internal/cpu: add AVX-512-CD and DQ, and derived "basic AVX-512"

This adds detection for the CD and DQ sub-features of x86 AVX-512.

Building on these, we also add a "derived" AVX-512 feature that
bundles together the basic usable subset of subfeatures. Despite the F
in AVX-512-F standing for "foundation", AVX-512-F+BW+DQ+VL together
really form the basic usable subset of AVX-512 functionality. These
have also all been supported together by almost every CPU, and are
guaranteed by GOAMD64=v4, so there's little point in separating them
out.

This is a cherry-pick of CL 680899 from the dev.simd branch.

Change-Id: I34356502bd1853ba2372e48db0b10d55cffe07a1
Reviewed-on: https://go-review.googlesource.com/c/go/+/693396
Reviewed-by: Cherry Mui <cherryyz@google.com>
Auto-Submit: Austin Clements <austin@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>

src/internal/cpu/cpu.go

@@ -31,9 +31,13 @@ var X86 struct {
 	HasADX              bool
 	HasAVX              bool
 	HasAVX2             bool
+	HasAVX512           bool // Virtual feature: F+CD+BW+DQ+VL
 	HasAVX512F          bool
+	HasAVX512CD         bool
 	HasAVX512BW         bool
+	HasAVX512DQ         bool
 	HasAVX512VL         bool
+	HasAVX512VPCLMULQDQ bool
 	HasBMI1             bool
 	HasBMI2             bool
 	HasERMS             bool
@@ -48,7 +52,6 @@ var X86 struct {
 	HasSSSE3            bool
 	HasSSE41            bool
 	HasSSE42            bool
-	HasAVX512VPCLMULQDQ bool
 	_                   CacheLinePad
 }
 
@@ -161,6 +164,10 @@ var RISCV64 struct {
 //go:linkname S390X
 //go:linkname RISCV64
 
+// doDerived, if non-nil, is called after processing GODEBUG to set "derived"
+// feature flags.
+var doDerived func()
+
 // Initialize examines the processor and sets the relevant variables above.
 // This is called by the runtime package early in program initialization,
 // before normal init functions are run. env is set by runtime if the OS supports
@@ -168,6 +175,9 @@ var RISCV64 struct {
 func Initialize(env string) {
 	doinit()
 	processOptions(env)
+	if doDerived != nil {
+		doDerived()
+	}
 }
 
 // options contains the cpu debug options that can be used in GODEBUG.

src/internal/cpu/cpu_x86.go

@@ -36,7 +36,9 @@ const (
 	cpuid_BMI2     = 1 << 8
 	cpuid_ERMS     = 1 << 9
 	cpuid_AVX512F  = 1 << 16
+	cpuid_AVX512DQ = 1 << 17
 	cpuid_ADX      = 1 << 19
+	cpuid_AVX512CD = 1 << 28
 	cpuid_SHA      = 1 << 29
 	cpuid_AVX512BW = 1 << 30
 	cpuid_AVX512VL = 1 << 31
@@ -89,7 +91,9 @@ func doinit() {
 		// they can be turned off.
 		options = append(options,
 			option{Name: "avx512f", Feature: &X86.HasAVX512F},
+			option{Name: "avx512cd", Feature: &X86.HasAVX512CD},
 			option{Name: "avx512bw", Feature: &X86.HasAVX512BW},
+			option{Name: "avx512dq", Feature: &X86.HasAVX512DQ},
 			option{Name: "avx512vl", Feature: &X86.HasAVX512VL},
 		)
 	}
@@ -154,7 +158,9 @@ func doinit() {
 
 	X86.HasAVX512F = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
 	if X86.HasAVX512F {
+		X86.HasAVX512CD = isSet(ebx7, cpuid_AVX512CD)
 		X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
+		X86.HasAVX512DQ = isSet(ebx7, cpuid_AVX512DQ)
 		X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
 		X86.HasAVX512VPCLMULQDQ = isSet(ecx7, cpuid_AVX512VPCLMULQDQ)
 	}
@@ -170,6 +176,17 @@ func doinit() {
 
 	_, _, _, edxExt1 := cpuid(0x80000001, 0)
 	X86.HasRDTSCP = isSet(edxExt1, cpuid_RDTSCP)
+
+	doDerived = func() {
+		// Rather than carefully gating on fundamental AVX-512 features, we have
+		// a virtual "AVX512" feature that captures F+CD+BW+DQ+VL. BW, DQ, and
+		// VL have a huge effect on which AVX-512 instructions are available,
+		// and these have all been supported on everything except the earliest
+		// Phi chips with AVX-512. No CPU has had CD without F, so we include
+		// it. GOAMD64=v4 also implies exactly this set, and these are all
+		// included in AVX10.1.
+		X86.HasAVX512 = X86.HasAVX512F && X86.HasAVX512CD && X86.HasAVX512BW && X86.HasAVX512DQ && X86.HasAVX512VL
+	}
 }
 
 func isSet(hwc uint32, value uint32) bool {