A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
riscv64: Fix `replicated_imm5` lowering rule (#12956)
Alex Crichton3c714e360379
diff --git a/cranelift/codegen/src/isa/aarch64/inst.isle b/cranelift/codegen/src/isa/aarch64/inst.isleindex 3be84db617..0ec5902780 100644--- a/cranelift/codegen/src/isa/aarch64/inst.isle+++ b/cranelift/codegen/src/isa/aarch64/inst.isle@@ -4112,13 +4112,13 @@;; as it will make it easier to create the immediates in the instructions below.(rule 5 (splat_const (u64_replicated_u32 n) (VectorSize.Size64x2))(splat_const n (VectorSize.Size32x4)))-(rule 5 (splat_const (u32_replicated_u16 n) (VectorSize.Size32x4))+(rule 5 (splat_const (u32_from_u64 (u32_replicated_u16 n)) (VectorSize.Size32x4))(splat_const n (VectorSize.Size16x8)))-(rule 5 (splat_const (u32_replicated_u16 n) (VectorSize.Size32x2))+(rule 5 (splat_const (u32_from_u64 (u32_replicated_u16 n)) (VectorSize.Size32x2))(splat_const n (VectorSize.Size16x4)))-(rule 5 (splat_const (u16_replicated_u8 n) (VectorSize.Size16x8))+(rule 5 (splat_const (u16_from_u64 (u16_replicated_u8 n)) (VectorSize.Size16x8))(splat_const n (VectorSize.Size8x16)))-(rule 5 (splat_const (u16_replicated_u8 n) (VectorSize.Size16x4))+(rule 5 (splat_const (u16_from_u64 (u16_replicated_u8 n)) (VectorSize.Size16x4))(splat_const n (VectorSize.Size8x8)));; Special cases for `vec_dup_imm` instructions where the input is eitherdiff --git a/cranelift/codegen/src/isa/riscv64/inst.isle b/cranelift/codegen/src/isa/riscv64/inst.isleindex 80521217ed..94e408d655 100644--- a/cranelift/codegen/src/isa/riscv64/inst.isle+++ b/cranelift/codegen/src/isa/riscv64/inst.isle@@ -2020,9 +2020,15 @@(decl imm5_from_i64 (Imm5) i64)(extern extractor imm5_from_i64 imm5_from_i64)-;; Construct a Imm5 from an i8+;; Construct a Imm5 from an iNN(decl pure partial i8_to_imm5 (i8) Imm5)(extern constructor i8_to_imm5 i8_to_imm5)+(decl pure partial i16_to_imm5 (i16) Imm5)+(rule (i16_to_imm5 (i8_from_i16 n)) (i8_to_imm5 n))+(decl pure partial i32_to_imm5 (i32) Imm5)+(rule (i32_to_imm5 (i8_from_i32 n)) (i8_to_imm5 n))+(decl pure partial i64_to_imm5 (i64) Imm5)+(rule (i64_to_imm5 (i8_from_i64 n)) (i8_to_imm5 n));; Helper to go directly from a `Value` to an `Imm5`.(decl imm5_from_value (Imm5) Value)@@ -2037,11 +2043,20 @@;; Constructor that matches a `Value` equivalent to a replicated Imm5 on all lanes.(decl pure partial replicated_imm5 (Value) Imm5)(rule (replicated_imm5 (splat _ (imm5_from_value n))) n)-(rule (replicated_imm5 (vconst _ (u128_from_constant n128)))+(rule (replicated_imm5 (vconst (multi_lane 64 _) (u128_from_constant n128)))(if-let (u128_replicated_u64 n64) n128)- (if-let (u64_replicated_u32 n32) n64)- (if-let (u32_replicated_u16 n16) n32)- (if-let (u16_replicated_u8 n8) n16)+ (if-let n (i64_to_imm5 (u64_cast_signed n64)))+ n)+(rule (replicated_imm5 (vconst (multi_lane 32 _) (u128_from_constant