Skip to content

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

[Cranelift] `(X | Y) + (X & Y) --> X + Y`, `(X & Y) + (X ^ Y) --> X | Y` (#11647)

Bongjun Jangd23f55ad8756
diff --git a/cranelift/codegen/src/opts/bitops.isle b/cranelift/codegen/src/opts/bitops.isle
index 9af24741c4..cf51c4da8b 100644
--- a/cranelift/codegen/src/opts/bitops.isle
+++ b/cranelift/codegen/src/opts/bitops.isle
@@ -204,3 +204,12 @@
(rule (simplify (band ty y (band ty x y))) (band ty x y))
(rule (simplify (band ty (bxor ty x (bnot ty y)) x)) (band ty x y))
+
+; (x & y) + (x ^ y) --> x | y
+(rule (simplify (iadd ty (band ty x y) (bxor ty x y))) (bor ty x y))
+(rule (simplify (iadd ty (bxor ty x y) (band ty x y))) (bor ty x y))
+
+; (x | y) + (x & y) --> x + y
+(rule (simplify (iadd ty (bor ty x y) (band ty x y))) (iadd ty x y))
+(rule (simplify (iadd ty (band ty x y) (bor ty x y))) (iadd ty x y))
+
diff --git a/cranelift/filetests/filetests/egraph/fold-add-and-xor.clif b/cranelift/filetests/filetests/egraph/fold-add-and-xor.clif
new file mode 100644
index 0000000000..d7639549e2
--- /dev/null
+++ b/cranelift/filetests/filetests/egraph/fold-add-and-xor.clif
@@ -0,0 +1,81 @@
+test optimize precise-output
+set opt_level=speed
+target x86_64
+
+function %fold_add_and_xor1(i8, i8) -> i8 {
+block0(v0: i8, v1: i8):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_and_xor1(i8, i8) -> i8 fast {
+; block0(v0: i8, v1: i8):
+; v5 = bor v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_and_xor2(i16, i16) -> i16 {
+block0(v0: i16, v1: i16):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_and_xor2(i16, i16) -> i16 fast {
+; block0(v0: i16, v1: i16):
+; v5 = bor v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_and_xor3(i32, i32) -> i32 {
+block0(v0: i32, v1: i32):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_and_xor3(i32, i32) -> i32 fast {
+; block0(v0: i32, v1: i32):
+; v5 = bor v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_and_xor4(i64, i64) -> i64 {
+block0(v0: i64, v1: i64):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_and_xor4(i64, i64) -> i64 fast {
+; block0(v0: i64, v1: i64):
+; v5 = bor v0, v1
+; return v5
+; }
+
+
+function %fold_add_and_xor4(i64, i64) -> i64 {
+block0(v0: i64, v1: i64):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v3, v2
+ return v4
+}
+
+; function %fold_add_and_xor4(i64, i64) -> i64 fast {
+; block0(v0: i64, v1: i64):
+; v5 = bor v0, v1
+; return v5
+; }
+
diff --git a/cranelift/filetests/filetests/egraph/fold-add-or-and.clif b/cranelift/filetests/filetests/egraph/fold-add-or-and.clif
new file mode 100644
index 0000000000..ef37358d1f
--- /dev/null
+++ b/cranelift/filetests/filetests/egraph/fold-add-or-and.clif
@@ -0,0 +1,82 @@
+test optimize precise-output
+set opt_level=speed
+target x86_64
+
+function %fold_add_or_and1(i8, i8) -> i8 {
+block0(v0: i8, v1: i8):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_or_and1(i8, i8) -> i8 fast {
+; block0(v0: i8, v1: i8):
+; v5 = iadd v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_or_and2(i16, i16) -> i16 {
+block0(v0: i16, v1: i16):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_or_and2(i16, i16) -> i16 fast {
+; block0(v0: i16, v1: i16):
+; v5 = iadd v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_or_and3(i32, i32) -> i32 {
+block0(v0: i32, v1: i32):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_or_and3(i32, i32) -> i32 fast {
+; block0(v0: i32, v1: i32):
+; v5 = iadd v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_or_and4(i64, i64) -> i64 {
+block0(v0: i64, v1: i64):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; function %fold_add_or_and4(i64, i64) -> i64 fast {
+; block0(v0: i64, v1: i64):
+; v5 = iadd v0, v1
+; return v5
+; }
+
+
+
+function %fold_add_or_and4(i64, i64) -> i64 {
+block0(v0: i64, v1: i64):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v3, v2
+ return v4
+}
+
+; function %fold_add_or_and4(i64, i64) -> i64 fast {
+; block0(v0: i64, v1: i64):
+; v5 = iadd v0, v1
+; return v5
+; }
+
diff --git a/cranelift/filetests/filetests/runtests/bitops.clif b/cranelift/filetests/filetests/runtests/bitops.clif
index 31d30b028c..fd89e9efc1 100644
--- a/cranelift/filetests/filetests/runtests/bitops.clif
+++ b/cranelift/filetests/filetests/runtests/bitops.clif
@@ -174,3 +174,23 @@ block0(v0: i32, v1: i32):
; run: %fold_and_xor_not(4, 5) == 4
+function %fold_add_or_and3(i32, i32) -> i32 {
+block0(v0: i32, v1: i32):
+ v2 = bor v0, v1
+ v3 = band v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; run: %fold_add_or_and3(4, 5) == 9
+
+function %fold_add_and_xor4(i64, i64) -> i64 {
+block0(v0: i64, v1: i64):
+ v2 = band v0, v1
+ v3 = bxor v0, v1
+ v4 = iadd v2, v3
+ return v4
+}
+
+; run: %fold_add_and_xor4(1, 2) == 3
+