More fixes

Author: Amanieu

crates/core_arch/src/aarch64/neon/generated.rs

@@ -4455,7 +4455,7 @@ pub unsafe fn vnegq_s64(a: int64x2_t) -> int64x2_t {
 #[target_feature(enable = "neon")]
 #[cfg_attr(test, assert_instr(neg))]
 pub unsafe fn vnegd_s64(a: i64) -> i64 {
-    -a
+    a.wrapping_neg()
 }
 
 /// Negate
@@ -7512,31 +7512,31 @@ pub unsafe fn vsubq_f64(a: float64x2_t, b: float64x2_t) -> float64x2_t {
 #[target_feature(enable = "neon")]
 #[cfg_attr(test, assert_instr(nop))]
 pub unsafe fn vsubd_s64(a: i64, b: i64) -> i64 {
-    a - b
+    a.wrapping_sub(b)
 }
 
 /// Subtract
 #[inline]
 #[target_feature(enable = "neon")]
 #[cfg_attr(test, assert_instr(nop))]
 pub unsafe fn vsubd_u64(a: u64, b: u64) -> u64 {
-    a - b
+    a.wrapping_sub(b)
 }
 
 /// Add
 #[inline]
 #[target_feature(enable = "neon")]
 #[cfg_attr(test, assert_instr(nop))]
 pub unsafe fn vaddd_s64(a: i64, b: i64) -> i64 {
-    a + b
+    a.wrapping_add(b)
 }
 
 /// Add
 #[inline]
 #[target_feature(enable = "neon")]
 #[cfg_attr(test, assert_instr(nop))]
 pub unsafe fn vaddd_u64(a: u64, b: u64) -> u64 {
-    a + b
+    a.wrapping_add(b)
 }
 
 /// Floating-point add across vector
@@ -11536,7 +11536,7 @@ pub unsafe fn vrshrn_high_n_u64<const N: i32>(a: uint32x2_t, b: uint64x2_t) -> u
 pub unsafe fn vrsrad_n_s64<const N: i32>(a: i64, b: i64) -> i64 {
     static_assert!(N : i32 where N >= 1 && N <= 64);
     let b: i64 = vrshrd_n_s64::<N>(b);
-    a + b
+    a.wrapping_add(b)
 }
 
 /// Ungisned rounding shift right and accumulate.
@@ -11547,7 +11547,7 @@ pub unsafe fn vrsrad_n_s64<const N: i32>(a: i64, b: i64) -> i64 {
 pub unsafe fn vrsrad_n_u64<const N: i32>(a: u64, b: u64) -> u64 {
     static_assert!(N : i32 where N >= 1 && N <= 64);
     let b: u64 = vrshrd_n_u64::<N>(b);
-    a + b
+    a.wrapping_add(b)
 }
 
 /// Rounding subtract returning high narrow

crates/core_arch/src/aarch64/neon/mod.rs

@@ -2814,7 +2814,7 @@ pub unsafe fn vshrd_n_u64<const N: i32>(a: u64) -> u64 {
 #[rustc_legacy_const_generics(2)]
 pub unsafe fn vsrad_n_s64<const N: i32>(a: i64, b: i64) -> i64 {
     static_assert!(N : i32 where N >= 1 && N <= 64);
-    a + vshrd_n_s64::<N>(b)
+    a.wrapping_add(vshrd_n_s64::<N>(b))
 }
 
 /// Unsigned shift right and accumulate
@@ -2824,7 +2824,7 @@ pub unsafe fn vsrad_n_s64<const N: i32>(a: i64, b: i64) -> i64 {
 #[rustc_legacy_const_generics(2)]
 pub unsafe fn vsrad_n_u64<const N: i32>(a: u64, b: u64) -> u64 {
     static_assert!(N : i32 where N >= 1 && N <= 64);
-    a + vshrd_n_u64::<N>(b)
+    a.wrapping_add(vshrd_n_u64::<N>(b))
 }
 
 /// Shift Left and Insert (immediate)

crates/intrinsic-test/missing_aarch64.txt

@@ -108,3 +108,26 @@ vqshluq_n_s32
 vqshluq_n_s64
 vqshluq_n_s8
 vqshlus_n_s32
+
+# These tests produce a different result from C but only in debug builds of
+# stdarch. This likely both a bug in stdarch (expanding to a different LLVM
+# intrinsic) and a bug in LLVM (incorrect optimization changing the behavior of
+# integer operations).
+vqrdmlah_lane_s16
+vqrdmlah_lane_s32
+vqrdmlah_laneq_s16
+vqrdmlah_laneq_s32
+vqrdmlah_s16
+vqrdmlah_s32
+vqrdmlahh_lane_s16
+vqrdmlahh_laneq_s16
+vqrdmlahh_s16
+vqrdmlahq_lane_s16
+vqrdmlahq_lane_s32
+vqrdmlahq_laneq_s16
+vqrdmlahq_laneq_s32
+vqrdmlahq_s16
+vqrdmlahq_s32
+vqrdmlahs_lane_s32
+vqrdmlahs_laneq_s32
+vqrdmlahs_s32

crates/intrinsic-test/src/main.rs

@@ -168,7 +168,7 @@ fn compile_c(c_filename: &str, intrinsic: &Intrinsic, compiler: &str, a32: bool)
     let output = Command::new("sh")
         .arg("-c")
         .arg(format!(
-            "{cpp} {cppflags} {arch_flags} -Wno-narrowing -O0 -target {target} -o c_programs/{intrinsic} {filename}",
+            "{cpp} {cppflags} {arch_flags} -Wno-narrowing -O2 -target {target} -o c_programs/{intrinsic} {filename}",
             target = if a32 { "armv7-unknown-linux-gnueabihf" } else { "aarch64-unknown-linux-gnu" },
             arch_flags = if a32 { "-march=armv8.6-a+crypto+crc+dotprod" } else { "-march=armv8.6-a+crypto+sha3+crc+dotprod" },
             filename = c_filename,

crates/stdarch-gen/neon.spec

@@ -2058,7 +2058,7 @@ generate int*_t
 
 /// Negate
 name = vneg
-multi_fn = -a
+multi_fn = a.wrapping_neg()
 a = 1
 validate -1
 
@@ -4055,7 +4055,7 @@ generate float*_t
 
 /// Subtract
 name = vsub
-multi_fn = a - b
+multi_fn = a.wrapping_sub(b)
 a = 3
 b = 2
 validate 1
@@ -4065,7 +4065,7 @@ generate i64, u64
 
 /// Add
 name = vadd
-multi_fn = a + b
+multi_fn = a.wrapping_add(b)
 a = 1
 b = 2
 validate 3
@@ -6613,7 +6613,7 @@ n-suffix
 constn = N
 multi_fn = static_assert-N-1-bits
 multi_fn = vrshr-nself-::<N>, b:in_t, b
-multi_fn = a + b
+multi_fn = a.wrapping_add(b)
 a = 1
 b = 4
 n = 2
@@ -6628,7 +6628,7 @@ n-suffix
 constn = N
 multi_fn = static_assert-N-1-bits
 multi_fn = vrshr-nself-::<N>, b:in_t, b
-multi_fn = a + b
+multi_fn = a.wrapping_add(b)
 a = 1
 b = 4
 n = 2