Merging r293310:

------------------------------------------------------------------------
r293310 | arsenm | 2017-01-27 09:42:26 -0800 (Fri, 27 Jan 2017) | 8 lines

AMDGPU: Enable FeatureFlatForGlobal on Volcanic Islands

Accomplishes what r292982 was supposed to, which ended up
only really making the necessary test changes.

This should be applied to the 4.0 branch.

Patch by Vedran Miletić <vedran@miletic.net>
------------------------------------------------------------------------

git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_40@293329 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index c5d7d85d40558a82ae3044ef4ad43adbd5b7cac3..13022009af169c543f1c89a26ee0f8cfbaa5b5ba 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -282,12 +282,6 @@ def FeatureEnableSIScheduler : SubtargetFeature<"si-scheduler",
   "Enable SI Machine Scheduler"
 >;
 
-def FeatureNoAddr64 : SubtargetFeature<"mubuf-no-addr64",
-  "NoAddr64",
-  "true",
-  "MUBUF instructions have addr64 bit"
->;
-
 // Unless +-flat-for-global is specified, turn on FlatForGlobal for
 // all OS-es on VI and newer hardware to avoid assertion failures due
 // to missing ADDR64 variants of MUBUF instructions.
@@ -297,8 +291,7 @@ def FeatureNoAddr64 : SubtargetFeature<"mubuf-no-addr64",
 def FeatureFlatForGlobal : SubtargetFeature<"flat-for-global",
   "FlatForGlobal",
   "true",
-  "Force to generate flat instruction for global",
-  [FeatureNoAddr64]
+  "Force to generate flat instruction for global"
 >;
 
 // Dummy feature used to disable assembler instructions.
@@ -350,8 +343,7 @@ def FeatureVolcanicIslands : SubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
    FeatureWavefrontSize64, FeatureFlatAddressSpace, FeatureGCN,
    FeatureGCN3Encoding, FeatureCIInsts, Feature16BitInsts,
    FeatureSMemRealTime, FeatureVGPRIndexMode, FeatureMovrel,
-   FeatureScalarStores, FeatureInv2PiInlineImm, 
-   FeatureNoAddr64
+   FeatureScalarStores, FeatureInv2PiInlineImm
   ]
 >;
 

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index e2ec839a9536f20ae521c82460e4715ae286b533..c35a67de1d7f8fab7cd59d52e9fff687fa734881 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -48,6 +48,13 @@ AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT,
 
   ParseSubtargetFeatures(GPU, FullFS);
 
+  // Unless +-flat-for-global is specified, turn on FlatForGlobal for all OS-es
+  // on VI and newer hardware to avoid assertion failures due to missing ADDR64
+  // variants of MUBUF instructions.
+  if (!hasAddr64() && !FS.contains("flat-for-global")) {
+    FlatForGlobal = true;
+  }
+
   // FIXME: I don't think think Evergreen has any useful support for
   // denormals, but should be checked. Should we issue a warning somewhere
   // if someone tries to enable these?
@@ -83,7 +90,6 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     FP64Denormals(false),
     FPExceptions(false),
     FlatForGlobal(false),
-    NoAddr64(false),
     UnalignedScratchAccess(false),
     UnalignedBufferAccess(false),
 

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.h b/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 51ca36fb7b35a2a29ddcd81574560a89ffb26570..0e3cb7dc1f8762d6305e31e5ab6f451b924cfda9 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -86,7 +86,6 @@ protected:
   bool FP64Denormals;
   bool FPExceptions;
   bool FlatForGlobal;
-  bool NoAddr64;
   bool UnalignedScratchAccess;
   bool UnalignedBufferAccess;
   bool EnableXNACK;

diff --git a/test/CodeGen/AMDGPU/ci-use-flat-for-global.ll b/test/CodeGen/AMDGPU/ci-use-flat-for-global.ll
deleted file mode 100644 (file)
index 8227d4c..0000000
--- a/test/CodeGen/AMDGPU/ci-use-flat-for-global.ll
+++ /dev/null
@@ -1,26 +0,0 @@
-; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=+flat-for-global < %s | FileCheck -check-prefix=HSA -check-prefix=HSA-DEFAULT -check-prefix=ALL %s
-; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=-flat-for-global < %s | FileCheck -check-prefix=HSA -check-prefix=HSA-NODEFAULT -check-prefix=ALL %s
-; RUN: llc -mtriple=amdgcn-- -mcpu=kaveri -mattr=-flat-for-global < %s | FileCheck -check-prefix=NOHSA-DEFAULT -check-prefix=ALL %s
-; RUN: llc -mtriple=amdgcn-- -mcpu=kaveri -mattr=+flat-for-global < %s | FileCheck -check-prefix=NOHSA-NODEFAULT -check-prefix=ALL %s
-
-
-; There are no stack objects even though flat is used by default, so
-; flat_scratch_init should be disabled.
-
-; ALL-LABEL: {{^}}test:
-; HSA: .amd_kernel_code_t
-; HSA: enable_sgpr_flat_scratch_init = 0
-; HSA: .end_amd_kernel_code_t
-
-; ALL-NOT: flat_scr
-
-; HSA-DEFAULT: flat_store_dword
-; HSA-NODEFAULT: buffer_store_dword
-
-; NOHSA-DEFAULT: buffer_store_dword
-; NOHSA-NODEFAULT: flat_store_dword
-define void @test(i32 addrspace(1)* %out) {
-entry:
-  store i32 0, i32 addrspace(1)* %out
-  ret void
-}

