compiler: fix min/max on floats by using intrinsics

This fixes two edge cases for min/max:
min/max(+0.0, -0.0) = -0.0, +0.0
min/max(number, NaN) = NaN, NaN

The compare and select method does not work here.
I switched to the llvm.minimum/llvm.maximum intrinsics which match the intended behavior.

The integer min/max were also swapped over to using intrinsics.
The compare and select path is now only used by strings.
This commit is contained in:
Nia Waldvogel
2025-12-28 09:44:58 -05:00
committed by Ron Evans
parent 8ef36ed939
commit bc9708f51a
4 changed files with 105 additions and 41 deletions
+59 -13
View File
@@ -1716,20 +1716,66 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
return llvmLen, nil
case "min", "max":
// min and max builtins, added in Go 1.21.
// We can simply reuse the existing binop comparison code, which has all
// the edge cases figured out already.
tok := token.LSS
if callName == "max" {
tok = token.GTR
}
result := argValues[0]
typ := argTypes[0]
for _, arg := range argValues[1:] {
cmp, err := b.createBinOp(tok, typ, typ, result, arg, pos)
if err != nil {
return result, err
// Find the corresponding intrinsic name.
ty := argTypes[0].Underlying().(*types.Basic)
llvmType := b.getLLVMType(ty)
info := ty.Info()
var prefix, delimeter, typeName string
if info&types.IsInteger != 0 {
// This is an integer value.
// Use the LLVM int min/max intrinsics.
prefix = "llvm.s"
if info&types.IsUnsigned != 0 {
prefix = "llvm.u"
}
result = b.CreateSelect(cmp, result, arg, "")
delimeter = ".i"
typeName = strconv.Itoa(llvmType.IntTypeWidth())
} else {
switch ty.Kind() {
case types.String:
// Strings do not have an equivalent intrinsic.
// Implement with compares and selects.
tok := token.LSS
if callName == "max" {
tok = token.GTR
}
result := argValues[0]
typ := argTypes[0]
for _, arg := range argValues[1:] {
cmp, err := b.createBinOp(tok, typ, typ, result, arg, pos)
if err != nil {
return result, err
}
result = b.CreateSelect(cmp, result, arg, "")
}
return result, nil
case types.Float32:
typeName = "f32"
case types.Float64:
typeName = "f64"
default:
return llvm.Value{}, b.makeError(pos, "todo: min/max: unknown type")
}
// There are a few edge cases with floating point min/max:
// min(-0.0, +0.0) = -0.0
// min(NaN, number) = NaN
// The llvm.minimum.*/llvm.maximum.* intrinsics match this behavior.
// Neither Go nor LLVM defines the bit representation of resulting NaNs.
prefix = "llvm."
delimeter = "imum."
}
intrinsicName := prefix + callName + delimeter + typeName
// Find or create the intrinsic.
llvmFn := b.mod.NamedFunction(intrinsicName)
if llvmFn.IsNil() {
fnType := llvm.FunctionType(llvmType, []llvm.Type{llvmType, llvmType}, false)
llvmFn = llvm.AddFunction(b.mod, intrinsicName, fnType)
}
// Call the intrinsic repeatedly to merge the arguments.
callType := llvmFn.GlobalValueType()
result := argValues[0]
for _, arg := range argValues[1:] {
result = b.CreateCall(callType, llvmFn, []llvm.Value{result, arg}, "")
}
return result, nil
case "panic":
+33 -27
View File
@@ -22,6 +22,9 @@ entry:
ret i32 %a
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.smin.i32(i32, i32) #3
; Function Attrs: nounwind
define hidden i32 @main.min2(i32 %a, i32 %b, ptr %context) unnamed_addr #2 {
entry:
@@ -53,6 +56,9 @@ entry:
ret i8 %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i8 @llvm.umin.i8(i8, i8) #3
; Function Attrs: nounwind
define hidden i32 @main.minUnsigned(i32 %a, i32 %b, ptr %context) unnamed_addr #2 {
entry:
@@ -60,22 +66,29 @@ entry:
ret i32 %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.umin.i32(i32, i32) #3
; Function Attrs: nounwind
define hidden float @main.minFloat32(float %a, float %b, ptr %context) unnamed_addr #2 {
entry:
%0 = fcmp olt float %a, %b
%1 = select i1 %0, float %a, float %b
ret float %1
