riscv64: Implement a few SIMD arithmetic ops (bytecodealliance#6268)

* riscv64: Swap order of `VecAluRRR` source registers

These were accidentally reversed from what we declare in the isle emit helper

* riscv64: Add SIMD `isub`

* riscv64: Add SIMD `imul`

* riscv64: Add `{u,s}mulhi`

* riscv64: Add `b{and,or,xor}`

* cranelift: Move `imul.i8x16` runtest to separate file

Looks like x86 does not implement it

* riscv64: Better formatting for `VecAluOpRRR`

* cranelift: Enable x86 SIMD tests with `has_sse41=false`

cranelift/codegen/src/isa/riscv64/inst.isle

@@ -326,8 +326,8 @@
     (VecAluRRR
       (op VecAluOpRRR)
       (vd WritableReg)
-      (vs1 Reg)
       (vs2 Reg)
+      (vs1 Reg)
       (vstate VState))
 
     (VecSetState

cranelift/codegen/src/isa/riscv64/inst/encode.rs

@@ -50,7 +50,7 @@ pub fn encode_valu(
 ) -> u32 {
     let funct6 = funct6 & 0b111111;
     let vm = vm & 0b1;
-    let funct7 = (funct6 << 6) | vm;
+    let funct7 = (funct6 << 1) | vm;
     encode_r_type(opcode, vd, funct3, vs1, vs2, funct7)
 }
 

cranelift/codegen/src/isa/riscv64/inst/vector.rs

@@ -214,27 +214,42 @@ impl fmt::Display for VState {
 
 impl VecAluOpRRR {
     pub fn opcode(&self) -> u32 {
-        match self {
-            VecAluOpRRR::Vadd => 0x57,
-        }
+        // Vector Opcode
+        0x57
     }
     pub fn funct3(&self) -> u32 {
         match self {
-            VecAluOpRRR::Vadd => 0b000,
+            // OPIVV
+            VecAluOpRRR::Vadd
+            | VecAluOpRRR::Vsub
+            | VecAluOpRRR::Vand
+            | VecAluOpRRR::Vor
+            | VecAluOpRRR::Vxor => 0b000,
+            // OPIMV
+            VecAluOpRRR::Vmul | VecAluOpRRR::Vmulh | VecAluOpRRR::Vmulhu => 0b010,
         }
     }
     pub fn funct6(&self) -> u32 {
+        // See: https://github.com/riscv/riscv-v-spec/blob/master/inst-table.adoc
         match self {
             VecAluOpRRR::Vadd => 0b000000,
+            VecAluOpRRR::Vsub => 0b000010,
+            VecAluOpRRR::Vmul => 0b100101,
+            VecAluOpRRR::Vmulh => 0b100111,
+            VecAluOpRRR::Vmulhu => 0b100100,
+            VecAluOpRRR::Vand => 0b001001,
+            VecAluOpRRR::Vor => 0b001010,
+            VecAluOpRRR::Vxor => 0b001011,
         }
     }
 }
 
 impl fmt::Display for VecAluOpRRR {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        match self {
-            VecAluOpRRR::Vadd => write!(f, "vadd.vv"),
-        }
+        let mut s = format!("{self:?}");
+        s.make_ascii_lowercase();
+        s.push_str(".vv");
+        f.write_str(&s)
     }
 }
 

cranelift/codegen/src/isa/riscv64/inst_vector.isle

@@ -59,6 +59,13 @@
 ;; Register to Register ALU Ops
 (type VecAluOpRRR (enum
   (Vadd)
+  (Vsub)
+  (Vmul)
+  (Vmulh)
+  (Vmulhu)
+  (Vand)
+  (Vor)
+  (Vxor)
 ))
 
