A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
[Cranelift] folds comparison over bitwise-not operands (#11638)
Bongjun Jangc16e6fed8b4f
diff --git a/cranelift/codegen/src/opts/icmp.isle b/cranelift/codegen/src/opts/icmp.isleindex 0714a52717..ceeb5adfb1 100644--- a/cranelift/codegen/src/opts/icmp.isle+++ b/cranelift/codegen/src/opts/icmp.isle@@ -298,3 +298,7 @@;; (x < y) ? x : y != x == x > y, for both signed and unsigned.(rule (simplify (ne cty (select ty (slt cty x y) x y) x)) (sgt cty x y))(rule (simplify (ne cty (select ty (ult cty x y) x y) x)) (ugt cty x y))++(rule (simplify (ult cty (bnot ty x) (bnot ty y))) (ugt cty x y))+(rule (simplify (slt cty (bnot ty x) (bnot ty y))) (sgt cty x y))+diff --git a/cranelift/filetests/filetests/egraph/icmp-new.clif b/cranelift/filetests/filetests/egraph/icmp-new.clifnew file mode 100644index 0000000000..d90067f3eb--- /dev/null+++ b/cranelift/filetests/filetests/egraph/icmp-new.clif@@ -0,0 +1,32 @@+test optimize precise-output+set opt_level=speed+target x86_64++function %test1(i8, i8) -> i8 {+block0(v0: i8, v2: i8):+ v1 = bnot v0+ v3 = bnot v2+ v4 = icmp ult v1, v3+ return v4+}++; function %test1(i8, i8) -> i8 fast {+; block0(v0: i8, v2: i8):+; v5 = icmp ugt v0, v2+; return v5+; }++function %test1(i8, i8) -> i8 {+block0(v0: i8, v2: i8):+ v1 = bnot v0+ v3 = bnot v2+ v4 = icmp slt v1, v3+ return v4+}++; function %test1(i8, i8) -> i8 fast {+; block0(v0: i8, v2: i8):+; v5 = icmp sgt v0, v2+; return v5+; }+diff --git a/cranelift/filetests/filetests/runtests/icmp.clif b/cranelift/filetests/filetests/runtests/icmp.clifindex 2d280cb817..fbcae5e674 100644--- a/cranelift/filetests/filetests/runtests/icmp.clif+++ b/cranelift/filetests/filetests/runtests/icmp.clif@@ -44,3 +44,24 @@ block0(v0: i32, v1: i32):; run: %ule_sub_swap(1, -88) == 0; run: %ule_sub_swap(1, 2) == 0; run: %ule_sub_swap(2, 1) == 1++function %fold_cmp_over_nots1(i8, i8) -> i8 {+block0(v0: i8, v2: i8):+ v1 = bnot v0+ v3 = bnot v2+ v4 = icmp ult v1, v3+ return v4+}++; run: %fold_cmp_over_nots1(2, 1) == 1++function %fold_cmp_over_nots2(i8, i8) -> i8 {+block0(v0: i8, v2: i8):+ v1 = bnot v0+ v3 = bnot v2+ v4 = icmp slt v1, v3+ return v4+}++; run: %fold_cmp_over_nots2(2, 1) == 1+