%0 = call float @llvm.minimum.f32(float %a, float %b)
ret float %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare float @llvm.minimum.f32(float, float) #3
; Function Attrs: nounwind
define hidden double @main.minFloat64(double %a, double %b, ptr %context) unnamed_addr #2 {
entry:
%0 = fcmp olt double %a, %b
%1 = select i1 %0, double %a, double %b
ret double %1
%0 = call double @llvm.minimum.f64(double %a, double %b)
ret double %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare double @llvm.minimum.f64(double, double) #3
; Function Attrs: nounwind
define hidden %runtime._string @main.minString(ptr readonly %a.data, i32 %a.len, ptr readonly %b.data, i32 %b.len, ptr %context) unnamed_addr #2 {
entry:
@@ -100,6 +113,9 @@ entry:
ret i32 %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.smax.i32(i32, i32) #3
; Function Attrs: nounwind
define hidden i32 @main.maxUint(i32 %a, i32 %b, ptr %context) unnamed_addr #2 {
entry:
@@ -107,14 +123,19 @@ entry:
ret i32 %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.umax.i32(i32, i32) #3
; Function Attrs: nounwind
define hidden float @main.maxFloat32(float %a, float %b, ptr %context) unnamed_addr #2 {
entry:
%0 = fcmp ogt float %a, %b
%1 = select i1 %0, float %a, float %b
ret float %1
%0 = call float @llvm.maximum.f32(float %a, float %b)
ret float %0
}
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare float @llvm.maximum.f32(float, float) #3
; Function Attrs: nounwind
define hidden %runtime._string @main.maxString(ptr readonly %a.data, i32 %a.len, ptr readonly %b.data, i32 %b.len, ptr %context) unnamed_addr #2 {
entry:
@@ -139,7 +160,7 @@ entry:
}
; Function Attrs: nocallback nofree nounwind willreturn memory(argmem: write)
declare void @llvm.memset.p0.i32(ptr nocapture writeonly, i8, i32, i1 immarg) #3
declare void @llvm.memset.p0.i32(ptr nocapture writeonly, i8, i32, i1 immarg) #4
; Function Attrs: nounwind
define hidden void @main.clearZeroSizedSlice(ptr %s.data, i32 %s.len, i32 %s.cap, ptr %context) unnamed_addr #2 {
@@ -156,24 +177,9 @@ entry:
declare void @runtime.hashmapClear(ptr dereferenceable_or_null(40), ptr) #1
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.smin.i32(i32, i32) #4
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i8 @llvm.umin.i8(i8, i8) #4
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.umin.i32(i32, i32) #4
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.smax.i32(i32, i32) #4
; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
declare i32 @llvm.umax.i32(i32, i32) #4
attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
attributes #1 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
attributes #2 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
attributes #3 = { nocallback nofree nounwind willreturn memory(argmem: write) }
attributes #4 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #3 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #4 = { nocallback nofree nounwind willreturn memory(argmem: write) }
attributes #5 = { nounwind }
+11 -1
View File
@@ -1,15 +1,25 @@
package main
import "math"
func main() {
// The new min/max builtins.
// With int:
ia := 1
ib := 5
ic := -3
println("min/max:", min(ia, ib, ic), max(ia, ib, ic))
// With float:
fa := 1.0
fb := 5.0
fc := -3.0
println("min/max:", min(ia, ib, ic), max(ia, ib, ic))
println("min/max:", min(fa, fb, fc), max(fa, fb, fc))
// Float +/- 0.0:
pos0 := 0.0
neg0 := -pos0
println("min/max:", min(pos0, neg0), max(pos0, neg0))
// Float NaN:
println("min/max:", min(math.NaN(), 12.0), max(math.NaN(), 12.0))
// The clear builtin, for slices.
s := []int{1, 2, 3, 4, 5}
+2
View File
@@ -1,5 +1,7 @@
min/max: -3 5
min/max: -3.000000e+000 +5.000000e+000
min/max: -0.000000e+000 +0.000000e+000
min/max: NaN NaN
cleared s[:3]: 0 0 0 4 5
cleared map: 0
added to cleared map: four 1