 
@@ -138,3 +145,38 @@
 (decl rv_vadd_vv (Reg Reg VState) Reg)
 (rule (rv_vadd_vv vs2 vs1 vstate)
   (vec_alu_rrr (VecAluOpRRR.Vadd) vs2 vs1 vstate))
+
+;; Helper for emitting the `vsub.vv` instruction.
+(decl rv_vsub_vv (Reg Reg VState) Reg)
+(rule (rv_vsub_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vsub) vs2 vs1 vstate))
+
+;; Helper for emitting the `vmul.vv` instruction.
+(decl rv_vmul_vv (Reg Reg VState) Reg)
+(rule (rv_vmul_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vmul) vs2 vs1 vstate))
+
+;; Helper for emitting the `vmulh.vv` instruction.
+(decl rv_vmulh_vv (Reg Reg VState) Reg)
+(rule (rv_vmulh_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vmulh) vs2 vs1 vstate))
+
+;; Helper for emitting the `vmulhu.vv` instruction.
+(decl rv_vmulhu_vv (Reg Reg VState) Reg)
+(rule (rv_vmulhu_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vmulhu) vs2 vs1 vstate))
+
+;; Helper for emitting the `vand.vv` instruction.
+(decl rv_vand_vv (Reg Reg VState) Reg)
+(rule (rv_vand_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vand) vs2 vs1 vstate))
+
+;; Helper for emitting the `vor.vv` instruction.
+(decl rv_vor_vv (Reg Reg VState) Reg)
+(rule (rv_vor_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vor) vs2 vs1 vstate))
+
+;; Helper for emitting the `vxor.vv` instruction.
+(decl rv_vxor_vv (Reg Reg VState) Reg)
+(rule (rv_vxor_vv vs2 vs1 vstate)
+  (vec_alu_rrr (VecAluOpRRR.Vxor) vs2 vs1 vstate))

cranelift/codegen/src/isa/riscv64/lower.isle

@@ -112,15 +112,19 @@
 ;;;; Rules for `isub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; Base case, simply subtracting things in registers.
 
-(rule -2 (lower (has_type (fits_in_64 ty) (isub x y)))
+(rule (lower (has_type (ty_int_ref_scalar_64 ty) (isub x y)))
   (rv_sub x y))
 
-(rule -1 (lower (has_type (fits_in_32 ty) (isub x y)))
+(rule 1 (lower (has_type (fits_in_32 (ty_int ty)) (isub x y)))
   (rv_subw x y))
 
-(rule (lower (has_type $I128 (isub x y)))
+(rule 2 (lower (has_type $I128 (isub x y)))
   (i128_sub x y))
 
+;; SIMD Vectors
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (isub x y)))
+  (rv_vsub_vv x y ty))
+
 ;;;; Rules for `ineg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; `i64` and smaller.
@@ -129,21 +133,14 @@
 
 ;;;; Rules for `imul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(rule -2 (lower (has_type (fits_in_64 ty) (imul x y)))
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (imul x y)))
   (rv_mul x y))
-(rule -1 (lower (has_type (fits_in_32 ty) (imul x y)))
-  (rv_mulw x y))
-
-;;;; Rules for `smulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule (lower (has_type (fits_in_64 ty) (smulhi x y)))
-  (lower_smlhi ty (ext_int_if_need $true x ty) (ext_int_if_need $true y ty)))
 
-;;;; Rules for `umulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule (lower (has_type (fits_in_64 ty) (umulhi x y)))
-  (lower_umlhi ty (ext_int_if_need $false x ty) (ext_int_if_need $false y ty)))
+(rule 1 (lower (has_type (fits_in_32 (ty_int ty)) (imul x y)))
+  (rv_mulw x y))
 
 ;; for I128
-(rule (lower (has_type $I128 (imul x y)))
+(rule 2 (lower (has_type $I128 (imul x y)))
   (let
     ((x_regs ValueRegs x)
       (x_lo Reg (value_regs_get x_regs 0))
@@ -169,6 +166,22 @@
       (dst_lo Reg (madd x_lo y_lo (zero_reg))))
     (value_regs dst_lo dst_hi)))
 
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (imul x y)))
+  (rv_vmul_vv x y ty))
+
+;;;; Rules for `smulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (smulhi x y)))
+  (lower_smlhi ty (ext_int_if_need $true x ty) (ext_int_if_need $true y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (smulhi x y)))
+  (rv_vmulh_vv x y ty))
+
+;;;; Rules for `umulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule 0 (lower (has_type (ty_int_ref_scalar_64 ty) (umulhi x y)))
+  (lower_umlhi ty (ext_int_if_need $false x ty) (ext_int_if_need $false y ty)))
+
+(rule 1 (lower (has_type (ty_vec_fits_in_register ty) (umulhi x y)))
+  (rv_vmulhu_vv x y ty))
 
 ;;;; Rules for `div` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
@@ -277,6 +290,10 @@
     (value_regs low high)))
 
