Add support for nneg flag with uitofp

As noted when #82404 was pushed (canonicalizing `sitofp` -> `uitofp`),
different signedness on fp casts can have dramatic performance
implications on different backends.

So, it makes to create a reliable means for the backend to pick its
cast signedness if either are correct.

Further, this allows us to start canonicalizing `sitofp`- > `uitofp`
which may easy middle end analysis.

Closes #86141

Author: goldsteinn

llvm/docs/LangRef.rst

@@ -11751,6 +11751,10 @@ Overview:
 The '``uitofp``' instruction regards ``value`` as an unsigned integer
 and converts that value to the ``ty2`` type.
 
+The ``nneg`` (non-negative) flag, if present, specifies that the
+operand is non-negative. This property may be used by optimization
+passes to later convert the ``uitofp`` into a ``sitofp``.
+
 Arguments:
 """"""""""
 
@@ -11768,6 +11772,9 @@ integer quantity and converts it to the corresponding floating-point
 value. If the value cannot be exactly represented, it is rounded using
 the default rounding mode.
 
+If the ``nneg`` flag is set, and the ``uitofp`` argument is negative,
+the result is a poison value.
+
 
 Example:
 """"""""
@@ -11777,6 +11784,9 @@ Example:
       %X = uitofp i32 257 to float         ; yields float:257.0
       %Y = uitofp i8 -1 to double          ; yields double:255.0
 
+      %a = uitofp nneg i32 256 to i32      ; yields float:256.0
+      %b = uitofp nneg i32 -256 to i32     ; yields i32 poison
+
 '``sitofp .. to``' Instruction
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

llvm/include/llvm/IR/IRBuilder.h

@@ -2068,11 +2068,17 @@ class IRBuilderBase {
     return CreateCast(Instruction::FPToSI, V, DestTy, Name);
   }
 
-  Value *CreateUIToFP(Value *V, Type *DestTy, const Twine &Name = ""){
+  Value *CreateUIToFP(Value *V, Type *DestTy, const Twine &Name = "",
+                      bool IsNonNeg = false) {
     if (IsFPConstrained)
       return CreateConstrainedFPCast(Intrinsic::experimental_constrained_uitofp,
                                      V, DestTy, nullptr, Name);
-    return CreateCast(Instruction::UIToFP, V, DestTy, Name);
+    if (Value *Folded = Folder.FoldCast(Instruction::UIToFP, V, DestTy))
+      return Folded;
+    Instruction *I = Insert(new UIToFPInst(V, DestTy), Name);
+    if (IsNonNeg)
+      I->setNonNeg();
+    return I;
   }
 
   Value *CreateSIToFP(Value *V, Type *DestTy, const Twine &Name = ""){

llvm/include/llvm/IR/InstrTypes.h

@@ -927,13 +927,19 @@ class CastInst : public UnaryInstruction {
   }
 };
 
-/// Instruction that can have a nneg flag (only zext).
+/// Instruction that can have a nneg flag (zext/uitofp).
 class PossiblyNonNegInst : public CastInst {
 public:
   enum { NonNeg = (1 << 0) };
 
   static bool classof(const Instruction *I) {
-    return I->getOpcode() == Instruction::ZExt;
+    switch (I->getOpcode()) {
+    case Instruction::ZExt:
+    case Instruction::UIToFP:
+      return true;
+    default:
+      return false;
+    }
   }
 
   static bool classof(const Value *V) {

llvm/lib/AsmParser/LLParser.cpp

@@ -6816,6 +6816,7 @@ int LLParser::parseInstruction(Instruction *&Inst, BasicBlock *BB,
   }
 
   // Casts.
+  case lltok::kw_uitofp:
   case lltok::kw_zext: {
     bool NonNeg = EatIfPresent(lltok::kw_nneg);
     bool Res = parseCast(Inst, PFS, KeywordVal);
@@ -6843,7 +6844,6 @@ int LLParser::parseInstruction(Instruction *&Inst, BasicBlock *BB,
   case lltok::kw_fpext:
   case lltok::kw_bitcast:
   case lltok::kw_addrspacecast:
-  case lltok::kw_uitofp:
   case lltok::kw_sitofp:
   case lltok::kw_fptoui:
   case lltok::kw_fptosi:

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -5039,7 +5039,7 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
       }
 
       if (OpNum < Record.size()) {
-        if (Opc == Instruction::ZExt) {
+        if (Opc == Instruction::ZExt || Opc == Instruction::UIToFP) {
           if (Record[OpNum] & (1 << bitc::PNNI_NON_NEG))
             cast<PossiblyNonNegInst>(I)->setNonNeg(true);
         } else if (Opc == Instruction::Trunc) {

llvm/lib/IR/Instruction.cpp

@@ -388,7 +388,7 @@ void Instruction::setIsExact(bool b) {
 }
 
 void Instruction::setNonNeg(bool b) {
-  assert(isa<PossiblyNonNegInst>(this) && "Must be zext");
+  assert(isa<PossiblyNonNegInst>(this) && "Must be zext/uitofp");
   SubclassOptionalData = (SubclassOptionalData & ~PossiblyNonNegInst::NonNeg) |
                          (b * PossiblyNonNegInst::NonNeg);
 }
@@ -408,7 +408,7 @@ bool Instruction::hasNoSignedWrap() const {
 }
 
 bool Instruction::hasNonNeg() const {
-  assert(isa<PossiblyNonNegInst>(this) && "Must be zext");
+  assert(isa<PossiblyNonNegInst>(this) && "Must be zext/uitofp");
   return (SubclassOptionalData & PossiblyNonNegInst::NonNeg) != 0;
 }
 
