[PATCH] D72865: [AMDGPU] allow multi-dword flat scratch access since GFX9

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec created this revision.
rampitec added a reviewer: arsenm.
Herald added subscribers: kerbowa, hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely, kzhuravl.
Herald added a project: LLVM.

This is supported starting with GFX9.


https://reviews.llvm.org/D72865

Files:
  llvm/lib/Target/AMDGPU/SIISelLowering.cpp
  llvm/test/CodeGen/AMDGPU/flat-address-space.ll


Index: llvm/test/CodeGen/AMDGPU/flat-address-space.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/flat-address-space.ll
+++ llvm/test/CodeGen/AMDGPU/flat-address-space.ll
@@ -1,7 +1,8 @@
 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=bonaire < %s | FileCheck -check-prefixes=CHECK,CIVI %s
 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tonga -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI,HSA %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,CIVI,HSA,CIVI-HSA %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,HSA,GFX9 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=CHECK,HSA,GFX10 %s
 
 ; CHECK-LABEL: {{^}}store_flat_i32:
 ; CHECK-DAG: s_load_dwordx2 s{{\[}}[[LO_SREG:[0-9]+]]:[[HI_SREG:[0-9]+]]],
@@ -145,8 +146,10 @@
 }
 
 ; CHECK-LABEL: flat_scratch_multidword_load:
-; HSA: flat_load_dword
-; HSA: flat_load_dword
+; CIVI-HSA: flat_load_dword v
+; CIVI-HSA: flat_load_dword v
+; GFX9:  flat_load_dwordx2
+; GFX10: flat_load_dwordx2
 ; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
 define amdgpu_kernel void @flat_scratch_multidword_load() {
   %scratch = alloca <2 x i32>, addrspace(5)
@@ -156,8 +159,10 @@
 }
 
 ; CHECK-LABEL: flat_scratch_multidword_store:
-; HSA: flat_store_dword
-; HSA: flat_store_dword
+; CIVI-HSA: flat_store_dword v
+; CIVI-HSA: flat_store_dword v
+; GFX9:  flat_store_dwordx2
+; GFX10: flat_store_dwordx2
 ; FIXME: These tests are broken for os = mesa3d, becasue it doesn't initialize flat_scr
 define amdgpu_kernel void @flat_scratch_multidword_store() {
   %scratch = alloca <2 x i32>, addrspace(5)
Index: llvm/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -7380,7 +7380,9 @@
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   // If there is a possibilty that flat instruction access scratch memory
   // then we need to use the same legalization rules we use for private.
-  if (AS == AMDGPUAS::FLAT_ADDRESS)
+  // Since GFX9 flat instructions can handle multi-dword.
+  if (AS == AMDGPUAS::FLAT_ADDRESS &&
+      getSubtarget()->getGeneration() < AMDGPUSubtarget::GFX9)
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 
@@ -7883,7 +7885,9 @@
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   // If there is a possibilty that flat instruction access scratch memory
   // then we need to use the same legalization rules we use for private.
-  if (AS == AMDGPUAS::FLAT_ADDRESS)
+  // Since GFX9 flat instructions can handle multi-dword.
+  if (AS == AMDGPUAS::FLAT_ADDRESS &&
+      getSubtarget()->getGeneration() < AMDGPUSubtarget::GFX9)
     AS = MFI->hasFlatScratchInit() ?
          AMDGPUAS::PRIVATE_ADDRESS : AMDGPUAS::GLOBAL_ADDRESS;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D72865.238582.patch
Type: text/x-patch
Size: 3263 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200116/d3c8e15d/attachment.bin>