Merge pull request #54 from zhouguangyuan0718/codex/fp-constant-bits

ir: preserve floating-point constant bit patterns

Author: xushiwei

IRBindings.cpp

@@ -14,16 +14,50 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Config/llvm-config.h"
 #include "llvm/IR/Attributes.h"
+#include "llvm/IR/Constants.h"
 #include "llvm/IR/DebugLoc.h"
 #include "llvm/IR/DebugInfoMetadata.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Module.h"
+#include <algorithm>
 
 using namespace llvm;
 
+LLVMValueRef LLVMGoConstFPFromBits(LLVMTypeRef Ty, const uint64_t *Words,
+                                 unsigned NumWords) {
+  if (!Ty || !unwrap(Ty)->isFloatingPointTy())
+    return nullptr;
+  auto *T = unwrap(Ty);
+  unsigned BitWidth = T->getScalarSizeInBits();
+  if (NumWords != (BitWidth + 63) / 64 || !Words)
+    return nullptr;
+  if (BitWidth % 64 && (Words[NumWords - 1] >> (BitWidth % 64)))
+    return nullptr;
+#if LLVM_VERSION_MAJOR >= 22
+  return LLVMConstFPFromBits(Ty, Words);
+#else
+  return wrap(ConstantFP::get(
+      T->getContext(), APFloat(T->getFltSemantics(),
+                              APInt(BitWidth, ArrayRef<uint64_t>(Words, NumWords)))));
+#endif
+}
+
+unsigned LLVMGoConstFPGetBits(LLVMValueRef Val, uint64_t *Words) {
+  if (!Val)
+    return 0;
+  auto *FP = dyn_cast<ConstantFP>(unwrap(Val));
+  if (!FP)
+    return 0;
+  APInt Bits = FP->getValueAPF().bitcastToAPInt();
+  unsigned NumWords = Bits.getNumWords();
+  if (Words)
+    std::copy_n(Bits.getRawData(), NumWords, Words);
+  return NumWords;
+}
+
 LLVMAttributeRef LLVMGoCreateConstantRangeAttribute(
     LLVMContextRef C, unsigned KindID, unsigned NumBits,
     const uint64_t *LowerWords, const uint64_t *UpperWords) {

IRBindings.h

@@ -33,6 +33,10 @@ struct LLVMDebugLocMetadata{
     LLVMMetadataRef InlinedAt;
 };
 
+LLVMValueRef LLVMGoConstFPFromBits(LLVMTypeRef Ty, const uint64_t *Words,
+                                 unsigned NumWords);
+unsigned LLVMGoConstFPGetBits(LLVMValueRef Val, uint64_t *Words);
+
 LLVMMetadataRef LLVMConstantAsMetadata(LLVMValueRef Val);
 
