AMDGPU: Ensure both wavesize features are not set (#159234)

Make sure we cannot be in a mode with both wavesizes. This
prevents assertions in a future change. This should probably
just be an error, but we do not have a good way to report
errors from the MCSubtargetInfo constructor.

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -1238,6 +1238,19 @@ def FeatureSetPrioIncWgInst : SubtargetFeature<"setprio-inc-wg-inst",
 // Subtarget Features (options and debugging)
 //===------------------------------------------------------------===//
 
+// Ugly hack to accomodate assembling modules with mixed
+// wavesizes. Ideally we would have a mapping symbol in assembly which
+// would keep track of which sections of code should be treated as
+// wave32 and wave64. Instead what users do is assemble with both
+// wavesizes enabled. We translate this into this special mode so this
+// only influences assembler behavior and nothing else.
+def FeatureAssemblerPermissiveWavesize : SubtargetFeature<
+  "assembler-permissive-wavesize",
+  "AssemblerPermissiveWavesize",
+  "true",
+  "allow parsing wave32 and wave64 variants of instructions"
+>;
+
 class FeatureMaxPrivateElementSize<int size> : SubtargetFeature<
   "max-private-element-size-"#size,
   "MaxPrivateElementSize",

llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

@@ -1246,6 +1246,12 @@ raw_ostream &operator <<(raw_ostream &OS, AMDGPUOperand::Modifiers Mods) {
 // AsmParser
 //===----------------------------------------------------------------------===//
 
+// TODO: define GET_SUBTARGET_FEATURE_NAME
+#define GET_REGISTER_MATCHER
+#include "AMDGPUGenAsmMatcher.inc"
+#undef GET_REGISTER_MATCHER
+#undef GET_SUBTARGET_FEATURE_NAME
+
 // Holds info related to the current kernel, e.g. count of SGPRs used.
 // Kernel scope begins at .amdgpu_hsa_kernel directive, ends at next
 // .amdgpu_hsa_kernel or at EOF.
@@ -1536,6 +1542,10 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
     return AMDGPU::isGFX10_BEncoding(getSTI());
   }
 
+  bool isWave32() const { return getAvailableFeatures()[Feature_isWave32Bit]; }
+
+  bool isWave64() const { return getAvailableFeatures()[Feature_isWave64Bit]; }
+
   bool hasInv2PiInlineImm() const {
     return getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm];
   }
@@ -1603,6 +1613,8 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
     return &MII;
   }
 
+  // FIXME: This should not be used. Instead, should use queries derived from
+  // getAvailableFeatures().
   const FeatureBitset &getFeatureBits() const {
     return getSTI().getFeatureBits();
   }
@@ -2259,9 +2271,8 @@ bool AMDGPUOperand::isSDWAInt32Operand() const {
 }
 
 bool AMDGPUOperand::isBoolReg() const {
-  auto FB = AsmParser->getFeatureBits();
-  return isReg() && ((FB[AMDGPU::FeatureWavefrontSize64] && isSCSrc_b64()) ||
-                     (FB[AMDGPU::FeatureWavefrontSize32] && isSCSrc_b32()));
+  return isReg() && ((AsmParser->isWave64() && isSCSrc_b64()) ||
+                     (AsmParser->isWave32() && isSCSrc_b32()));
 }
 
 uint64_t AMDGPUOperand::applyInputFPModifiers(uint64_t Val, unsigned Size) const
@@ -5025,9 +5036,8 @@ bool AMDGPUAsmParser::validateDPP(const MCInst &Inst,
 
 // Check if VCC register matches wavefront size
 bool AMDGPUAsmParser::validateVccOperand(MCRegister Reg) const {
-  auto FB = getFeatureBits();
-  return (FB[AMDGPU::FeatureWavefrontSize64] && Reg == AMDGPU::VCC) ||
-    (FB[AMDGPU::FeatureWavefrontSize32] && Reg == AMDGPU::VCC_LO);
+  return (Reg == AMDGPU::VCC && isWave64()) ||
+         (Reg == AMDGPU::VCC_LO && isWave32());
 }
 
