Try to make f16 legal instead

llvm · Oct 31, 2024 · 44dffa4 · 44dffa4
1 parent 9c0bc36
commit 44dffa4
Show file tree

Hide file tree

Showing 23 changed files with 906 additions and 246 deletions.
diff --git a/clang/lib/Basic/Targets/SystemZ.h b/clang/lib/Basic/Targets/SystemZ.h
@@ -93,6 +93,11 @@ class LLVM_LIBRARY_VISIBILITY SystemZTargetInfo : public TargetInfo {
     MaxAtomicPromoteWidth = MaxAtomicInlineWidth = 128;
 
     // True if the backend supports operations on the half LLVM IR type.
+    // By setting this to false, conversions will happen for _Float16 around
+    // a statement by default with operations done in float. However, if
+    // -ffloat16-excess-precision=none is given, no conversions will be made
+    // and instead the backend will promote each half operation to float
+    // individually.
     HasLegalHalfType = false;
     // Allow half arguments and return values.
     HalfArgsAndReturns = true;

diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
@@ -16548,7 +16548,7 @@ ExprResult Sema::BuildVAArgExpr(SourceLocation BuiltinLoc,
           PromoteType = QualType();
       }
     }
-    if (TInfo->getType()->isFloat16Type() || TInfo->getType()->isFloat32Type())
+    if (TInfo->getType()->isSpecificBuiltinType(BuiltinType::Float))
       PromoteType = Context.DoubleTy;
     if (!PromoteType.isNull())
       DiagRuntimeBehavior(TInfo->getTypeLoc().getBeginLoc(), E,

diff --git a/compiler-rt/test/builtins/CMakeLists.txt b/compiler-rt/test/builtins/CMakeLists.txt
@@ -56,7 +56,7 @@ foreach(arch ${BUILTIN_TEST_ARCH})
       string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")
     endif()
   else()
-    if (${arch} MATCHES "arm|armhf|aarch64|arm64|i?86|x86_64|AMD64|riscv32|riscv64" AND COMPILER_RT_HAS_${arch}_FLOAT16)
+    if (${arch} MATCHES "arm|armhf|aarch64|arm64|i?86|x86_64|AMD64|riscv32|riscv64|s390x" AND COMPILER_RT_HAS_${arch}_FLOAT16)
       list(APPEND BUILTINS_TEST_TARGET_CFLAGS -DCOMPILER_RT_HAS_FLOAT16)
       string(REPLACE ";" " " BUILTINS_TEST_TARGET_CFLAGS "${BUILTINS_TEST_TARGET_CFLAGS}")
     endif()

diff --git a/llvm/lib/IR/RuntimeLibcalls.cpp b/llvm/lib/IR/RuntimeLibcalls.cpp
@@ -327,6 +327,8 @@ void RuntimeLibcallsInfo::initLibcalls(const Triple &TT) {
     setLibcallName(RTLIB::SRL_I128, nullptr);
     setLibcallName(RTLIB::SHL_I128, nullptr);
     setLibcallName(RTLIB::SRA_I128, nullptr);
+    setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
+    setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
   }
 
   if (TT.isX86()) {

diff --git a/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp b/llvm/lib/Target/SystemZ/AsmParser/SystemZAsmParser.cpp
@@ -60,6 +60,7 @@ enum RegisterKind {
   GRH32Reg,
   GR64Reg,
   GR128Reg,
+  FP16Reg,
   FP32Reg,
   FP64Reg,
   FP128Reg,
@@ -356,6 +357,7 @@ class SystemZOperand : public MCParsedAsmOperand {
   bool isADDR32() const { return isReg(GR32Reg); }
   bool isADDR64() const { return isReg(GR64Reg); }
   bool isADDR128() const { return false; }
+  bool isFP16() const { return isReg(FP16Reg); }
   bool isFP32() const { return isReg(FP32Reg); }
   bool isFP64() const { return isReg(FP64Reg); }
   bool isFP128() const { return isReg(FP128Reg); }
@@ -534,6 +536,9 @@ class SystemZAsmParser : public MCTargetAsmParser {
   ParseStatus parseADDR128(OperandVector &Operands) {
     llvm_unreachable("Shouldn't be used as an operand");
   }
+  ParseStatus parseFP16(OperandVector &Operands) {
+    return parseRegister(Operands, FP16Reg);
+  }
   ParseStatus parseFP32(OperandVector &Operands) {
     return parseRegister(Operands, FP32Reg);
   }
@@ -829,6 +834,7 @@ ParseStatus SystemZAsmParser::parseRegister(OperandVector &Operands,
   case GR128Reg:
     Group = RegGR;
     break;
+  case FP16Reg:
   case FP32Reg:
   case FP64Reg:
   case FP128Reg:
@@ -882,6 +888,7 @@ ParseStatus SystemZAsmParser::parseRegister(OperandVector &Operands,
   case GRH32Reg: Regs = SystemZMC::GRH32Regs; break;
   case GR64Reg:  Regs = SystemZMC::GR64Regs;  break;
   case GR128Reg: Regs = SystemZMC::GR128Regs; break;
+  case FP16Reg:  Regs = SystemZMC::FP16Regs;  break;
   case FP32Reg:  Regs = SystemZMC::FP32Regs;  break;
   case FP64Reg:  Regs = SystemZMC::FP64Regs;  break;
   case FP128Reg: Regs = SystemZMC::FP128Regs; break;

diff --git a/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZMCTargetDesc.cpp b/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZMCTargetDesc.cpp
@@ -61,6 +61,13 @@ const unsigned SystemZMC::GR128Regs[16] = {
   SystemZ::R12Q, 0, SystemZ::R14Q, 0
 };
 
+const unsigned SystemZMC::FP16Regs[16] = {
+  SystemZ::F0H, SystemZ::F1H, SystemZ::F2H, SystemZ::F3H,
+  SystemZ::F4H, SystemZ::F5H, SystemZ::F6H, SystemZ::F7H,
+  SystemZ::F8H, SystemZ::F9H, SystemZ::F10H, SystemZ::F11H,
+  SystemZ::F12H, SystemZ::F13H, SystemZ::F14H, SystemZ::F15H
+};
+
 const unsigned SystemZMC::FP32Regs[16] = {
   SystemZ::F0S, SystemZ::F1S, SystemZ::F2S, SystemZ::F3S,
   SystemZ::F4S, SystemZ::F5S, SystemZ::F6S, SystemZ::F7S,

diff --git a/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZMCTargetDesc.h b/llvm/lib/Target/SystemZ/MCTargetDesc/SystemZMCTargetDesc.h
@@ -43,6 +43,7 @@ extern const unsigned GR32Regs[16];
 extern const unsigned GRH32Regs[16];
 extern const unsigned GR64Regs[16];
 extern const unsigned GR128Regs[16];
+extern const unsigned FP16Regs[16];
 extern const unsigned FP32Regs[16];
 extern const unsigned FP64Regs[16];
 extern const unsigned FP128Regs[16];

diff --git a/llvm/lib/Target/SystemZ/SystemZCallingConv.td b/llvm/lib/Target/SystemZ/SystemZCallingConv.td
@@ -50,7 +50,8 @@ def RetCC_SystemZ_ELF : CallingConv<[
   // other floating-point argument registers available for code that
   // doesn't care about the ABI.  All floating-point argument registers
   // are call-clobbered, so we can use all of them here.
-  CCIfType<[f16, f32], CCAssignToReg<[F0S, F2S, F4S, F6S]>>,
+  CCIfType<[f16], CCAssignToReg<[F0H, F2H, F4H, F6H]>>,
+  CCIfType<[f32], CCAssignToReg<[F0S, F2S, F4S, F6S]>>,
   CCIfType<[f64], CCAssignToReg<[F0D, F2D, F4D, F6D]>>,
 
   // Similarly for vectors, with V24 being the ABI-compliant choice.
@@ -115,7 +116,8 @@ def CC_SystemZ_ELF : CallingConv<[
   CCIfType<[i64], CCAssignToReg<[R2D, R3D, R4D, R5D, R6D]>>,
 
   // The first 4 float and double arguments are passed in even registers F0-F6.
-  CCIfType<[f16, f32], CCAssignToReg<[F0S, F2S, F4S, F6S]>>,
+  CCIfType<[f16], CCAssignToReg<[F0H, F2H, F4H, F6H]>>,
+  CCIfType<[f32], CCAssignToReg<[F0S, F2S, F4S, F6S]>>,
   CCIfType<[f64], CCAssignToReg<[F0D, F2D, F4D, F6D]>>,
 
   // The first 8 named vector arguments are passed in V24-V31.  Sub-128 vectors