compiler: replace some math operation bodies with fast intrinsics

Instead of changing the calls, replace the function bodies themselves.
This is useful for a number of reasons, see
https://github.com/tinygo-org/tinygo/pull/2920 for more information.

I have removed the math intrinsics tests because they are no longer
useful. Instead, I think `tinygo test math` should suffice.

compiler/compiler.go

@@ -790,6 +790,10 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 			}
 			// Create the function definition.
 			b := newBuilder(c, irbuilder, member)
+			if _, ok := mathToLLVMMapping[member.RelString(nil)]; ok {
+				b.defineMathOp()
+				continue
+			}
 			if member.Blocks == nil {
 				// Try to define this as an intrinsic function.
 				b.defineIntrinsicFunction()
@@ -1024,7 +1028,7 @@ func (c *compilerContext) getEmbedFileString(file *loader.EmbedFile) llvm.Value
 // parameters, create basic blocks, and set up debug information.
 // This is separated out from createFunction() so that it is also usable to
 // define compiler intrinsics like the atomic operations in sync/atomic.
-func (b *builder) createFunctionStart() {
+func (b *builder) createFunctionStart(intrinsic bool) {
 	if b.DumpSSA {
 		fmt.Printf("\nfunc %s:\n", b.fn)
 	}
@@ -1097,20 +1101,20 @@ func (b *builder) createFunctionStart() {
 	}
 
 	// Pre-create all basic blocks in the function.
-	for _, block := range b.fn.DomPreorder() {
-		llvmBlock := b.ctx.AddBasicBlock(b.llvmFn, block.Comment)
-		b.blockEntries[block] = llvmBlock
-		b.blockExits[block] = llvmBlock
-	}
 	var entryBlock llvm.BasicBlock
-	if len(b.fn.Blocks) != 0 {
+	if intrinsic {
-		// Normal functions have an entry block.
-		entryBlock = b.blockEntries[b.fn.Blocks[0]]
-	} else {
 		// This function isn't defined in Go SSA. It is probably a compiler
 		// intrinsic (like an atomic operation). Create the entry block
 		// manually.
 		entryBlock = b.ctx.AddBasicBlock(b.llvmFn, "entry")
+	} else {
+		for _, block := range b.fn.DomPreorder() {
+			llvmBlock := b.ctx.AddBasicBlock(b.llvmFn, block.Comment)
+			b.blockEntries[block] = llvmBlock
+			b.blockExits[block] = llvmBlock
+		}
+		// Normal functions have an entry block.
+		entryBlock = b.blockEntries[b.fn.Blocks[0]]
 	}
 	b.SetInsertPointAtEnd(entryBlock)
 
@@ -1192,7 +1196,7 @@ func (b *builder) createFunctionStart() {
 // function must not yet be defined, otherwise this function will create a
 // diagnostic.
 func (b *builder) createFunction() {
-	b.createFunctionStart()
+	b.createFunctionStart(false)
 
 	// Fill blocks with instructions.
 	for _, block := range b.fn.DomPreorder() {
@@ -1654,8 +1658,6 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 		// applied) function call. If it is anonymous, it may be a closure.
 		name := fn.RelString(nil)
 		switch {
-		case name == "math.Ceil" || name == "math.Floor" || name == "math.Sqrt" || name == "math.Trunc":
-			return b.createMathOp(instr), nil
 		case name == "device.Asm" || name == "device/arm.Asm" || name == "device/arm64.Asm" || name == "device/avr.Asm" || name == "device/riscv.Asm":
 			return b.createInlineAsm(instr.Args)
 		case name == "device.AsmFull" || name == "device/arm.AsmFull" || name == "device/arm64.AsmFull" || name == "device/avr.AsmFull" || name == "device/riscv.AsmFull":

compiler/compiler_test.go

@@ -27,12 +27,6 @@ type testCase struct {
 func TestCompiler(t *testing.T) {
 	t.Parallel()
 