n128)))+ (if-let (u128_replicated_u32 n32) n128)+ (if-let n (i32_to_imm5 (u32_cast_signed n32)))+ n)+(rule (replicated_imm5 (vconst (multi_lane 16 _) (u128_from_constant n128)))+ (if-let (u128_replicated_u16 n16) n128)+ (if-let n (i16_to_imm5 (u16_cast_signed n16)))+ n)+(rule (replicated_imm5 (vconst (multi_lane 8 _) (u128_from_constant n128)))+ (if-let (u128_replicated_u8 n8) n128)(if-let n (i8_to_imm5 (u8_cast_signed n8)))n)diff --git a/cranelift/codegen/src/isle_prelude.rs b/cranelift/codegen/src/isle_prelude.rsindex c63309fbd3..a5dfce0ad4 100644--- a/cranelift/codegen/src/isle_prelude.rs+++ b/cranelift/codegen/src/isle_prelude.rs@@ -792,27 +792,25 @@ macro_rules! isle_common_prelude_methods {}}- fn u64_replicated_u32(&mut self, val: u64) -> Option<u64> {+ fn u64_replicated_u32(&mut self, val: u64) -> Option<u32> {let low32 = val as u32 as u64;if (low32 | (low32 << 32)) == val {- Some(low32)+ Some(val as u32)} else {None}}- fn u32_replicated_u16(&mut self, val: u64) -> Option<u64> {- let val = val as u32;+ fn u32_replicated_u16(&mut self, val: u32) -> Option<u16> {let low16 = val as u16 as u32;if (low16 | (low16 << 16)) == val {- Some(low16.into())+ Some(low16 as u16)} else {None}}- fn u16_replicated_u8(&mut self, val: u64) -> Option<u8> {- let val = val as u16;+ fn u16_replicated_u8(&mut self, val: u16) -> Option<u8> {let low8 = val as u8 as u16;if (low8 | (low8 << 8)) == val {Some(low8 as u8)diff --git a/cranelift/codegen/src/prelude.isle b/cranelift/codegen/src/prelude.isleindex 76a9773f42..8beecf6cd2 100644--- a/cranelift/codegen/src/prelude.isle+++ b/cranelift/codegen/src/prelude.isle@@ -115,13 +115,21 @@;; lower half of the input(decl u128_replicated_u64 (u64) u128)(extern extractor u128_replicated_u64 u128_replicated_u64)-(decl u64_replicated_u32 (u64) u64)+(decl u64_replicated_u32 (u32) u64)(extern extractor u64_replicated_u32 u64_replicated_u32)-(decl u32_replicated_u16 (u64) u64)+(decl u32_replicated_u16 (u16) u32)(extern extractor u32_replicated_u16 u32_replicated_u16)-(decl u16_replicated_u8 (u8) u64)+(decl u16_replicated_u8 (u8) u16)(extern extractor u16_replicated_u8 u16_replicated_u8)+;; Convenience extractors to go directly from u128 to something smaller+(decl u128_replicated_u32 (u32) u128)+(extractor (u128_replicated_u32 n) (u128_replicated_u64 (u64_replicated_u32 n)))+(decl u128_replicated_u16 (u16) u128)+(extractor (u128_replicated_u16 n) (u128_replicated_u32 (u32_replicated_u16 n)))+(decl u128_replicated_u8 (u8) u128)+(extractor (u128_replicated_u8 n) (u128_replicated_u16 (u16_replicated_u8 n)))+;; Get the low and high bits of a `u128` as `u64`s.