AMDGPU: Assume ECC is enabled by default if supported

The test should really be checking for the property directly in the
code object headers, but there are problems with this. I don't see
this directly represented in the text form, and for the binary
emission this is depending on a function level subtarget feature to
emit a global flag.

llvm-svn: 357558

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -281,6 +281,12 @@ def FeatureDot2Insts : SubtargetFeature<"dot2-insts",
   "Has v_dot2_f32_f16, v_dot2_i32_i16, v_dot2_u32_u16, v_dot4_u32_u8, v_dot8_u32_u4 instructions"
 >;
 
+def FeatureDoesNotSupportSRAMECC : SubtargetFeature<"no-sram-ecc-support",
+  "DoesNotSupportSRAMECC",
+  "true",
+  "Hardware does not support SRAM ECC"
+>;
+
 def FeatureSRAMECC : SubtargetFeature<"sram-ecc",
   "EnableSRAMECC",
   "true",
@@ -439,14 +445,16 @@ def FeatureSouthernIslands : GCNSubtargetFeatureGeneration<"SOUTHERN_ISLANDS",
     "southern-islands",
   [FeatureFP64, FeatureLocalMemorySize32768, FeatureMIMG_R128,
   FeatureWavefrontSize64,
-  FeatureLDSBankCount32, FeatureMovrel, FeatureTrigReducedRange]
+  FeatureLDSBankCount32, FeatureMovrel, FeatureTrigReducedRange,
+  FeatureDoesNotSupportSRAMECC]
 >;
 
 def FeatureSeaIslands : GCNSubtargetFeatureGeneration<"SEA_ISLANDS",
     "sea-islands",
   [FeatureFP64, FeatureLocalMemorySize65536, FeatureMIMG_R128,
   FeatureWavefrontSize64, FeatureFlatAddressSpace,
-  FeatureCIInsts, FeatureMovrel, FeatureTrigReducedRange]
+  FeatureCIInsts, FeatureMovrel, FeatureTrigReducedRange,
+  FeatureDoesNotSupportSRAMECC]
 >;
 
 def FeatureVolcanicIslands : GCNSubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
@@ -457,7 +465,7 @@ def FeatureVolcanicIslands : GCNSubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
    FeatureSMemRealTime, FeatureVGPRIndexMode, FeatureMovrel,
    FeatureScalarStores, FeatureInv2PiInlineImm,
    FeatureSDWA, FeatureSDWAOutModsVOPC, FeatureSDWAMac, FeatureDPP,
-   FeatureIntClamp, FeatureTrigReducedRange
+   FeatureIntClamp, FeatureTrigReducedRange, FeatureDoesNotSupportSRAMECC
   ]
 >;
 
@@ -550,19 +558,22 @@ def FeatureISAVersion9_0_0 : FeatureSet<
   [FeatureGFX9,
    FeatureMadMixInsts,
    FeatureLDSBankCount32,
-   FeatureCodeObjectV3]>;
+   FeatureCodeObjectV3,
+   FeatureDoesNotSupportSRAMECC]>;
 
 def FeatureISAVersion9_0_2 : FeatureSet<
   [FeatureGFX9,
    FeatureMadMixInsts,
    FeatureLDSBankCount32,
    FeatureXNACK,
+   FeatureDoesNotSupportSRAMECC,
    FeatureCodeObjectV3]>;
 
 def FeatureISAVersion9_0_4 : FeatureSet<
   [FeatureGFX9,
    FeatureLDSBankCount32,
    FeatureFmaMixInsts,
+   FeatureDoesNotSupportSRAMECC,
    FeatureCodeObjectV3]>;
 
 def FeatureISAVersion9_0_6 : FeatureSet<

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

@@ -64,7 +64,7 @@ R600Subtarget::initializeSubtargetDependencies(const Triple &TT,
 
 GCNSubtarget &
 GCNSubtarget::initializeSubtargetDependencies(const Triple &TT,
-                                                 StringRef GPU, StringRef FS) {
+                                              StringRef GPU, StringRef FS) {
   // Determine default and user-specified characteristics
   // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
   // enabled, but some instructions do not respect them and they run at the
@@ -77,7 +77,8 @@ GCNSubtarget::initializeSubtargetDependencies(const Triple &TT,
   // Similarly we want enable-prt-strict-null to be on by default and not to
   // unset everything else if it is disabled
 
-  SmallString<256> FullFS("+promote-alloca,+load-store-opt,");
+  // Assuming ECC is enabled is the conservative default.
+  SmallString<256> FullFS("+promote-alloca,+load-store-opt,+sram-ecc,");
 
   if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA.
     FullFS += "+flat-for-global,+unaligned-buffer-access,+trap-handler,";
@@ -129,6 +130,14 @@ GCNSubtarget::initializeSubtargetDependencies(const Triple &TT,
 
   HasFminFmaxLegacy = getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS;
 
+  // ECC is on by default, but turn it off if the hardware doesn't support it
+  // anyway. This matters for the gfx9 targets with d16 loads, but don't support
+  // ECC.
+  if (DoesNotSupportSRAMECC && EnableSRAMECC) {
+    ToggleFeature(AMDGPU::FeatureSRAMECC);
+    EnableSRAMECC = false;
+  }
+
   return *this;
 }
 
