[llvm] [AArch64][Codegen]Transform saturating smull to sqdmulh (PR #143671)

[David Green via llvm-commits]

================
@@ -0,0 +1,58 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64-none-elf < %s | FileCheck %s
+
+define <8 x i16> @saturating_int16(<8 x i16> %a, <8 x i16> %b) {
+; CHECK-LABEL: saturating_int16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sqdmulh v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    ret
+  %as = sext <8 x i16> %a to <8 x i32>
+  %bs = sext <8 x i16> %b to <8 x i32>
+  %m = mul <8 x i32> %bs, %as
+  %sh = ashr <8 x i32> %m, splat (i32 15)
+  %ma = tail call <8 x i32> @llvm.smin.v8i32(<8 x i32> %sh, <8 x i32> splat (i32 32767))
+  %t = trunc <8 x i32> %ma to <8 x i16>
+  ret <8 x i16> %t
+}
+
+define <4 x i32> @saturating_int32(<4 x i32> %a, <4 x i32> %b) {
+; CHECK-LABEL: saturating_int32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sqdmulh v0.4s, v1.4s, v0.4s
+; CHECK-NEXT:    ret
+  %as = sext <4 x i32> %a to <4 x i64>
+  %bs = sext <4 x i32> %b to <4 x i64>
+  %m = mul <4 x i64> %bs, %as
+  %sh = ashr <4 x i64> %m, splat (i64 31)
+  %ma = tail call <4 x i64> @llvm.smin.v8i32(<4 x i64> %sh, <4 x i64> splat (i64 2147483647))
+  %t = trunc <4 x i64> %ma to <4 x i32>
+  ret <4 x i32> %t
+}
+
+define <8 x i16> @saturating_q15(<8 x i16> %a, <8 x i16> %b) {
+; CHECK-LABEL: saturating_q15:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sqdmulh v0.8h, v1.8h, v0.8h
----------------
davemgreen wrote:

I'm not sure if this is valid on it's own, as it would saturate at a different point.

https://github.com/llvm/llvm-project/pull/143671