[dev.simd] simd: add carryless multiply

now with comments, and also a test.

choice of data types, method names, etc, are all up for comment.
It's NOT commutative, because of the immediate operand (unless we
swap the bits of the immediate).

Change-Id: I730a6938c6803d0b93544445db65eadc51783e42
Reviewed-on: https://go-review.googlesource.com/c/go/+/726963
Reviewed-by: Junyang Shao <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>

Author: dr2chase

src/cmd/compile/internal/amd64/simdssa.go

@@ -1333,6 +1333,9 @@
 (blendMaskedInt16x32 x y mask) => (VPBLENDMWMasked512 x y (VPMOVVec16x32ToM <types.TypeMask> mask))
 (blendMaskedInt32x16 x y mask) => (VPBLENDMDMasked512 x y (VPMOVVec32x16ToM <types.TypeMask> mask))
 (blendMaskedInt64x8 x y mask) => (VPBLENDMQMasked512 x y (VPMOVVec64x8ToM <types.TypeMask> mask))
+(carrylessMultiplyUint64x2 ...) => (VPCLMULQDQ128 ...)
+(carrylessMultiplyUint64x4 ...) => (VPCLMULQDQ256 ...)
+(carrylessMultiplyUint64x8 ...) => (VPCLMULQDQ512 ...)
 (concatSelectedConstantFloat32x4 ...) => (VSHUFPS128 ...)
 (concatSelectedConstantFloat64x2 ...) => (VSHUFPD128 ...)
 (concatSelectedConstantInt32x4 ...) => (VSHUFPS128 ...)

src/cmd/compile/internal/ssa/_gen/simdAMD64.rules

@@ -19,3 +19,5 @@
   documentation: !string |-
     // NAME computes element-wise GF(2^8) multiplication with
     // reduction polynomial x^8 + x^4 + x^3 + x + 1.
+- go: carrylessMultiply
+  commutative: false

src/cmd/compile/internal/ssa/_gen/simdAMD64ops.go

@@ -30,3 +30,63 @@
   - *uint8
   out:
   - *uint8
+
+- go: carrylessMultiply
+  documentation: !string |-
+    // NAME computes one of four possible Galois polynomial
+    // products of selected high and low halves of x and y,
+    // depending on the value of xyHiLo, returning the 128-bit
+    // product in the concatenated two elements of the result.
+    // Bit 0 selects the low (0) or high (1) element of x and
+    // bit 4 selects the low (0x00) or high (0x10) element of y.
+  asm: V?PCLMULQDQ
+  in:
+  - go: Uint64x2
+  - go: Uint64x2
+  - class: immediate
+    immOffset: 0
+    name: xyHiLo
+  out:
+  - go: Uint64x2
+    overwriteElementBits: 64
+  hideMaskMethods: true
+
+- go: carrylessMultiply
+  documentation: !string |-
+    // NAME computes one of two possible Galois polynomial
+    // products of selected high and low halves of each of the two
+    // 128-bit lanes of x and y, depending on the value of xyHiLo,
+    // and returns the four 128-bit products in the result's lanes.
+    // Bit 0 selects the low (0) or high (1) elements of x's lanes and
+    // bit 4 selects the low (0x00) or high (0x10) elements of y's lanes.
+  asm: V?PCLMULQDQ
+  in:
+  - go: Uint64x4
+  - go: Uint64x4
+  - class: immediate
+    immOffset: 0
+    name: xyHiLo
+  out:
+  - go: Uint64x4
+    overwriteElementBits: 64
+  hideMaskMethods: true
+
+- go: carrylessMultiply
+  documentation: !string |-
+    // NAME computes one of four possible Galois polynomial
+    // products of selected high and low halves of each of the four
+    // 128-bit lanes of x and y, depending on the value of xyHiLo,
+    // and returns the four 128-bit products in the result's lanes.
+    // Bit 0 selects the low (0) or high (1) elements of x's lanes and
+    // bit 4 selects the low (0x00) or high (0x10) elements of y's lanes.
+  asm: V?PCLMULQDQ
+  in:
+  - go: Uint64x8
+  - go: Uint64x8
+  - class: immediate
+    immOffset: 0
+    name: xyHiLo
+  out:
+  - go: Uint64x8
+    overwriteElementBits: 64
+  hideMaskMethods: true

src/cmd/compile/internal/ssa/_gen/simdgenericOps.go

@@ -83,6 +83,9 @@ in: !repeat
   - {class: vreg, go: Int64x4,    base: "int",   elemBits: 128, bits: 256, lanes: 4}
   - {class: vreg, go: Uint64x4,   base: "uint",  elemBits: 128, bits: 256, lanes: 4}
 
+# Special for carryless multiply
+  - {class: vreg, go: Uint64x8,   base: "uint",  elemBits: 128, bits: 512, lanes: 8}
+
 # Special shapes just to make VAES(ENC|DEC)(LAST)?512 work.
 # The elemBits field of these shapes are wrong, it would be overwritten by overwriteElemBits.
   - {class: vreg, go: Int8x32,    base: "int",   elemBits: 128, bits: 512, lanes: 32}