST.isXNACKEnabled())
+      BaseIdx -= 4;
+

     unsigned BaseReg(AMDGPU::SGPR_32RegClass.getRegister(BaseIdx));
     return getMatchingSuperReg(BaseReg, AMDGPU::sub0, &AMDGPU::SReg_128RegClass);
   }
 
   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
     unsigned BaseReg(AMDGPU::SGPR_32RegClass.getRegister(BaseIdx));
     return getMatchingSuperReg(BaseReg, AMDGPU::sub0, &AMDGPU::SReg_128RegClass);
   }
 
   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {

-    // 98/99 need to be reserved for flat_scr, and 100/101 for vcc. This is the
-    // next sgpr128 down.
+    // 98/99 need to be reserved for flat_scr or 96/97 for flat_scr and
+    // 98/99 for xnack_mask, and 100/101 for vcc. This is the next sgpr128 down
+    // either way.

     return AMDGPU::SGPR92_SGPR93_SGPR94_SGPR95;
   }
 
     return AMDGPU::SGPR92_SGPR93_SGPR94_SGPR95;
   }
 


@@ -54,13 +58,25 @@ unsigned SIRegisterInfo::reservedPrivateSegmentWaveByteOffsetReg(
   const MachineFunction &MF) const {
   const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
   if (ST.hasSGPRInitBug()) {
   const MachineFunction &MF) const {
   const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
   if (ST.hasSGPRInitBug()) {

-    unsigned Idx = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4 - 5;
+    unsigned Idx;
+
+    if (!ST.isXNACKEnabled())
+      Idx = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4 - 5;
+    else
+      Idx = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 6 - 1;
+

     return AMDGPU::SGPR_32RegClass.getRegister(Idx);
   }
 
   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
     return AMDGPU::SGPR_32RegClass.getRegister(Idx);
   }
 
   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {

-    // Next register before reservations for flat_scr and vcc.
-    return AMDGPU::SGPR97;
+    if (!ST.isXNACKEnabled()) {
+      // Next register before reservations for flat_scr and vcc.
+      return AMDGPU::SGPR97;
+    } else {
+      // Next register before reservations for flat_scr, xnack_mask, vcc,
+      // and scratch resource.
+      return AMDGPU::SGPR91;
+    }

   }
 
   return AMDGPU::SGPR95;
   }
 
   return AMDGPU::SGPR95;


@@ -86,6 +102,9 @@ BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
     // for VCC/FLAT_SCR.
     reserveRegisterTuples(Reserved, AMDGPU::SGPR98_SGPR99);
     reserveRegisterTuples(Reserved, AMDGPU::SGPR100_SGPR101);
     // for VCC/FLAT_SCR.
     reserveRegisterTuples(Reserved, AMDGPU::SGPR98_SGPR99);
     reserveRegisterTuples(Reserved, AMDGPU::SGPR100_SGPR101);

+
+    if (ST.isXNACKEnabled())
+      reserveRegisterTuples(Reserved, AMDGPU::SGPR96_SGPR97);

   }
 
   // Tonga and Iceland can only allocate a fixed number of SGPRs due
   }
 
   // Tonga and Iceland can only allocate a fixed number of SGPRs due


@@ -93,9 +112,11 @@ BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
   if (ST.hasSGPRInitBug()) {
     unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
     // Reserve some SGPRs for FLAT_SCRATCH and VCC (4 SGPRs).
   if (ST.hasSGPRInitBug()) {
     unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
     // Reserve some SGPRs for FLAT_SCRATCH and VCC (4 SGPRs).

-    // Assume XNACK_MASK is unused.

     unsigned Limit = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4;
 
     unsigned Limit = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4;
 

+    if (ST.isXNACKEnabled())
+      Limit -= 2;
+

     for (unsigned i = Limit; i < NumSGPRs; ++i) {
       unsigned Reg = AMDGPU::SGPR_32RegClass.getRegister(i);
       reserveRegisterTuples(Reserved, Reg);
     for (unsigned i = Limit; i < NumSGPRs; ++i) {
       unsigned Reg = AMDGPU::SGPR_32RegClass.getRegister(i);
       reserveRegisterTuples(Reserved, Reg);

diff --git a/test/CodeGen/AMDGPU/flat-scratch-reg.ll b/test/CodeGen/AMDGPU/flat-scratch-reg.ll
index e2ae3353ae1d3b7a5f3dbd2cabf60c9fd53936b3..9aea7c7734314591db9ffa42fc402777a77cf542 100644 (file)
--- a/test/CodeGen/AMDGPU/flat-scratch-reg.ll
+++ b/test/CodeGen/AMDGPU/flat-scratch-reg.ll
@@ -1,8 +1,10 @@
-; RUN: llc < %s -march=amdgcn -mcpu=kaveri -verify-machineinstrs | FileCheck %s --check-prefix=GCN --check-prefix=CI
-; RUN: llc < %s -march=amdgcn -mcpu=fiji -verify-machineinstrs | FileCheck %s --check-prefix=GCN --check-prefix=VI
+; RUN: llc < %s -march=amdgcn -mcpu=kaveri -verify-machineinstrs | FileCheck %s --check-prefix=GCN --check-prefix=CI --check-prefix=NO-XNACK
+; RUN: llc < %s -march=amdgcn -mcpu=fiji -verify-machineinstrs | FileCheck %s --check-prefix=GCN --check-prefix=VI --check-prefix=NO-XNACK
+; RUN: llc < %s -march=amdgcn -mcpu=carrizo -mattr=+xnack -verify-machineinstrs | FileCheck %s --check-prefix=GCN --check-prefix=XNACK

 
 ; GCN-LABEL: {{^}}no_vcc_no_flat:
 
 ; GCN-LABEL: {{^}}no_vcc_no_flat:

-; GCN: ; NumSgprs: 8
+; NO-XNACK: ; NumSgprs: 8
+; XNACK: ; NumSgprs: 12

 define void @no_vcc_no_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7}"()
 define void @no_vcc_no_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7}"()


@@ -10,7 +12,8 @@ entry:
 }
 
 ; GCN-LABEL: {{^}}vcc_no_flat:
 }
 
 ; GCN-LABEL: {{^}}vcc_no_flat:

-; GCN: ; NumSgprs: 10
+; NO-XNACK: ; NumSgprs: 10
+; XNACK: ; NumSgprs: 12

 define void @vcc_no_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{VCC}"()
 define void @vcc_no_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{VCC}"()


@@ -19,7 +22,8 @@ entry:
 
 ; GCN-LABEL: {{^}}no_vcc_flat:
 ; CI: ; NumSgprs: 12
 
 ; GCN-LABEL: {{^}}no_vcc_flat:
 ; CI: ; NumSgprs: 12

-; VI: ; NumSgprs: 14
+; VI: ; NumSgprs: 12
+; XNACK: ; NumSgprs: 14

 define void @no_vcc_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{FLAT_SCR}"()
 define void @no_vcc_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{FLAT_SCR}"()


@@ -28,7 +32,8 @@ entry:
 
 ; GCN-LABEL: {{^}}vcc_flat:
 ; CI: ; NumSgprs: 12
 
 ; GCN-LABEL: {{^}}vcc_flat:
 ; CI: ; NumSgprs: 12

-; VI: ; NumSgprs: 14
+; VI: ; NumSgprs: 12
+; XNACK: ; NumSgprs: 14

 define void @vcc_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{VCC},~{FLAT_SCR}"()
 define void @vcc_flat() {
 entry:
   call void asm sideeffect "", "~{SGPR7},~{VCC},~{FLAT_SCR}"()