A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
Fix subtle ISLE extractor issue leading to incorrect matching in `*mul_overflow` lowering. (#14295)
Chris Fallinad985bc64f65
diff --git a/cranelift/codegen/src/isa/aarch64/inst.isle b/cranelift/codegen/src/isa/aarch64/inst.isleindex b91d765700..281ab5d80a 100644--- a/cranelift/codegen/src/isa/aarch64/inst.isle+++ b/cranelift/codegen/src/isa/aarch64/inst.isle@@ -4992,7 +4992,7 @@;; give an opportunistic def of the other output.(rule 4(is_nonzero- (second_result uadd @ (uadd_overflow (ty_32_or_64 ty) x y)))+ (is_second_result uadd @ (uadd_overflow (ty_32_or_64 ty) x y)))(if-let (first_result sum_value) uadd)(let ((producer ProducesFlags (alu_rrr_with_flags_paired ty x y (ALUOp.AddS))))(CondResult.Cond@@ -5012,7 +5012,7 @@;; cmp out, out, uxt{b,h}(rule 5(is_nonzero- (second_result umul @ (umul_overflow (fits_in_16 ty) a b)))+ (is_second_result umul @ (umul_overflow (fits_in_16 ty) a b)))(if-let (first_result prod_value) umul)(let ((a_uext Reg (put_in_reg_zext32 a))(b_uext Reg (put_in_reg_zext32 b))@@ -5031,7 +5031,7 @@;; cmp out, out, uxtw(rule 6(is_nonzero- (second_result umul @ (umul_overflow $I32 a b)))+ (is_second_result umul @ (umul_overflow $I32 a b)))(if-let (first_result prod_value) umul)(let ((dst WritableReg (temp_writable_reg $I64))(producer ProducesFlags@@ -5048,7 +5048,7 @@;; cmp tmp, #0(rule 7(is_nonzero- (second_result umul @ (umul_overflow $I64 a b)))+ (is_second_result umul @ (umul_overflow $I64 a b)))(let ((tmp Reg (umulh $I64 a b))(producer ProducesFlags (cmp64_imm tmp (u8_into_imm12 0))))(CondResult.Cond producer (Cond.Ne))))@@ -5057,7 +5057,7 @@;; cmp out, out, sxt{b,h}(rule 8(is_nonzero- (second_result smul @ (smul_overflow (fits_in_16 ty) a b)))+ (is_second_result smul @ (smul_overflow (fits_in_16 ty) a b)))(if-let (first_result prod_value) smul)(let ((a_sext Reg (put_in_reg_sext32 a))(b_sext Reg (put_in_reg_sext32 b))@@ -5076,7 +5076,7 @@;; cmp out, out, sxtw(rule 9(is_nonzero- (second_result smul @ (smul_overflow $I32 a b)))+ (is_second_result smul @ (smul_overflow $I32 a b)))(if-let (first_result prod_value) smul)(let ((dst WritableReg (temp_writable_reg $I64))(producer ProducesFlags@@ -5093,7 +5093,7 @@;; cmp tmp, out, asr #63(rule 10(is_nonzero- (second_result smul @ (smul_overflow $I64 a b)))+ (is_second_result smul @ (smul_overflow $I64 a b)))(if-let (first_result prod_value) smul)(let ((prod Reg (put_in_reg prod_value))(tmp Reg (smulh $I64 a b))diff --git a/cranelift/codegen/src/isa/x64/inst.isle b/cranelift/codegen/src/isa/x64/inst.isleindex 62d14633ac..6c1e9d1d7c 100644--- a/cranelift/codegen/src/isa/x64/inst.isle+++ b/cranelift/codegen/src/isa/x64/inst.isle@@ -3882,7 +3882,7 @@;; opportunistic def of the other output.(rule 3(is_nonzero- (second_result uadd @ (uadd_overflow (fits_in_64 ty) x y)))+ (is_second_result uadd @ (uadd_overflow (fits_in_64 ty) x y)))(if-let (first_result sum_value) uadd)(let ((producer ProducesFlags (x64_add_with_flags_paired ty x y)))(CondResult.CC@@ -3895,7 +3895,7 @@;; `mul` of narrow types), so no extra compare is needed.