(decl u128_low_bits (u128) u64)(extern constructor u128_low_bits u128_low_bits)diff --git a/cranelift/filetests/filetests/isa/riscv64/simd-iadd-opt.clif b/cranelift/filetests/filetests/isa/riscv64/simd-iadd-opt.clifnew file mode 100644index 0000000000..28c29d74a2--- /dev/null+++ b/cranelift/filetests/filetests/isa/riscv64/simd-iadd-opt.clif@@ -0,0 +1,330 @@+test compile precise-output+set enable_multi_ret_implicit_sret+set unwind_info=false+target riscv64 has_v+++function %iadd_const_i8x16(i8x16) -> i8x16 {+block0(v0: i8x16):+ v2 = vconst.i8x16 0x01010101_01010101_01010101_01010101+ v3 = iadd v0, v2++ v4 = vconst.i8x16 0x00010001_00010001_00010001_00010001+ v5 = iadd v3, v4++ v6 = vconst.i8x16 0x00000001_00000001_00000001_00000001+ v7 = iadd v5, v6++ v8 = vconst.i8x16 0x00000000_00000001_00000000_00000001+ v9 = iadd v7, v8++ return v9+}++; VCode:+; addi sp,sp,-16+; sd ra,8(sp)+; sd fp,0(sp)+; mv fp,sp+; block0:+; vle8.v v8,-16(incoming_arg) #avl=16, #vtype=(e8, m1, ta, ma)+; vadd.vi v10,v8,1 #avl=16, #vtype=(e8, m1, ta, ma)+; vle8.v v11,[const(2)] #avl=16, #vtype=(e8, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=16, #vtype=(e8, m1, ta, ma)+; vle8.v v11,[const(1)] #avl=16, #vtype=(e8, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=16, #vtype=(e8, m1, ta, ma)+; vle8.v v11,[const(0)] #avl=16, #vtype=(e8, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=16, #vtype=(e8, m1, ta, ma)+; vse8.v v10,0(a0) #avl=16, #vtype=(e8, m1, ta, ma)+; ld ra,8(sp)+; ld fp,0(sp)+; addi sp,sp,16+; ret+;+; Disassembled:+; block0: ; offset 0x0+; addi sp, sp, -0x10+; sd ra, 8(sp)+; sd s0, 0(sp)+; mv s0, sp+; block1: ; offset 0x10+; .byte 0x57, 0x70, 0x08, 0xcc+; addi t6, sp, 0x10+; .byte 0x07, 0x84, 0x0f, 0x02+; .byte 0x57, 0xb5, 0x80, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0x85, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0x85, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0x85, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; .byte 0x27, 0x05, 0x05, 0x02+; ld ra, 8(sp)+; ld s0, 0(sp)+; addi sp, sp, 0x10+; ret+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00++function %iadd_const_i16x8(i16x8) -> i16x8 {+block0(v0: i16x8):+ v2 = vconst.i16x8 0x01010101_01010101_01010101_01010101+ v3 = iadd v0, v2++ v4 = vconst.i16x8 0x00010001_00010001_00010001_00010001+ v5 = iadd v3, v4++ v6 = vconst.i16x8 0x00000001_00000001_00000001_00000001+ v7 = iadd v5, v6++ v8 = vconst.i16x8 0x00000000_00000001_00000000_00000001+ v9 = iadd v7, v8++ return v9+}++; VCode:+; addi sp,sp,-16+; sd ra,8(sp)+; sd fp,0(sp)+; mv fp,sp+; block0:+; vle8.v v8,-16(incoming_arg) #avl=16, #vtype=(e8, m1, ta, ma)+; vle16.v v10,[const(2)] #avl=8, #vtype=(e16, m1, ta, ma)+; vadd.vv v10,v8,v10 #avl=8, #vtype=(e16, m1, ta, ma)+; vadd.vi v10,v10,1 #avl=8, #vtype=(e16, m1, ta, ma)+; vle16.v v11,[const(1)] #avl=8, #vtype=(e16, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=8, #vtype=(e16, m1, ta, ma)+; vle16.v v11,[const(0)] #avl=8, #vtype=(e16, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=8, #vtype=(e16, m1, ta, ma)+; vse8.v