 LLVMAttributeRef LLVMGoCreateConstantRangeAttribute(

float_bits_test.go

@@ -0,0 +1,115 @@
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+package llvm
+
+import (
+	"os"
+	"path/filepath"
+	"reflect"
+	"testing"
+)
+
+func TestFloatBits(t *testing.T) {
+	// Parse independently specified IR, then rebuild in another context. Values
+	// include precision below float64's significand and NaN payload/signaling bits.
+	cases := []struct {
+		name, ir string
+		words    []uint64
+	}{
+		{"half", "half 0xH8000", []uint64{0x8000}},
+		{"bfloat", "bfloat 0xR7FC1", []uint64{0x7fc1}},
+		{"float_nan", "float 0x7FF82468A0000000", []uint64{0x7fc12345}},
+		{"double_negative_zero", "double 0x8000000000000000", []uint64{0x8000000000000000}},
+		{"double_snan", "double 0x7FF0000000000001", []uint64{0x7ff0000000000001}},
+		{"double_infinity", "double 0x7FF0000000000000", []uint64{0x7ff0000000000000}},
+		{"double_subnormal", "double 0x0000000000000001", []uint64{1}},
+		{"fp80_precision", "x86_fp80 0xK3FFF8000000000000001", []uint64{0x8000000000000001, 0x3fff}},
+		{"fp80_nan", "x86_fp80 0xK7FFFC000000000000123", []uint64{0xc000000000000123, 0x7fff}},
+		{"fp128_precision", "fp128 0xL00000000000000013FFF000000000000", []uint64{1, 0x3fff000000000000}},
+		{"fp128_nan", "fp128 0xL00000000000001237FFF800000000000", []uint64{0x123, 0x7fff800000000000}},
+		{"ppc_fp128_low_only", "ppc_fp128 0xM00000000000000003FF0000000000000", []uint64{0, 0x3ff0000000000000}},
+		{"ppc_fp128_negative_zero", "ppc_fp128 0xM80000000000000000000000000000000", []uint64{0x8000000000000000, 0}},
+		{"ppc_fp128_noncanonical", "ppc_fp128 0xM3FF00000000000003FF0000000000000", []uint64{0x3ff0000000000000, 0x3ff0000000000000}},
+		{"ppc_fp128_nan", "ppc_fp128 0xM7FF80000000001230000000000000000", []uint64{0x7ff8000000000123, 0}},
+		{"ppc_fp128_precision", "ppc_fp128 0xM3FF00000000000003C90000000000000", []uint64{0x3ff0000000000000, 0x3c90000000000000}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			srcCtx, dstCtx := NewContext(), NewContext()
+			defer dstCtx.Dispose()
+			path := filepath.Join(t.TempDir(), "float.ll")
+			if err := os.WriteFile(path, []byte("@value = constant "+tc.ir+"\n"), 0600); err != nil {
+				t.Fatal(err)
+			}
+			buf, err := NewMemoryBufferFromFile(path)
+			if err != nil {
+				t.Fatal(err)
+			}
+			src, err := srcCtx.ParseIR(buf)
+			if err != nil {
+				t.Fatal(err)
+			}
+			value := src.NamedGlobal("value").Initializer()
+			words := value.FloatBits()
+			if !reflect.DeepEqual(words, tc.words) {
+				t.Fatalf("bits = %x, want %x", words, tc.words)
+			}
+			// Reparse just to obtain the identical destination-context type.
+			buf, err = NewMemoryBufferFromFile(path)
+			if err != nil {
+				t.Fatal(err)
+			}
+			dst, err := dstCtx.ParseIR(buf)
+			if err != nil {
+				t.Fatal(err)
+			}
+			defer dst.Dispose()
+			expected := dst.NamedGlobal("value").Initializer()
+			cloned := ConstFloatFromBits(expected.Type(), words)
+			if cloned != expected {
+				t.Fatalf("rebuilt %s, want %s", cloned, expected)
+			}
+			src.Dispose()
+			srcCtx.Dispose()
+			if got := cloned.FloatBits(); !reflect.DeepEqual(got, tc.words) {
+				t.Fatalf("after source disposal: %x", got)
+			}
+			words[0] ^= 1
+			if got := cloned.FloatBits(); !reflect.DeepEqual(got, tc.words) {
+				t.Fatalf("returned slice aliases constant: %x", got)
+			}
+			if err := VerifyModule(dst, ReturnStatusAction); err != nil {
+				t.Fatal(err)
+			}
+		})
+	}
+}
+
+func TestFloatBitsInvalidInput(t *testing.T) {
+	ctx := NewContext()
+	defer ctx.Dispose()
+	for name, call := range map[string]func(){
+		"nil_type":              func() { ConstFloatFromBits(Type{}, []uint64{0}) },
+		"integer_type":          func() { ConstFloatFromBits(ctx.Int64Type(), []uint64{0}) },
+		"vector_type":           func() { ConstFloatFromBits(VectorType(ctx.DoubleType(), 2), []uint64{0, 0}) },
+		"empty":                 func() { ConstFloatFromBits(ctx.DoubleType(), nil) },
+		"too_few":               func() { ConstFloatFromBits(ctx.FP128Type(), []uint64{0}) },
+		"too_many":              func() { ConstFloatFromBits(ctx.DoubleType(), []uint64{0, 0}) },
+		"unused_high_bits":      func() { ConstFloatFromBits(ctx.FloatType(), []uint64{1 << 32}) },
+		"fp80_unused_high_bits": func() { ConstFloatFromBits(ctx.X86FP80Type(), []uint64{0, 1 << 16}) },
+		"nil_value":             func() { Value{}.FloatBits() },
+		"integer_value":         func() { ConstInt(ctx.Int64Type(), 0, false).FloatBits() },
+		"undef":                 func() { Undef(ctx.DoubleType()).FloatBits() },
+	} {
+		t.Run(name, func(t *testing.T) {
+			defer func() {
+				if recover() == nil {
+					t.Fatal("expected panic")
+				}
+			}()
+			call()
+		})
+	}
+}

ir.go

@@ -927,6 +927,37 @@ func ConstFloatFromString(t Type, str string) (v Value) {
 	return
 }
 
+// ConstFloatFromBits constructs a scalar floating-point constant without rounding
+// through float64. words holds the raw representation, least significant word
+// first, independently of host byte order. Its length must be ceil(bitWidth/64),
+// and unused high bits in the final word must be zero. It panics for an invalid
+// type or representation. For ppc_fp128, the first word is the leading double
+// and the second word is the trailing double, as in LLVM's APFloat representation.
+func ConstFloatFromBits(t Type, words []uint64) (v Value) {
+	var data *C.uint64_t
+	if len(words) != 0 {
+		data = (*C.uint64_t)(unsafe.Pointer(&words[0]))
+	}
+	v.C = C.LLVMGoConstFPFromBits(t.C, data, C.uint(len(words)))
+	if v.IsNil() {
+		panic("llvm: invalid floating-point type or bit representation")
+	}
+	return
+}
+
+// FloatBits returns a copy of a scalar floating-point constant's raw APFloat
+// representation in least-significant-word-first order. Unused high bits in the
+// final word are zero. It panics if v is not a scalar floating-point constant.
+func (v Value) FloatBits() []uint64 {
+	n := C.LLVMGoConstFPGetBits(v.C, nil)
+	if n == 0 {
+		panic("llvm: FloatBits requires a scalar floating-point constant")
+	}
+	words := make([]uint64, int(n))
+	C.LLVMGoConstFPGetBits(v.C, (*C.uint64_t)(unsafe.Pointer(&words[0])))
+	return words
+}
+
 func (v Value) ZExtValue() uint64 { return uint64(C.LLVMConstIntGetZExtValue(v.C)) }
 func (v Value) SExtValue() int64  { return int64(C.LLVMConstIntGetSExtValue(v.C)) }
 func (v Value) DoubleValue() (result float64, inexact bool) {