 // One unique literal can be used. VOP3 literal is only allowed in GFX10+
@@ -5717,7 +5727,7 @@ bool AMDGPUAsmParser::checkUnsupportedInstruction(StringRef Mnemo,
   // Check if this instruction may be used with a different wavesize.
   if (isGFX10Plus() && getFeatureBits()[AMDGPU::FeatureWavefrontSize64] &&
       !getFeatureBits()[AMDGPU::FeatureWavefrontSize32]) {
-
+    // FIXME: Use getAvailableFeatures, and do not manually recompute
     FeatureBitset FeaturesWS32 = getFeatureBits();
     FeaturesWS32.flip(AMDGPU::FeatureWavefrontSize64)
         .flip(AMDGPU::FeatureWavefrontSize32);
@@ -6472,10 +6482,10 @@ bool AMDGPUAsmParser::ParseAMDKernelCodeTValue(StringRef ID,
     if (C.code_properties & AMD_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32) {
       if (!isGFX10Plus())
         return TokError("enable_wavefront_size32=1 is only allowed on GFX10+");
-      if (!getFeatureBits()[AMDGPU::FeatureWavefrontSize32])
+      if (!isWave32())
         return TokError("enable_wavefront_size32=1 requires +WavefrontSize32");
     } else {
-      if (!getFeatureBits()[AMDGPU::FeatureWavefrontSize64])
+      if (!isWave64())
         return TokError("enable_wavefront_size32=0 requires +WavefrontSize64");
     }
   }
@@ -6484,10 +6494,10 @@ bool AMDGPUAsmParser::ParseAMDKernelCodeTValue(StringRef ID,
     if (C.wavefront_size == 5) {
       if (!isGFX10Plus())
         return TokError("wavefront_size=5 is only allowed on GFX10+");
-      if (!getFeatureBits()[AMDGPU::FeatureWavefrontSize32])
+      if (!isWave32())
         return TokError("wavefront_size=5 requires +WavefrontSize32");
     } else if (C.wavefront_size == 6) {
-      if (!getFeatureBits()[AMDGPU::FeatureWavefrontSize64])
+      if (!isWave64())
         return TokError("wavefront_size=6 requires +WavefrontSize64");
     }
   }
@@ -10390,7 +10400,6 @@ LLVMInitializeAMDGPUAsmParser() {
   RegisterMCAsmParser<AMDGPUAsmParser> B(getTheGCNTarget());
 }
 
-#define GET_REGISTER_MATCHER
 #define GET_MATCHER_IMPLEMENTATION
 #define GET_MNEMONIC_SPELL_CHECKER
 #define GET_MNEMONIC_CHECKER

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -99,6 +99,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool EnableDS128 = false;
   bool EnablePRTStrictNull = false;
   bool DumpCode = false;
+  bool AssemblerPermissiveWavesize = false;
 
   // Subtarget statically properties set by tablegen
   bool FP64 = false;

llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUMCTargetDesc.cpp

@@ -82,20 +82,36 @@ createAMDGPUMCSubtargetInfo(const Triple &TT, StringRef CPU, StringRef FS) {
   MCSubtargetInfo *STI =
       createAMDGPUMCSubtargetInfoImpl(TT, CPU, /*TuneCPU*/ CPU, FS);
 
+  bool IsWave64 = STI->hasFeature(AMDGPU::FeatureWavefrontSize64);
+  bool IsWave32 = STI->hasFeature(AMDGPU::FeatureWavefrontSize32);
+
   // FIXME: We should error for the default target.
   if (STI->getFeatureBits().none())
     STI->ToggleFeature(AMDGPU::FeatureSouthernIslands);
 
-  if (!STI->hasFeature(AMDGPU::FeatureWavefrontSize64) &&
-      !STI->hasFeature(AMDGPU::FeatureWavefrontSize32)) {
+  if (!IsWave64 && !IsWave32) {
     // If there is no default wave size it must be a generation before gfx10,
     // these have FeatureWavefrontSize64 in their definition already. For gfx10+
     // set wave32 as a default.
     STI->ToggleFeature(AMDGPU::isGFX10Plus(*STI)
                            ? AMDGPU::FeatureWavefrontSize32
                            : AMDGPU::FeatureWavefrontSize64);
+  } else if (IsWave64 && IsWave32) {
+    // The wave size is mutually exclusive. If both somehow end up set, wave32
+    // wins if supported.
+    STI->ToggleFeature(AMDGPU::supportsWave32(*STI)
+                           ? AMDGPU::FeatureWavefrontSize64
+                           : AMDGPU::FeatureWavefrontSize32);
+
+    // If both wavesizes were manually requested, hack in a feature to permit
+    // assembling modules with mixed wavesizes.
+    STI->ToggleFeature(AMDGPU::FeatureAssemblerPermissiveWavesize);
   }
 
+  assert((STI->hasFeature(AMDGPU::FeatureWavefrontSize64) !=
+          STI->hasFeature(AMDGPU::FeatureWavefrontSize32)) &&
+         "wavesize features are mutually exclusive");
+
   return STI;
 }
 

llvm/lib/Target/AMDGPU/SIInstrInfo.td

@@ -7,9 +7,11 @@
 //===----------------------------------------------------------------------===//
 
 def isWave32 : Predicate<"Subtarget->isWave32()">,
-  AssemblerPredicate <(all_of FeatureWavefrontSize32)>;
+  AssemblerPredicate <(any_of FeatureWavefrontSize32,
+                              FeatureAssemblerPermissiveWavesize)>;
 def isWave64 : Predicate<"Subtarget->isWave64()">,
-  AssemblerPredicate <(all_of FeatureWavefrontSize64)>;
+  AssemblerPredicate <(any_of FeatureWavefrontSize64,
+                              FeatureAssemblerPermissiveWavesize)>;
 
 class AMDGPUMnemonicAlias<string From, string To, string VariantName = "">
     : MnemonicAlias<From, To, VariantName>, PredicateControl;

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

@@ -1568,6 +1568,11 @@ bool hasArchitectedFlatScratch(const MCSubtargetInfo &STI);
 bool hasMAIInsts(const MCSubtargetInfo &STI);
 bool hasVOPD(const MCSubtargetInfo &STI);
 bool hasDPPSrc1SGPR(const MCSubtargetInfo &STI);
+
+inline bool supportsWave32(const MCSubtargetInfo &STI) {
+  return AMDGPU::isGFX10Plus(STI) && !AMDGPU::isGFX1250(STI);
+}
+
 int getTotalNumVGPRs(bool has90AInsts, int32_t ArgNumAGPR, int32_t ArgNumVGPR);
 unsigned hasKernargPreload(const MCSubtargetInfo &STI);
 bool hasSMRDSignedImmOffset(const MCSubtargetInfo &ST);

llvm/test/MC/AMDGPU/wavesize-feature-unsupported-target.s

@@ -0,0 +1,23 @@
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx1250 -mattr=+wavefrontsize64 -o - %s | FileCheck -check-prefix=GFX1250 %s
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx900 -mattr=+wavefrontsize32 -o - %s | FileCheck -check-prefix=GFX900 %s
+
+// Make sure setting both modes is supported at the same time.
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32,+wavefrontsize64 %s | FileCheck -check-prefixes=GFX10 %s
+
+// Test that there is no assertion when using an explicit
+// wavefrontsize attribute on a target which does not support it.
+
+// GFX1250: v_add_f64_e32 v[0:1], 1.0, v[0:1]
+// GFX900: v_add_f64 v[0:1], 1.0, v[0:1]
+// GFX10: v_add_f64 v[0:1], 1.0, v[0:1]
+v_add_f64 v[0:1], 1.0, v[0:1]
+
+// GFX1250: v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+// GFX900: v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+// GFX10: v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+
+// GFX1250: v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+// GFX900: v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+// GFX10: v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+v_cndmask_b32 v1, v2, v3, s[0:1]

llvm/test/MC/Disassembler/AMDGPU/gfx10_vopc.txt

@@ -1,6 +1,6 @@
 # RUN: llvm-mc -triple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32 -disassemble -show-encoding < %s | FileCheck -strict-whitespace -check-prefix=W32 %s
 # RUN: llvm-mc -triple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64 -disassemble -show-encoding < %s | FileCheck -strict-whitespace -check-prefix=W64 %s
-
+# RUN: llvm-mc -triple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize32,+wavefrontsize64 -disassemble -show-encoding < %s | FileCheck -strict-whitespace -check-prefix=W32 %s
 
 # W32: v_cmp_class_f32_e32 vcc_lo, -1, v2      ; encoding: [0xc1,0x04,0x10,0x7d]
 # W64: v_cmp_class_f32_e32 vcc, -1, v2         ; encoding: [0xc1,0x04,0x10,0x7d]

llvm/test/MC/Disassembler/AMDGPU/gfx1250_wave64_feature.s

@@ -0,0 +1,13 @@
+# RUN: llvm-mc -triple=amdgcn -mcpu=gfx1250 -mattr=+wavefrontsize64 -disassemble -o - %s | FileCheck %s
+
+# Make sure there's no assertion when trying to use an unsupported
+# wave64 on a wave32-only target
+
+# CHECK: v_add_f64_e32 v[0:1], 1.0, v[0:1]
+0xf2,0x00,0x00,0x04
+
+# CHECK: v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+0x00,0x00,0x4a,0xd4,0xf2,0x02,0x00,0x00
+
+# CHECK: v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+0x01,0x00,0x01,0xd5,0x02,0x07,0x02,0x00

llvm/test/MC/Disassembler/AMDGPU/gfx9_wave32_feature.txt

@@ -0,0 +1,13 @@
+# RUN: llvm-mc -triple=amdgcn -mcpu=gfx900 -mattr=+wavefrontsize32 -disassemble -o - %s | FileCheck %s
+
+# Make sure there's no assertion when trying to use an unsupported
+# wave32 on a wave64-only target
+
+# CHECK: v_add_f64 v[0:1], 1.0, v[0:1]
+0x00,0x00,0x80,0xd2,0xf2,0x00,0x02,0x00
+
+# CHECK: v_cmp_eq_u32_e64 s[0:1], 1.0, s1
+0x00,0x00,0xca,0xd0,0xf2,0x02,0x00,0x00
+
+# CHECK: v_cndmask_b32_e64 v1, v2, v3, s[0:1]
+0x01,0x00,0x00,0xd1,0x02,0x07,0x02,0x00