 
+(rule 7 (lower (has_type (ty_vec_fits_in_register ty) (band x y)))
+  (rv_vand_vv x y ty))
+
+
 ;;;; Rules for `or` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 (rule -1 (lower (has_type (ty_int ty) (bor x y)))
   (gen_or ty x y))
@@ -320,6 +337,8 @@
       (high Reg (rv_orn (value_regs_get x 1) (value_regs_get y 1))))
     (value_regs low high)))
 
+(rule 7 (lower (has_type (ty_vec_fits_in_register ty) (bor x y)))
+  (rv_vor_vv x y ty))
 
 ;;;; Rules for `xor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 (rule -1 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x y)))
@@ -341,6 +360,8 @@
 (rule (lower (has_type $F64 (bxor x y)))
   (lower_float_binary (AluOPRRR.Xor) x y $F64))
 
+(rule 3 (lower (has_type (ty_vec_fits_in_register ty) (bxor x y)))
+  (rv_vxor_vv x y ty))
 
 ;;;; Rules for `bnot` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 (rule -1 (lower (has_type (ty_int ty) (bnot x)))

cranelift/filetests/filetests/isa/riscv64/simd-band.clif

@@ -0,0 +1,73 @@
+test compile precise-output
+set unwind_info=false
+target riscv64 has_v
+
+
+function %band_i8x16(i8x16, i8x16) -> i8x16 {
+block0(v0: i8x16, v1: i8x16):
+    v2 = band v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vand.vv v10,v10,v11 #avl=16, #vtype=(e8, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x08, 0xcc
+;   .byte 0x57, 0x85, 0xa5, 0x26
+;   ret
+
+function %band_i16x8(i16x8, i16x8) -> i16x8 {
+block0(v0: i16x8, v1: i16x8):
+    v2 = band v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vand.vv v10,v10,v11 #avl=8, #vtype=(e16, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x84, 0xcc
+;   .byte 0x57, 0x85, 0xa5, 0x26
+;   ret
+
+function %band_i32x4(i32x4, i32x4) -> i32x4 {
+block0(v0: i32x4, v1: i32x4):
+    v2 = band v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vand.vv v10,v10,v11 #avl=4, #vtype=(e32, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x02, 0xcd
+;   .byte 0x57, 0x85, 0xa5, 0x26
+;   ret
+
+function %band_i64x2(i64x2, i64x2) -> i64x2 {
+block0(v0: i64x2, v1: i64x2):
+    v2 = band v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vand.vv v10,v10,v11 #avl=2, #vtype=(e64, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x81, 0xcd
+;   .byte 0x57, 0x85, 0xa5, 0x26
+;   ret
+

cranelift/filetests/filetests/isa/riscv64/simd-bor.clif

@@ -0,0 +1,73 @@
+test compile precise-output
+set unwind_info=false
+target riscv64 has_v
+
+
+function %bor_i8x16(i8x16, i8x16) -> i8x16 {
+block0(v0: i8x16, v1: i8x16):
+    v2 = bor v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vor.vv v10,v10,v11 #avl=16, #vtype=(e8, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x08, 0xcc
+;   .byte 0x57, 0x85, 0xa5, 0x2a
+;   ret
+
+function %bor_i16x8(i16x8, i16x8) -> i16x8 {
+block0(v0: i16x8, v1: i16x8):
+    v2 = bor v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vor.vv v10,v10,v11 #avl=8, #vtype=(e16, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x84, 0xcc
+;   .byte 0x57, 0x85, 0xa5, 0x2a
+;   ret
+
+function %bor_i32x4(i32x4, i32x4) -> i32x4 {
+block0(v0: i32x4, v1: i32x4):
+    v2 = bor v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vor.vv v10,v10,v11 #avl=4, #vtype=(e32, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x02, 0xcd
+;   .byte 0x57, 0x85, 0xa5, 0x2a
+;   ret
+
+function %bor_i64x2(i64x2, i64x2) -> i64x2 {
+block0(v0: i64x2, v1: i64x2):
+    v2 = bor v0, v1
+    return v2
+}
+
+; VCode:
+; block0:
+;   vor.vv v10,v10,v11 #avl=2, #vtype=(e64, m1, ta, ma)
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   .byte 0x57, 0x70, 0x81, 0xcd
+;   .byte 0x57, 0x85, 0xa5, 0x2a
+;   ret
+