reflect: add hooks for dealing with narrow width floats

Currently on amd64 and arm64, float32 values just live in the bottom 32
bits of the register, so reflect simply places them there in a RegArgs
for reflectcall to load them. This works fine because both of these
platforms don't care what the upper 32 bits are, and have instructions
to operate on float32 values specifically that we use. In sum, the
representation of the float32 in memory is identical to that of the
representation in a register.

On other platforms, however, the representation of FP values differ
depending on whether they're in memory or in a register. For instance,
on ppc64, all floating point values get promoted to a float64 when
loaded to a register (i.e. there's only one set of FP instructions). As
another example, on riscv64, narrow-width floats in registers need to be
NaN-boxed.

What all this means is that for supporting the register ABI on these
platforms, reflect needs to do a little extra work to ensure that the
representation of FP values in a RegArgs matches the representation it
takes on in a register. For this purpose, this change abstracts away the
action of storing values into a RegArgs a little bit and adds a
platform-specific hook which currently does nothing but copy the value.

For #40724.

Change-Id: I65dcc7d86d5602a584f86026ac204564617f4c5a
Reviewed-on: https://go-review.googlesource.com/c/go/+/347566
Trust: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>

src/internal/abi/abi.go

@@ -19,6 +19,14 @@ import (
 // when it may not be safe to keep them only in the integer
 // register space otherwise.
 type RegArgs struct {
+	// Values in these slots should be precisely the bit-by-bit
+	// representation of how they would appear in a register.
+	//
+	// This means that on big endian arches, integer values should
+	// be in the top bits of the slot. Floats are usually just
+	// directly represented, but some architectures treat narrow
+	// width floating point values specially (e.g. they're promoted
+	// first, or they need to be NaN-boxed).
 	Ints   [IntArgRegs]uintptr  // untyped integer registers
 	Floats [FloatArgRegs]uint64 // untyped float registers
 
@@ -56,26 +64,6 @@ func (r *RegArgs) IntRegArgAddr(reg int, argSize uintptr) unsafe.Pointer {
 	return unsafe.Pointer(uintptr(unsafe.Pointer(&r.Ints[reg])) + offset)
 }
 
-// FloatRegArgAddr returns a pointer inside of r.Floats[reg] that is appropriately
-// offset for an argument of size argSize.
-//
-// argSize must be non-zero, fit in a register, and a power-of-two.
-//
-// This method is a helper for dealing with the endianness of different CPU
-// architectures, since sub-word-sized arguments in big endian architectures
-// need to be "aligned" to the upper edge of the register to be interpreted
-// by the CPU correctly.
-func (r *RegArgs) FloatRegArgAddr(reg int, argSize uintptr) unsafe.Pointer {
-	if argSize > EffectiveFloatRegSize || argSize == 0 || argSize&(argSize-1) != 0 {
-		panic("invalid argSize")
-	}
-	offset := uintptr(0)
-	if goarch.BigEndian {
-		offset = EffectiveFloatRegSize - argSize
-	}
-	return unsafe.Pointer(uintptr(unsafe.Pointer(&r.Floats[reg])) + offset)
-}
-
 // IntArgRegBitmap is a bitmap large enough to hold one bit per
 // integer argument/return register.
 type IntArgRegBitmap [(IntArgRegs + 7) / 8]uint8

src/reflect/abi.go

@@ -467,3 +467,45 @@ func newAbiDesc(t *funcType, rcvr *rtype) abiDesc {
 	out.stackBytes -= retOffset
 	return abiDesc{in, out, stackCallArgsSize, retOffset, spill, stackPtrs, inRegPtrs, outRegPtrs}
 }
+
+// intFromReg loads an argSize sized integer from reg and places it at to.
+//
+// argSize must be non-zero, fit in a register, and a power-of-two.
+func intFromReg(r *abi.RegArgs, reg int, argSize uintptr, to unsafe.Pointer) {
+	memmove(to, r.IntRegArgAddr(reg, argSize), argSize)
+}
+
+// intToReg loads an argSize sized integer and stores it into reg.
+//
+// argSize must be non-zero, fit in a register, and a power-of-two.
+func intToReg(r *abi.RegArgs, reg int, argSize uintptr, from unsafe.Pointer) {
+	memmove(r.IntRegArgAddr(reg, argSize), from, argSize)
+}
+
+// floatFromReg loads a float value from its register representation in r.
+//
+// argSize must be 4 or 8.
+func floatFromReg(r *abi.RegArgs, reg int, argSize uintptr, to unsafe.Pointer) {
+	switch argSize {
+	case 4:
+		*(*float32)(to) = archFloat32FromReg(r.Floats[reg])
+	case 8:
+		*(*float64)(to) = *(*float64)(unsafe.Pointer(&r.Floats[reg]))
+	default:
+		panic("bad argSize")
+	}
+}
+
+// floatToReg stores a float value in its register representation in r.
+//
+// argSize must be either 4 or 8.
+func floatToReg(r *abi.RegArgs, reg int, argSize uintptr, from unsafe.Pointer) {
+	switch argSize {
+	case 4:
+		r.Floats[reg] = archFloat32ToReg(*(*float32)(from))
+	case 8:
+		r.Floats[reg] = *(*uint64)(from)
+	default:
+		panic("bad argSize")
+	}
+}

src/reflect/float32reg_generic.go

@@ -0,0 +1,21 @@
+// Copyright 2021 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package reflect
+
+import "unsafe"
+
+// This file implements a straightforward conversion of a float32
+// value into its representation in a register. This conversion
+// applies for amd64 and arm64. It is also chosen for the case of
+// zero argument registers, but is not used.
+
+func archFloat32FromReg(reg uint64) float32 {
+	i := uint32(reg)
+	return *(*float32)(unsafe.Pointer(&i))
+}
+
+func archFloat32ToReg(val float32) uint64 {
+	return uint64(*(*uint32)(unsafe.Pointer(&val)))
+}