@@ -441,6 +441,7 @@ void Instruction::dropPoisonGeneratingFlags() {
     cast<GetElementPtrInst>(this)->setIsInBounds(false);
     break;
 
+  case Instruction::UIToFP:
   case Instruction::ZExt:
     setNonNeg(false);
     break;

llvm/lib/IR/Operator.cpp

@@ -44,6 +44,7 @@ bool Operator::hasPoisonGeneratingFlags() const {
     // Note: inrange exists on constexpr only
     return GEP->isInBounds() || GEP->getInRange() != std::nullopt;
   }
+  case Instruction::UIToFP:
   case Instruction::ZExt:
     if (auto *NNI = dyn_cast<PossiblyNonNegInst>(this))
       return NNI->hasNonNeg();

llvm/test/Assembler/flags.ll

@@ -256,6 +256,13 @@ define i64 @test_zext(i32 %a) {
   ret i64 %res
 }
 
+define float @test_uitofp(i32 %a) {
+; CHECK: %res = uitofp nneg i32 %a to float
+  %res = uitofp nneg i32 %a to float
+  ret float %res
+}
+
+
 define i64 @test_or(i64 %a, i64 %b) {
 ; CHECK: %res = or disjoint i64 %a, %b
   %res = or disjoint i64 %a, %b

llvm/test/Bitcode/flags.ll

@@ -18,6 +18,8 @@ second:                                           ; preds = %first
   %z = add i32 %a, 0                              ; <i32> [#uses=0]
   %hh = zext nneg i32 %a to i64
   %ll = zext i32 %s to i64
+  %ff = uitofp nneg i32 %a to float
+  %bb = uitofp i32 %s to float
   %jj = or disjoint i32 %a, 0
   %oo = or i32 %a, 0
   %tu = trunc nuw i32 %a to i16
@@ -39,6 +41,8 @@ first:                                                    ; preds = %entry
   %zz = add i32 %a, 0                                     ; <i32> [#uses=0]
   %kk = zext nneg i32 %a to i64
   %rr = zext i32 %ss to i64
+  %ww = uitofp nneg i32 %a to float
+  %xx = uitofp i32 %ss to float
   %mm = or disjoint i32 %a, 0
   %nn = or i32 %a, 0
   %tuu = trunc nuw i32 %a to i16

llvm/test/Transforms/InstCombine/freeze.ll

@@ -1127,6 +1127,17 @@ define i32 @freeze_zext_nneg(i8 %x) {
   ret i32 %fr
 }
 
+define float @freeze_uitofp_nneg(i8 %x) {
+; CHECK-LABEL: @freeze_uitofp_nneg(
+; CHECK-NEXT:    [[X_FR:%.*]] = freeze i8 [[X:%.*]]
+; CHECK-NEXT:    [[UITOFP:%.*]] = uitofp i8 [[X_FR]] to float
+; CHECK-NEXT:    ret float [[UITOFP]]
+;
+  %uitofp = uitofp nneg i8 %x to float
+  %fr = freeze float %uitofp
+  ret float %fr
+}
+
 define i32 @propagate_drop_flags_or(i32 %arg) {
 ; CHECK-LABEL: @propagate_drop_flags_or(
 ; CHECK-NEXT:    [[ARG_FR:%.*]] = freeze i32 [[ARG:%.*]]

llvm/test/Transforms/SimplifyCFG/HoistCode.ll

@@ -125,6 +125,37 @@ F:
   ret i32 %z2
 }
 
+
+define float @hoist_uitofp_flags_preserve(i1 %C, i8 %x) {
+; CHECK-LABEL: @hoist_uitofp_flags_preserve(
+; CHECK-NEXT:  common.ret:
+; CHECK-NEXT:    [[Z1:%.*]] = uitofp nneg i8 [[X:%.*]] to float
+; CHECK-NEXT:    ret float [[Z1]]
+;
+  br i1 %C, label %T, label %F
+T:
+  %z1 = uitofp nneg i8 %x to float
+  ret float %z1
+F:
+  %z2 = uitofp nneg i8 %x to float
+  ret float %z2
+}
+
+define float @hoist_uitofp_flags_drop(i1 %C, i8 %x) {
+; CHECK-LABEL: @hoist_uitofp_flags_drop(
+; CHECK-NEXT:  common.ret:
+; CHECK-NEXT:    [[Z1:%.*]] = uitofp i8 [[X:%.*]] to float
+; CHECK-NEXT:    ret float [[Z1]]
+;
+  br i1 %C, label %T, label %F
+T:
+  %z1 = uitofp nneg i8 %x to float
+  ret float %z1
+F:
+  %z2 = uitofp i8 %x to float
+  ret float %z2
+}
+
 define i32 @hoist_or_flags_preserve(i1 %C, i32 %x, i32 %y) {
 ; CHECK-LABEL: @hoist_or_flags_preserve(
 ; CHECK-NEXT:  common.ret: