Skip to content

A lightweight WebAssembly runtime that is fast, secure, and standards-compliant

Fix a bug that caused float bitwise operations to fail on aarch64. (#11811)

Masashi Yoshimura3859a683f5e7
diff --git a/cranelift/codegen/src/isa/aarch64/inst.isle b/cranelift/codegen/src/isa/aarch64/inst.isle
index 2d1eb8f1b2..1c7bc76fdd 100644
--- a/cranelift/codegen/src/isa/aarch64/inst.isle
+++ b/cranelift/codegen/src/isa/aarch64/inst.isle
@@ -1413,6 +1413,12 @@
(rule (vector_size (dynamic_lane 32 4)) (VectorSize.Size32x4))
(rule (vector_size (dynamic_lane 64 2)) (VectorSize.Size64x2))
+;; Helper for converting the `ScalarSize` of a float value to the corresponding `VectorSize`
+(decl float_vector_size_in_64 (Type) VectorSize)
+(rule (float_vector_size_in_64 $F16) (VectorSize.Size16x4))
+(rule (float_vector_size_in_64 $F32) (VectorSize.Size32x2))
+(rule (float_vector_size_in_64 $F64) (VectorSize.Size8x8))
+
;; A floating-point unit (FPU) operation with one arg.
(type FPUOp1
(enum
diff --git a/cranelift/codegen/src/isa/aarch64/lower.isle b/cranelift/codegen/src/isa/aarch64/lower.isle
index c32edb2b1f..33c06c4c4a 100644
--- a/cranelift/codegen/src/isa/aarch64/lower.isle
+++ b/cranelift/codegen/src/isa/aarch64/lower.isle
@@ -1383,9 +1383,12 @@
;;;; Rules for `band` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule band_fits_in_64 -3 (lower (has_type (fits_in_64 ty) (band x y)))
+(rule band_fits_in_64 -4 (lower (has_type (fits_in_64 (ty_int ty)) (band x y)))
(alu_rs_imm_logic_commutative (ALUOp.And) ty x y))
+(rule -3 (lower (has_type (fits_in_64 (ty_scalar_float ty)) (band x y)))
+ (and_vec x y (float_vector_size_in_64 ty)))
+
;; Implementation of `band` for vector types.
(rule -2 (lower (has_type (ty_vec64 ty) (band x y)))
(and_vec x y (vector_size ty)))
@@ -1417,9 +1420,12 @@
;;;; Rules for `bor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule bor_fits_in_64 -3 (lower (has_type (fits_in_64 ty) (bor x y)))
+(rule bor_fits_in_64 -4 (lower (has_type (fits_in_64 (ty_int ty)) (bor x y)))
(alu_rs_imm_logic_commutative (ALUOp.Orr) ty x y))
+(rule -3 (lower (has_type (fits_in_64 (ty_scalar_float ty)) (bor x y)))
+ (orr_vec x y (float_vector_size_in_64 ty)))
+
;; Implementation of `bor` for vector types.
(rule -2 (lower (has_type (ty_vec64 ty) (bor x y)))
(orr_vec x y (vector_size ty)))
@@ -1472,9 +1478,12 @@
;;;; Rules for `bxor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-(rule bxor_fits_in_64 -3 (lower (has_type (fits_in_64 ty) (bxor x y)))
+(rule bxor_fits_in_64 -4 (lower (has_type (fits_in_64 (ty_int ty)) (bxor x y)))
(alu_rs_imm_logic_commutative (ALUOp.Eor) ty x y))
+(rule -3 (lower (has_type (fits_in_64 (ty_scalar_float ty)) (bxor x y)))
+ (eor_vec x y (float_vector_size_in_64 ty)))
+
;; Implementation of `bxor` for vector types.
(rule -2 (lower (has_type (ty_vec64 ty) (bxor x y)))
(eor_vec x y (vector_size ty)))
diff --git a/cranelift/filetests/filetests/isa/aarch64/floating-point-bitwise.clif b/cranelift/filetests/filetests/isa/aarch64/floating-point-bitwise.clif
new file mode 100644
index 0000000000..b0645e470f
--- /dev/null
+++ b/cranelift/filetests/filetests/isa/aarch64/floating-point-bitwise.clif
@@ -0,0 +1,93 @@
+test compile precise-output
+set unwind_info=false
+set preserve_frame_pointers=true
+target aarch64
+
+function %f1() -> f64 fast {
+
+block0:
+ v1 = f64const 0x1.11044b092d153p-1
+ v2 = bxor.f64 v1, v1
+ return v2
+}
+
+; VCode:
+; stp fp, lr, [sp, #-16]!
+; mov fp, sp
+; block0:
+; ldr d1, [const(0)]
+; eor v0.8b, v1.8b, v1.8b
+; ldp fp, lr, [sp], #16
+; ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+; stp x29, x30, [sp, #-0x10]!
+; mov x29, sp
+; block1: ; offset 0x8
+; ldr d1, #0x18
+; eor v0.8b, v1.8b, v1.8b
+; ldp x29, x30, [sp], #0x10
+; ret
+; adrp x19, #0xffffffff25a29000
+; .byte 0x44, 0x10, 0xe1, 0x3f
+
+function %f2() -> f32 fast {
+
+block0:
+ v1 = f32const 0x1.11044cp-1
+ v2 = bxor.f32 v1, v1
+ return v2
+}
+
+; VCode:
+; stp fp, lr, [sp, #-16]!
+; mov fp, sp
+; block0:
+; movz w2, #33318
+; movk w2, w2, #16136, LSL #16
+; fmov s5, w2
+; eor v0.8b, v5.8b, v5.8b
+; ldp fp, lr, [sp], #16
+; ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+; stp x29, x30, [sp, #-0x10]!
+; mov x29, sp
+; block1: ; offset 0x8
+; mov w2, #0x8226
+; movk w2, #0x3f08, lsl #16
+; fmov s5, w2
+; eor v0.8b, v5.8b, v5.8b
+; ldp x29, x30, [sp], #0x10
+; ret
+
+function %f3() -> f16 fast {
+
+block0:
+ v1 = f16const 0x1.11p-1
+ v2 = bxor.f16 v1, v1
+ return v2
+}
+
+; VCode:
+; stp fp, lr, [sp, #-16]!
+; mov fp, sp
+; block0:
+; movz w1, #14404
+; fmov s3, w1
+; eor v0.8b, v3.8b, v3.8b
+; ldp fp, lr, [sp], #16
+; ret
+;
+; Disassembled:
+; block0: ; offset 0x0
+; stp x29, x30, [sp, #-0x10]!
+; mov x29, sp
+; block1: ; offset 0x8
+; mov w1, #0x3844
+; fmov s3, w1
+; eor v0.8b, v3.8b, v3.8b
+; ldp x29, x30, [sp], #0x10
+; ret
\ No newline at end of file