Skip to content

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

Fix AArch64 subword atomic unsigned min/max (#14029)

subotacbb4984627c4f
diff --git a/cranelift/codegen/src/isa/aarch64/inst/emit.rs b/cranelift/codegen/src/isa/aarch64/inst/emit.rs
index b10e412763..a603e8d6e0 100644
--- a/cranelift/codegen/src/isa/aarch64/inst/emit.rs
+++ b/cranelift/codegen/src/isa/aarch64/inst/emit.rs
@@ -1508,7 +1508,14 @@ impl MachInstEmit for Inst {
},
_ => None,
};
-
+ let zero_ext = match op {
+ AtomicRMWLoopOp::Umin | AtomicRMWLoopOp::Umax => match ty {
+ I16 => Some(ExtendOp::UXTH),
+ I8 => Some(ExtendOp::UXTB),
+ _ => None,
+ },
+ _ => None,
+ };
// sxt{b|h} the loaded result if necessary.
if sign_ext.is_some() {
let (_, from_bits) = sign_ext.unwrap();
@@ -1561,8 +1568,7 @@ impl MachInstEmit for Inst {
_ => unreachable!(),
};
- if sign_ext.is_some() {
- let (extendop, _) = sign_ext.unwrap();
+ if let Some(extendop) = sign_ext.map(|(op, _)| op).or(zero_ext) {
Inst::AluRRRExtend {
alu_op: ALUOp::SubS,
size,
diff --git a/cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs b/cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
index 930080fd4c..04d7e77def 100644
--- a/cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
+++ b/cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
@@ -7284,7 +7284,7 @@ fn test_aarch64_binemit() {
scratch1: writable_xreg(24),
scratch2: writable_xreg(28),
},
- "3BFF5F087F031A6B7C339A9A3CFF180898FFFFB5",
+ "3BFF5F087F033A6B7C339A9A3CFF180898FFFFB5",
"atomic_rmw_loop_umin_8 addr=x25 operand=x26 oldval=x27 scratch1=x24 scratch2=x28",
));
insns.push((
@@ -7298,7 +7298,7 @@ fn test_aarch64_binemit() {
scratch1: writable_xreg(24),
scratch2: writable_xreg(28),
},
- "3BFF5F487F031A6B7C839A9A3CFF184898FFFFB5",
+ "3BFF5F487F233A6B7C839A9A3CFF184898FFFFB5",
"atomic_rmw_loop_umax_16 addr=x25 operand=x26 oldval=x27 scratch1=x24 scratch2=x28",
));
diff --git a/cranelift/filetests/filetests/isa/aarch64/atomic-rmw.clif b/cranelift/filetests/filetests/isa/aarch64/atomic-rmw.clif
index d1c620f3cb..29cd380c0f 100644
--- a/cranelift/filetests/filetests/isa/aarch64/atomic-rmw.clif
+++ b/cranelift/filetests/filetests/isa/aarch64/atomic-rmw.clif
@@ -1306,7 +1306,7 @@ block0(v0: i64, v1: i16):
; mov x25, x0
; mov x26, x1
; ldaxrh w27, [x25] ; trap: heap_oob
-; cmp w27, w26
+; cmp w27, w26, uxth
; csel x28, x27, x26, hi
; stlxrh w24, w28, [x25] ; trap: heap_oob
; cbnz x24, #0x1c
@@ -1349,7 +1349,7 @@ block0(v0: i64, v1: i8):
; mov x25, x0
; mov x26, x1
; ldaxrb w27, [x25] ; trap: heap_oob
-; cmp w27, w26
+; cmp w27, w26, uxtb
; csel x28, x27, x26, hi
; stlxrb w24, w28, [x25] ; trap: heap_oob
; cbnz x24, #0x1c
@@ -1652,7 +1652,7 @@ block0(v0: i64, v1: i16):
; mov x25, x0
; mov x26, x1
; ldaxrh w27, [x25] ; trap: heap_oob
-; cmp w27, w26
+; cmp w27, w26, uxth
; csel x28, x27, x26, lo
; stlxrh w24, w28, [x25] ; trap: heap_oob
; cbnz x24, #0x1c
@@ -1695,7 +1695,7 @@ block0(v0: i64, v1: i8):
; mov x25, x0
; mov x26, x1
; ldaxrb w27, [x25] ; trap: heap_oob
-; cmp w27, w26
+; cmp w27, w26, uxtb
; csel x28, x27, x26, lo
; stlxrb w24, w28, [x25] ; trap: heap_oob
; cbnz x24, #0x1c
diff --git a/cranelift/filetests/filetests/runtests/atomic-rmw-subword-little.clif b/cranelift/filetests/filetests/runtests/atomic-rmw-subword-little.clif
index 391e6851cb..6bff6b3faf 100644
--- a/cranelift/filetests/filetests/runtests/atomic-rmw-subword-little.clif
+++ b/cranelift/filetests/filetests/runtests/atomic-rmw-subword-little.clif
@@ -900,3 +900,31 @@ block0(v0: i32, v1: i64, v2: i8):
; run: %atomic_rmw_xchg_little_i8_no_res(0x12345678, 1, 0xff) == 0x1234ff78
; run: %atomic_rmw_xchg_little_i8_no_res(0x12345678, 0, 0x11) == 0x12345611
; run: %atomic_rmw_xchg_little_i8_no_res(0x12345678, 0, 0xff) == 0x123456ff
+
+function %atomic_rmw_umin_i8_reduced_operand(i32) -> i8 {
+ ss0 = explicit_slot 1
+
+block0(v0: i32):
+ v1 = stack_addr.i64 ss0
+ v2 = iconst.i8 32
+ store.i8 v2, v1
+ v3 = ireduce.i8 v0
+ v4 = atomic_rmw.i8 umin v1, v3
+ v5 = load.i8 v1
+ return v5
+}
+; run: %atomic_rmw_umin_i8_reduced_operand(0x110) == 16
+
+function %atomic_rmw_umax_i8_reduced_operand(i32) -> i8 {
+ ss0 = explicit_slot 1
+
+block0(v0: i32):
+ v1 = stack_addr.i64 ss0
+ v2 = iconst.i8 32
+ store.i8 v2, v1
+ v3 = ireduce.i8 v0
+ v4 = atomic_rmw.i8 umax v1, v3
+ v5 = load.i8 v1
+ return v5
+}
+; run: %atomic_rmw_umax_i8_reduced_operand(0x110) == 32