v10,0(a0) #avl=16, #vtype=(e8, m1, ta, ma)+; ld ra,8(sp)+; ld fp,0(sp)+; addi sp,sp,16+; ret+;+; Disassembled:+; block0: ; offset 0x0+; addi sp, sp, -0x10+; sd ra, 8(sp)+; sd s0, 0(sp)+; mv s0, sp+; block1: ; offset 0x10+; .byte 0x57, 0x70, 0x08, 0xcc+; addi t6, sp, 0x10+; .byte 0x07, 0x84, 0x0f, 0x02+; .byte 0x57, 0x70, 0x84, 0xcc+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x07, 0xd5, 0x0f, 0x02+; .byte 0x57, 0x05, 0x85, 0x02+; .byte 0x57, 0xb5, 0xa0, 0x02+; auipc t6, 0+; addi t6, t6, 0x4c+; .byte 0x87, 0xd5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; auipc t6, 0+; addi t6, t6, 0x4c+; .byte 0x87, 0xd5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; .byte 0x57, 0x70, 0x08, 0xcc+; .byte 0x27, 0x05, 0x05, 0x02+; ld ra, 8(sp)+; ld s0, 0(sp)+; addi sp, sp, 0x10+; ret+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00++function %iadd_const_i32x4(i32x4) -> i32x4 {+block0(v0: i32x4):+ v2 = vconst.i32x4 0x01010101_01010101_01010101_01010101+ v3 = iadd v0, v2++ v4 = vconst.i32x4 0x00010001_00010001_00010001_00010001+ v5 = iadd v3, v4++ v6 = vconst.i32x4 0x00000001_00000001_00000001_00000001+ v7 = iadd v5, v6++ v8 = vconst.i32x4 0x00000000_00000001_00000000_00000001+ v9 = iadd v7, v8++ return v9+}++; VCode:+; addi sp,sp,-16+; sd ra,8(sp)+; sd fp,0(sp)+; mv fp,sp+; block0:+; vle8.v v8,-16(incoming_arg) #avl=16, #vtype=(e8, m1, ta, ma)+; vle32.v v10,[const(2)] #avl=4, #vtype=(e32, m1, ta, ma)+; vadd.vv v10,v8,v10 #avl=4, #vtype=(e32, m1, ta, ma)+; vle32.v v11,[const(1)] #avl=4, #vtype=(e32, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=4, #vtype=(e32, m1, ta, ma)+; vadd.vi v10,v10,1 #avl=4, #vtype=(e32, m1, ta, ma)+; vle32.v v11,[const(0)] #avl=4, #vtype=(e32, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=4, #vtype=(e32, m1, ta, ma)+; vse8.v v10,0(a0) #avl=16, #vtype=(e8, m1, ta, ma)+; ld ra,8(sp)+; ld fp,0(sp)+; addi sp,sp,16+; ret+;+; Disassembled:+; block0: ; offset 0x0+; addi sp, sp, -0x10+; sd ra, 8(sp)+; sd s0, 0(sp)+; mv s0, sp+; block1: ; offset 0x10+; .byte 0x57, 0x70, 0x08, 0xcc+; addi t6, sp, 0x10+; .byte 0x07, 0x84, 0x0f, 0x02+; .byte 0x57, 0x70, 0x02, 0xcd+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x07, 0xe5, 0x0f, 0x02+; .byte 0x57, 0x05, 0x85, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0xe5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; .byte 0x57, 0xb5, 0xa0, 0x02+; auipc t6, 0+; addi t6, t6, 0x4c+; .byte 0x87, 0xe5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; .byte 0x57, 0x70, 0x08, 0xcc+; .byte 0x27, 0x05, 0x05, 0x02+; ld ra, 8(sp)+; ld s0, 0(sp)+; addi sp, sp, 0x10+; ret+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x00, 0x00, 0x00, 0x00++function %iadd_const_i64x2(i64x2) -> i64x2 {+block0(v0: i64x2):+ v2 = vconst.i64x2 0x01010101_01010101_01010101_01010101+ v3 = iadd v0, v2++ v4 = vconst.i64x2 0x00010001_00010001_00010001_00010001+ v5 = iadd v3, v4++ v6 = vconst.i64x2 0x00000001_00000001_00000001_00000001+ v7 = iadd v5, v6++ v8 = vconst.i64x2 0x00000000_00000001_00000000_00000001+ v9 = iadd v7, v8++ return v9+}++; VCode:+; addi sp,sp,-16+; sd ra,8(sp)+; sd fp,0(sp)+; mv fp,sp+; block0:+; vle8.v v8,-16(incoming_arg) #avl=16, #vtype=(e8, m1, ta, ma)+; vle64.v v10,[const(2)] #avl=2, #vtype=(e64, m1, ta, ma)+; vadd.vv v10,v8,v10 #avl=2, #vtype=(e64, m1, ta, ma)+; vle64.v v11,[const(1)] #avl=2, #vtype=(e64, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=2, #vtype=(e64, m1, ta, ma)+; vle64.v v11,[const(0)] #avl=2, #vtype=(e64, m1, ta, ma)+; vadd.vv v10,v10,v11 #avl=2, #vtype=(e64, m1, ta, ma)+; vadd.vi v10,v10,1 #avl=2, #vtype=(e64, m1, ta, ma)+; vse8.v v10,0(a0) #avl=16, #vtype=(e8, m1, ta, ma)+; ld ra,8(sp)+; ld fp,0(sp)+; addi sp,sp,16+; ret+;+; Disassembled:+; block0: ; offset 0x0+; addi sp, sp, -0x10+; sd ra, 8(sp)+; sd s0, 0(sp)+; mv s0, sp+; block1: ; offset 0x10+; .byte 0x57, 0x70, 0x08, 0xcc+; addi t6, sp, 0x10+; .byte 0x07, 0x84, 0x0f, 0x02+; .byte 0x57, 0x70, 0x81, 0xcd+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x07, 0xf5, 0x0f, 0x02+; .byte 0x57, 0x05, 0x85, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0xf5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; auipc t6, 0+; addi t6, t6, 0x50+; .byte 0x87, 0xf5, 0x0f, 0x02+; .byte 0x57, 0x85, 0xa5, 0x02+; .byte 0x57, 0xb5, 0xa0, 0x02+; .byte 0x57, 0x70, 0x08, 0xcc+; .byte 0x27, 0x05, 0x05, 0x02+; ld ra, 8(sp)+; ld s0, 0(sp)+; addi sp, sp, 0x10+; ret+; .byte 0x00, 0x00, 0x00, 0x00+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x01, 0x01, 0x01+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x01, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+; .byte 0x01, 0x00, 0x00, 0x00+diff --git a/crates/test-util/src/wast.rs b/crates/test-util/src/wast.rsindex dc570ce435..6992ecccbb 100644--- a/crates/test-util/src/wast.rs+++ b/crates/test-util/src/wast.rs@@ -637,6 +637,7 @@ impl WastTest {"misc_testsuite/simd/sse-cannot-fold-unaligned-loads.wast","misc_testsuite/winch/issue-10331.wast","misc_testsuite/winch/replace_lane.wast",+ "misc_testsuite/simd/riscv64-replicated-imm5-works.wast","spec_testsuite/simd_align.wast","spec_testsuite/simd_boolean.wast","spec_testsuite/simd_conversions.wast",diff --git a/tests/misc_testsuite/simd/riscv64-replicated-imm5-works.wast b/tests/misc_testsuite/simd/riscv64-replicated-imm5-works.wastnew file mode 100644index 0000000000..bed4ea723e--- /dev/null+++ b/tests/misc_testsuite/simd/riscv64-replicated-imm5-works.wast@@ -0,0 +1,12 @@+;;! simd = true++(module+ (memory (export "mem") 1)+ (func (export "add_and_extract") (param v128) (result i32)+ (i32x4.extract_lane 0+ (i32x4.add+ (local.get 0)+ (i32x4.splat (i32.const 0x01010101))))))++(assert_return (invoke "add_and_extract" (v128.const i32x4 0 0 0 0))+ (i32.const 0x01010101))