[clang] [Clang][AArch64] Lower NEON vaddv/vminv/vmaxv builtins to llvm.vector.reduce intrinsics. (PR #165400)

[David Sherwood via cfe-commits]

================
@@ -1243,27 +1251,43 @@ static const ARMVectorIntrinsicInfo AArch64SISDIntrinsicMap[] = {
   NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
   NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
   NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
-  NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
-  NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
+  NEONMAP1(vmaxv_s16, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxv_s32, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxv_s8, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxv_u16, vector_reduce_umax, Add1ArgType),
+  NEONMAP1(vmaxv_u32, vector_reduce_umax, Add1ArgType),
+  NEONMAP1(vmaxv_u8, vector_reduce_umax, Add1ArgType),
   NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
   NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
-  NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
-  NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
+  NEONMAP1(vmaxvq_s16, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxvq_s32, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxvq_s8, vector_reduce_smax, Add1ArgType),
+  NEONMAP1(vmaxvq_u16, vector_reduce_umax, Add1ArgType),
+  NEONMAP1(vmaxvq_u32, vector_reduce_umax, Add1ArgType),
+  NEONMAP1(vmaxvq_u8, vector_reduce_umax, Add1ArgType),
   NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
   NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
   NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
   NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
-  NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
-  NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
+  NEONMAP1(vminv_s16, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminv_s32, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminv_s8, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminv_u16, vector_reduce_umin, Add1ArgType),
+  NEONMAP1(vminv_u32, vector_reduce_umin, Add1ArgType),
+  NEONMAP1(vminv_u8, vector_reduce_umin, Add1ArgType),
   NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
   NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
-  NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
-  NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
+  NEONMAP1(vminvq_s16, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminvq_s32, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminvq_s8, vector_reduce_smin, Add1ArgType),
+  NEONMAP1(vminvq_u16, vector_reduce_umin, Add1ArgType),
+  NEONMAP1(vminvq_u32, vector_reduce_umin, Add1ArgType),
+  NEONMAP1(vminvq_u8, vector_reduce_umin, Add1ArgType),
----------------
david-arm wrote:

Just an observation, but when SVE is available we can also use the SVE umaxv,uminv,etc. for 64-bit elements, which NEON seems to be missing.

https://github.com/llvm/llvm-project/pull/165400