-	// Determine LLVM version.
-	llvmMajor, err := strconv.Atoi(strings.SplitN(llvm.Version, ".", 2)[0])
-	if err != nil {
-		t.Fatal("could not parse LLVM version:", llvm.Version)
-	}
-
 	// Determine which tests to run, depending on the Go and LLVM versions.
 	tests := []testCase{
 		{"basic.go", "", ""},
@@ -47,14 +41,8 @@ func TestCompiler(t *testing.T) {
 		{"goroutine.go", "wasm", "asyncify"},
 		{"goroutine.go", "cortex-m-qemu", "tasks"},
 		{"channel.go", "", ""},
-		{"intrinsics.go", "cortex-m-qemu", ""},
-		{"intrinsics.go", "wasm", ""},
 		{"gc.go", "", ""},
 	}
-	if llvmMajor >= 12 {
-		tests = append(tests, testCase{"intrinsics.go", "cortex-m-qemu", ""})
-		tests = append(tests, testCase{"intrinsics.go", "wasm", ""})
-	}
 
 	for _, tc := range tests {
 		name := tc.file

compiler/intrinsics.go

@@ -7,7 +7,6 @@ import (
 	"strconv"
 	"strings"
 
-	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )
 
@@ -29,7 +28,7 @@ func (b *builder) defineIntrinsicFunction() {
 	case strings.HasPrefix(name, "runtime/volatile.Store"):
 		b.createVolatileStore()
 	case strings.HasPrefix(name, "sync/atomic.") && token.IsExported(b.fn.Name()):
-		b.createFunctionStart()
+		b.createFunctionStart(true)
 		returnValue := b.createAtomicOp(b.fn.Name())
 		if !returnValue.IsNil() {
 			b.CreateRet(returnValue)
@@ -44,7 +43,7 @@ func (b *builder) defineIntrinsicFunction() {
 // specially by optimization passes possibly resulting in better generated code,
 // and will otherwise be lowered to regular libc memcpy/memmove calls.
 func (b *builder) createMemoryCopyImpl() {
-	b.createFunctionStart()
+	b.createFunctionStart(true)
 	fnName := "llvm." + b.fn.Name() + ".p0i8.p0i8.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
 	llvmFn := b.mod.NamedFunction(fnName)
 	if llvmFn.IsNil() {
@@ -64,7 +63,7 @@ func (b *builder) createMemoryCopyImpl() {
 // memory, declaring the function if needed. These calls will be lowered to
 // regular libc memset calls if they aren't optimized out in a different way.
 func (b *builder) createMemoryZeroImpl() {
-	b.createFunctionStart()
+	b.createFunctionStart(true)
 	fnName := "llvm.memset.p0i8.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
 	llvmFn := b.mod.NamedFunction(fnName)
 	if llvmFn.IsNil() {
@@ -88,10 +87,19 @@ var mathToLLVMMapping = map[string]string{
 	"math.Trunc": "llvm.trunc.f64",
 }
 
-// createMathOp lowers the given call as a LLVM math intrinsic. It returns the
+// defineMathOp defines a math function body as a call to a LLVM intrinsic,
-// resulting value.
+// instead of the regular Go implementation. This allows LLVM to reason about
-func (b *builder) createMathOp(call *ssa.CallCommon) llvm.Value {
+// the math operation and (depending on the architecture) allows it to lower the
-	llvmName := mathToLLVMMapping[call.StaticCallee().RelString(nil)]
+// operation to very fast floating point instructions. If this is not possible,
+// LLVM will emit a call to a libm function that implements the same operation.
+//
+// One example of an optimization that LLVM can do is to convert
+// float32(math.Sqrt(float64(v))) to a 32-bit floating point operation, which is
+// beneficial on architectures where 64-bit floating point operations are (much)
+// more expensive than 32-bit ones.
+func (b *builder) defineMathOp() {
+	b.createFunctionStart(true)
+	llvmName := mathToLLVMMapping[b.fn.RelString(nil)]
 	if llvmName == "" {
 		panic("unreachable: unknown math operation") // sanity check
 	}
@@ -104,9 +112,10 @@ func (b *builder) createMathOp(call *ssa.CallCommon) llvm.Value {
 		llvmFn = llvm.AddFunction(b.mod, llvmName, llvmType)
 	}
 	// Create a call to the intrinsic.
-	args := make([]llvm.Value, len(call.Args))
+	args := make([]llvm.Value, len(b.fn.Params))
-	for i, arg := range call.Args {
+	for i, param := range b.fn.Params {
-		args[i] = b.getValue(arg)
+		args[i] = b.getValue(param)
 	}
-	return b.CreateCall(llvmFn, args, "")
+	result := b.CreateCall(llvmFn, args, "")
+	b.CreateRet(result)
 }

compiler/testdata/intrinsics-cortex-m-qemu.ll

@@ -1,36 +0,0 @@
-; ModuleID = 'intrinsics.go'
-source_filename = "intrinsics.go"
-target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
-target triple = "thumbv7m-unknown-unknown-eabi"
-
-declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*) #0
-
-; Function Attrs: nounwind
-define hidden void @main.init(i8* %context) unnamed_addr #1 {
-entry:
-  ret void
-}
-
-; Function Attrs: nounwind
-define hidden double @main.mySqrt(double %x, i8* %context) unnamed_addr #1 {
-entry:
-  %0 = call double @llvm.sqrt.f64(double %x)
-  ret double %0
-}
-
-; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
-declare double @llvm.sqrt.f64(double) #2
-
-; Function Attrs: nounwind
-define hidden double @main.myTrunc(double %x, i8* %context) unnamed_addr #1 {
-entry:
-  %0 = call double @llvm.trunc.f64(double %x)
-  ret double %0
-}
-
-; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
-declare double @llvm.trunc.f64(double) #2
-
-attributes #0 = { "target-features"="+armv7-m,+hwdiv,+soft-float,+strict-align,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
-attributes #1 = { nounwind "target-features"="+armv7-m,+hwdiv,+soft-float,+strict-align,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
-attributes #2 = { nofree nosync nounwind readnone speculatable willreturn }

compiler/testdata/intrinsics-wasm.ll

@@ -1,38 +0,0 @@
-; ModuleID = 'intrinsics.go'
-source_filename = "intrinsics.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
-target triple = "wasm32-unknown-wasi"
-
-declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*) #0
-
-declare void @runtime.trackPointer(i8* nocapture readonly, i8*) #0
-
-; Function Attrs: nounwind
-define hidden void @main.init(i8* %context) unnamed_addr #1 {
-entry:
-  ret void
-}
-
-; Function Attrs: nounwind
-define hidden double @main.mySqrt(double %x, i8* %context) unnamed_addr #1 {
-entry:
-  %0 = call double @llvm.sqrt.f64(double %x)
-  ret double %0
-}
-
-; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
-declare double @llvm.sqrt.f64(double) #2
-
-; Function Attrs: nounwind
-define hidden double @main.myTrunc(double %x, i8* %context) unnamed_addr #1 {
-entry:
-  %0 = call double @llvm.trunc.f64(double %x)
-  ret double %0
-}
-
-; Function Attrs: nofree nosync nounwind readnone speculatable willreturn
-declare double @llvm.trunc.f64(double) #2
-
-attributes #0 = { "target-features"="+bulk-memory,+nontrapping-fptoint,+sign-ext" }
-attributes #1 = { nounwind "target-features"="+bulk-memory,+nontrapping-fptoint,+sign-ext" }
-attributes #2 = { nofree nosync nounwind readnone speculatable willreturn }

compiler/testdata/intrinsics.go

@@ -1,14 +0,0 @@
-package main
-
-// Test how intrinsics are lowered: either as regular calls to the math
-// functions or as LLVM builtins (such as llvm.sqrt.f64).
-
-import "math"
-
-func mySqrt(x float64) float64 {
-	return math.Sqrt(x)
-}
-
-func myTrunc(x float64) float64 {
-	return math.Trunc(x)
-}

compiler/volatile.go

@@ -6,7 +6,7 @@ package compiler
 // createVolatileLoad is the implementation of the intrinsic function
 // runtime/volatile.LoadT().
 func (b *builder) createVolatileLoad() {
-	b.createFunctionStart()
+	b.createFunctionStart(true)
 	addr := b.getValue(b.fn.Params[0])
 	b.createNilCheck(b.fn.Params[0], addr, "deref")
 	val := b.CreateLoad(addr, "")
@@ -17,7 +17,7 @@ func (b *builder) createVolatileLoad() {
 // createVolatileStore is the implementation of the intrinsic function
 // runtime/volatile.StoreT().
 func (b *builder) createVolatileStore() {
-	b.createFunctionStart()
+	b.createFunctionStart(true)
 	addr := b.getValue(b.fn.Params[0])
 	val := b.getValue(b.fn.Params[1])
 	b.createNilCheck(b.fn.Params[0], addr, "deref")