@@ -206,6 +215,7 @@ GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     HasDot1Insts(false),
     HasDot2Insts(false),
     EnableSRAMECC(false),
+    DoesNotSupportSRAMECC(false),
     FlatAddressSpace(false),
     FlatInstOffsets(false),
     FlatGlobalInsts(false),

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

@@ -332,6 +332,7 @@ protected:
   bool HasDot1Insts;
   bool HasDot2Insts;
   bool EnableSRAMECC;
+  bool DoesNotSupportSRAMECC;
   bool FlatAddressSpace;
   bool FlatInstOffsets;
   bool FlatGlobalInsts;

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

@@ -84,6 +84,10 @@ class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
     AMDGPU::FeatureTrapHandler,
     AMDGPU::FeatureCodeObjectV3,
 
+    // The default assumption needs to be ecc is enabled, but no directly
+    // exposed operations depend on it, so it can be safely inlined.
+    AMDGPU::FeatureSRAMECC,
+
     // Perf-tuning features
     AMDGPU::FeatureFastFMAF32,
     AMDGPU::HalfRate64Ops

llvm/test/CodeGen/AMDGPU/load-hi16.ll

@@ -1,5 +1,5 @@
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-sroa=0 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX900 %s
-; RUN: llc -march=amdgcn -mcpu=gfx906 -amdgpu-sroa=0 -mattr=-promote-alloca,+sram-ecc -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX906,NO-D16-HI %s
+; RUN: llc -march=amdgcn -mcpu=gfx906 -amdgpu-sroa=0 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX906,NO-D16-HI %s
 ; RUN: llc -march=amdgcn -mcpu=fiji -amdgpu-sroa=0 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX803,NO-D16-HI %s
 
 ; GCN-LABEL: {{^}}load_local_lo_hi_v2i16_multi_use_lo:

llvm/test/CodeGen/AMDGPU/sram-ecc-default.ll

@@ -0,0 +1,24 @@
+; RUN: llc -march=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -mattr=+sram-ecc < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -mattr=-sram-ecc < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx902 -mattr=+sram-ecc < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx904 -mattr=+sram-ecc < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx906 -mattr=+sram-ecc < %s | FileCheck -check-prefixes=GCN,ECC %s
+; RUN: llc -march=amdgcn -mcpu=gfx906 -mattr=-sram-ecc < %s | FileCheck -check-prefixes=GCN,NO-ECC %s
+
+; Make sure the correct set of targets are marked with
+; FeatureDoesNotSupportSRAMECC, and +sram-ecc is ignored if it's never
+; supported.
+
+; GCN-LABEL: {{^}}load_global_hi_v2i16_reglo_vreg:
+; NO-ECC: global_load_short_d16_hi
+; ECC: global_load_ushort
+define void @load_global_hi_v2i16_reglo_vreg(i16 addrspace(1)* %in, i16 %reg) {
+entry:
+  %gep = getelementptr inbounds i16, i16 addrspace(1)* %in, i64 -2047
+  %load = load i16, i16 addrspace(1)* %gep
+  %build0 = insertelement <2 x i16> undef, i16 %reg, i32 0
+  %build1 = insertelement <2 x i16> %build0, i16 %load, i32 1
+  store <2 x i16> %build1, <2 x i16> addrspace(1)* undef
+  ret void
+}

llvm/test/Transforms/Inline/AMDGPU/inline-target-feature-sram-ecc.ll

@@ -0,0 +1,70 @@
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -inline < %s | FileCheck %s
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -S -passes='cgscc(inline)' < %s | FileCheck %s
+
+; sram-ecc can be safely ignored when inlining, since no intrinisics
+; or other directly exposed operations depend on it.
+
+define i32 @func_default() #0 {
+  ret i32 0
+}
+
+define i32 @func_ecc_enabled() #1 {
+  ret i32 0
+}
+
+define i32 @func_ecc_disabled() #2 {
+  ret i32 0
+}
+
+; CHECK-LABEL: @default_call_default(
+; CHECK-NEXT: ret i32 0
+define i32 @default_call_default() #0 {
+  %call = call i32 @func_default()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_enabled_call_default(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_enabled_call_default() #1 {
+  %call = call i32 @func_default()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_enabled_call_ecc_enabled(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_enabled_call_ecc_enabled() #1 {
+  %call = call i32 @func_ecc_enabled()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_enabled_call_ecc_disabled(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_enabled_call_ecc_disabled() #1 {
+  %call = call i32 @func_ecc_disabled()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_disabled_call_default(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_disabled_call_default() #2 {
+  %call = call i32 @func_default()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_disabled_call_ecc_enabled(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_disabled_call_ecc_enabled() #2 {
+  %call = call i32 @func_ecc_enabled()
+  ret i32 %call
+}
+
+; CHECK-LABEL: @ecc_disabled_call_ecc_disabled(
+; CHECK-NEXT: ret i32 0
+define i32 @ecc_disabled_call_ecc_disabled() #2 {
+  %call = call i32 @func_ecc_disabled()
+  ret i32 %call
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind "target-features"="+sram-ecc" }
+attributes #2 = { nounwind "target-features"="-sram-ecc" }