diff --git a/test/CodeGen/AMDGPU/fcanonicalize.f16.ll b/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
index fb693de0e39dba49ecddbd8b6de2a293252fec2b..ad3992f4cd03f2a806a072dc3ef86f63393dc7ff 100644 (file)
--- a/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
+++ b/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
@@ -167,6 +167,6 @@ define void @test_fold_canonicalize_snan3_value_f16(half addrspace(1)* %out) #1
 }
 
 attributes #0 = { nounwind readnone }
-attributes #1 = { nounwind }
-attributes #2 = { nounwind "target-features"="-fp16-denormals,-fp16-denormals" }
-attributes #3 = { nounwind "target-features"="+fp16-denormals,+fp64-denormals" }
+attributes #1 = { nounwind "target-features"="-flat-for-global" }
+attributes #2 = { nounwind "target-features"="-flat-for-global,-fp16-denormals,-fp16-denormals" }
+attributes #3 = { nounwind "target-features"="-flat-for-global,+fp16-denormals,+fp64-denormals" }

diff --git a/test/CodeGen/AMDGPU/flat-for-global-subtarget-feature.ll b/test/CodeGen/AMDGPU/flat-for-global-subtarget-feature.ll
new file mode 100644 (file)
index 0000000..df9ba00
--- /dev/null
+++ b/test/CodeGen/AMDGPU/flat-for-global-subtarget-feature.ll
@@ -0,0 +1,54 @@
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=+flat-for-global < %s | FileCheck -check-prefix=HSA -check-prefix=HSA-DEFAULT -check-prefix=ALL %s
+; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=-flat-for-global < %s | FileCheck -check-prefix=HSA -check-prefix=HSA-NODEFAULT -check-prefix=ALL %s
+; RUN: llc -mtriple=amdgcn-- -mcpu=tonga < %s | FileCheck -check-prefix=HSA-NOADDR64 -check-prefix=ALL %s
+; RUN: llc -mtriple=amdgcn-- -mcpu=kaveri -mattr=-flat-for-global < %s | FileCheck -check-prefix=NOHSA-DEFAULT -check-prefix=ALL %s
+; RUN: llc -mtriple=amdgcn-- -mcpu=kaveri -mattr=+flat-for-global < %s | FileCheck -check-prefix=NOHSA-NODEFAULT -check-prefix=ALL %s
+; RUN: llc -mtriple=amdgcn-- -mcpu=tonga < %s | FileCheck -check-prefix=NOHSA-NOADDR64 -check-prefix=ALL %s
+
+
+; There are no stack objects even though flat is used by default, so
+; flat_scratch_init should be disabled.
+
+; ALL-LABEL: {{^}}test:
+; HSA: .amd_kernel_code_t
+; HSA: enable_sgpr_flat_scratch_init = 0
+; HSA: .end_amd_kernel_code_t
+
+; ALL-NOT: flat_scr
+
+; HSA-DEFAULT: flat_store_dword
+; HSA-NODEFAULT: buffer_store_dword
+; HSA-NOADDR64: flat_store_dword
+
+; NOHSA-DEFAULT: buffer_store_dword
+; NOHSA-NODEFAULT: flat_store_dword
+; NOHSA-NOADDR64: flat_store_dword
+define void @test(i32 addrspace(1)* %out) {
+entry:
+  store i32 0, i32 addrspace(1)* %out
+  ret void
+}
+
+; HSA-DEFAULT: flat_store_dword
+; HSA-NODEFAULT: buffer_store_dword
+; HSA-NOADDR64: flat_store_dword
+
+; NOHSA-DEFAULT: buffer_store_dword
+; NOHSA-NODEFAULT: flat_store_dword
+; NOHSA-NOADDR64: flat_store_dword
+define void @test_addr64(i32 addrspace(1)* %out) {
+entry:
+  %out.addr = alloca i32 addrspace(1)*, align 4
+
+  store i32 addrspace(1)* %out, i32 addrspace(1)** %out.addr, align 4
+  %ld0 = load i32 addrspace(1)*, i32 addrspace(1)** %out.addr, align 4
+
+  %arrayidx = getelementptr inbounds i32, i32 addrspace(1)* %ld0, i32 0
+  store i32 1, i32 addrspace(1)* %arrayidx, align 4
+
+  %ld1 = load i32 addrspace(1)*, i32 addrspace(1)** %out.addr, align 4
+  %arrayidx1 = getelementptr inbounds i32, i32 addrspace(1)* %ld1, i32 1
+  store i32 2, i32 addrspace(1)* %arrayidx1, align 4
+
+  ret void
+}

diff --git a/test/CodeGen/AMDGPU/v_madak_f16.ll b/test/CodeGen/AMDGPU/v_madak_f16.ll
index 7707d6b1baad469acf3b57c5d47f1cfa8dd6a14d..df220d7a977b255a372c232a3f74df8323ab9c01 100644 (file)
--- a/test/CodeGen/AMDGPU/v_madak_f16.ll
+++ b/test/CodeGen/AMDGPU/v_madak_f16.ll
@@ -1,5 +1,5 @@
 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
-; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-fp16-denormals,-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
 ; GCN-LABEL: {{^}}madak_f16
 ; GCN: buffer_load_ushort v[[A_F16:[0-9]+]]