(rule 4(is_nonzero- (second_result umul @ (umul_overflow _ x y @ (value_type (ty_int_ref_16_to_64 ty)))))+ (is_second_result umul @ (umul_overflow _ x y @ (value_type (ty_int_ref_16_to_64 ty)))))(if-let (first_result prod_value) umul)(let ((producer ProducesFlags (x64_mul_lo_with_flags_paired ty false x y)))(CondResult.CC@@ -3904,7 +3904,7 @@(rule 5(is_nonzero- (second_result umul @ (umul_overflow _ x y @ (value_type $I8))))+ (is_second_result umul @ (umul_overflow _ x y @ (value_type $I8))))(if-let (first_result prod_value) umul)(let ((producer ProducesFlags (x64_mul8_with_flags_paired false x y)))(CondResult.CC@@ -3913,7 +3913,7 @@(rule 6(is_nonzero- (second_result smul @ (smul_overflow _ x y @ (value_type (ty_int_ref_16_to_64 ty)))))+ (is_second_result smul @ (smul_overflow _ x y @ (value_type (ty_int_ref_16_to_64 ty)))))(if-let (first_result prod_value) smul)(let ((producer ProducesFlags (x64_mul_lo_with_flags_paired ty true x y)))(CondResult.CC@@ -3922,7 +3922,7 @@(rule 7(is_nonzero- (second_result smul @ (smul_overflow _ x y @ (value_type $I8))))+ (is_second_result smul @ (smul_overflow _ x y @ (value_type $I8))))(if-let (first_result prod_value) smul)(let ((producer ProducesFlags (x64_mul8_with_flags_paired true x y)))(CondResult.CCdiff --git a/cranelift/codegen/src/machinst/isle.rs b/cranelift/codegen/src/machinst/isle.rsindex 0d4288bd6d..a8c81dd3cb 100644--- a/cranelift/codegen/src/machinst/isle.rs+++ b/cranelift/codegen/src/machinst/isle.rs@@ -213,6 +213,20 @@ macro_rules! isle_lower_prelude_methods {.copied()}+ #[inline]+ fn is_second_result(&mut self, val: Value) -> Option<Value> {+ let inst = self.def_inst(val)?;+ let is_match = self+ .lower_ctx+ .dfg()+ .inst_results(inst)+ .iter()+ .skip(1)+ .next()+ == Some(&val);+ if is_match { Some(val) } else { None }+ }+#[inline]fn second_result_used(&mut self, inst: Inst) -> bool {let second_result = self.lower_ctx.dfg().inst_results(inst).get(1).copied();diff --git a/cranelift/codegen/src/prelude_lower.isle b/cranelift/codegen/src/prelude_lower.isleindex 46ce41d47a..3e0d88672d 100644--- a/cranelift/codegen/src/prelude_lower.isle+++ b/cranelift/codegen/src/prelude_lower.isle@@ -301,10 +301,23 @@(extern extractor first_result first_result);; Extract the second result value of the given instruction.+;;+;; Be careful using this: if used to match on a *`Value`*, the+;; auto-converter will take that `Value` to its defining `Inst`, then+;; the inner result of this extractor will be the second def of the+;; value; in other words, it will *fetch* the second result but will+;; not *assert/only match if* the initial `Value` is the second+;; result. Use `is_second_result` for that.(spec (second_result value) (provide (= result value)))(decl second_result (Value) Inst)(extern extractor second_result second_result)+;; Determine whether the given value is the second result of its+;; defining instruction.+(spec (is_second_result value) (provide (= result value)))+(decl is_second_result (Value) Value)+(extern extractor is_second_result is_second_result)+;; Returns whether an instruction's second result still has uses to serve in;; the lowering scan. Overflow operations use this while lowering their first;; (non-boolean) result.diff --git a/cranelift/filetests/filetests/runtests/issue-14293.clif b/cranelift/filetests/filetests/runtests/issue-14293.clifnew file mode 100644index 0000000000..54ed9aca90--- /dev/null+++ b/cranelift/filetests/filetests/runtests/issue-14293.clif@@ -0,0 +1,15 @@+test run+set opt_level=speed_and_size+target x86_64+target aarch64++function %a(i64) -> i8 {+block0(v0: i64):+ v1, v2 = smul_overflow.i64 v0, v0+ v3 = iconst.i64 0+ v4 = icmp eq v1, v3+ return v4+}++; run: %a(0) == 1+; run: %a(8192) == 0