src/reflect/value.go

@@ -508,7 +508,7 @@ func (v Value) call(op string, in []Value) []Value {
 				// Copy values to "integer registers."
 				if v.flag&flagIndir != 0 {
 					offset := add(v.ptr, st.offset, "precomputed value offset")
-					memmove(regArgs.IntRegArgAddr(st.ireg, st.size), offset, st.size)
+					intToReg(&regArgs, st.ireg, st.size, offset)
 				} else {
 					if st.kind == abiStepPointer {
 						// Duplicate this pointer in the pointer area of the
@@ -524,7 +524,7 @@ func (v Value) call(op string, in []Value) []Value {
 					panic("attempted to copy pointer to FP register")
 				}
 				offset := add(v.ptr, st.offset, "precomputed value offset")
-				memmove(regArgs.FloatRegArgAddr(st.freg, st.size), offset, st.size)
+				floatToReg(&regArgs, st.freg, st.size, offset)
 			default:
 				panic("unknown ABI part kind")
 			}
@@ -610,13 +610,13 @@ func (v Value) call(op string, in []Value) []Value {
 				switch st.kind {
 				case abiStepIntReg:
 					offset := add(s, st.offset, "precomputed value offset")
-					memmove(offset, regArgs.IntRegArgAddr(st.ireg, st.size), st.size)
+					intFromReg(&regArgs, st.ireg, st.size, offset)
 				case abiStepPointer:
 					s := add(s, st.offset, "precomputed value offset")
 					*((*unsafe.Pointer)(s)) = regArgs.Ptrs[st.ireg]
 				case abiStepFloatReg:
 					offset := add(s, st.offset, "precomputed value offset")
-					memmove(offset, regArgs.FloatRegArgAddr(st.freg, st.size), st.size)
+					floatFromReg(&regArgs, st.freg, st.size, offset)
 				case abiStepStack:
 					panic("register-based return value has stack component")
 				default:
@@ -698,13 +698,13 @@ func callReflect(ctxt *makeFuncImpl, frame unsafe.Pointer, retValid *bool, regs
 					switch st.kind {
 					case abiStepIntReg:
 						offset := add(v.ptr, st.offset, "precomputed value offset")
-						memmove(offset, regs.IntRegArgAddr(st.ireg, st.size), st.size)
+						intFromReg(regs, st.ireg, st.size, offset)
 					case abiStepPointer:
 						s := add(v.ptr, st.offset, "precomputed value offset")
 						*((*unsafe.Pointer)(s)) = regs.Ptrs[st.ireg]
 					case abiStepFloatReg:
 						offset := add(v.ptr, st.offset, "precomputed value offset")
-						memmove(offset, regs.FloatRegArgAddr(st.freg, st.size), st.size)
+						floatFromReg(regs, st.freg, st.size, offset)
 					case abiStepStack:
 						panic("register-based return value has stack component")
 					default:
@@ -784,7 +784,7 @@ func callReflect(ctxt *makeFuncImpl, frame unsafe.Pointer, retValid *bool, regs
 					// Copy values to "integer registers."
 					if v.flag&flagIndir != 0 {
 						offset := add(v.ptr, st.offset, "precomputed value offset")
-						memmove(regs.IntRegArgAddr(st.ireg, st.size), offset, st.size)
+						intToReg(regs, st.ireg, st.size, offset)
 					} else {
 						// Only populate the Ints space on the return path.
 						// This is safe because out is kept alive until the
@@ -799,7 +799,7 @@ func callReflect(ctxt *makeFuncImpl, frame unsafe.Pointer, retValid *bool, regs
 						panic("attempted to copy pointer to FP register")
 					}
 					offset := add(v.ptr, st.offset, "precomputed value offset")
-					memmove(regs.FloatRegArgAddr(st.freg, st.size), offset, st.size)
+					floatToReg(regs, st.freg, st.size, offset)
 				default:
 					panic("unknown ABI part kind")
 				}
@@ -982,9 +982,9 @@ func callMethod(ctxt *methodValue, frame unsafe.Pointer, retValid *bool, regs *a
 					methodRegs.Ptrs[mStep.ireg] = *(*unsafe.Pointer)(from)
 					fallthrough // We need to make sure this ends up in Ints, too.
 				case abiStepIntReg:
-					memmove(methodRegs.IntRegArgAddr(mStep.ireg, mStep.size), from, mStep.size)
+					intToReg(&methodRegs, mStep.ireg, mStep.size, from)
 				case abiStepFloatReg:
-					memmove(methodRegs.FloatRegArgAddr(mStep.freg, mStep.size), from, mStep.size)
+					floatToReg(&methodRegs, mStep.freg, mStep.size, from)
 				default:
 					panic("unexpected method step")
 				}
@@ -1000,9 +1000,9 @@ func callMethod(ctxt *methodValue, frame unsafe.Pointer, retValid *bool, regs *a
 					// Do the pointer copy directly so we get a write barrier.
 					*(*unsafe.Pointer)(to) = valueRegs.Ptrs[vStep.ireg]
 				case abiStepIntReg:
-					memmove(to, valueRegs.IntRegArgAddr(vStep.ireg, vStep.size), vStep.size)
+					intFromReg(valueRegs, vStep.ireg, vStep.size, to)
 				case abiStepFloatReg:
-					memmove(to, valueRegs.FloatRegArgAddr(vStep.freg, vStep.size), vStep.size)
+					floatFromReg(valueRegs, vStep.freg, vStep.size, to)
 				default:
 					panic("unexpected value step")
 				}