[AArch64] Try to fold uaddlv and uaddlp

[jaykang10]

Summary

gcc generates less instructions than llvm from below intrinsic example.

#include <arm_neon.h>

unsigned foo(uint16x8_t b) {
    return vaddlvq_u32(vpadalq_u16(vdupq_n_u32(0), b));
}

gcc output

foo:
	uaddlv	s31, v0.8h
	fmov	x0, d31
	ret

llvm output

foo:
	uaddlp	v0.4s, v0.8h
	uaddlv	d0, v0.4s
	fmov	x0, d0

We could do uaddlv(uaddlp(x)) ==> uaddlv(x).
After adding tablegen pattern for it, the llvm output is as below.

foo:
	uaddlv	s0, v0.8h
	fmov	x0, d0

[dmgreen]

llvm/lib/Target/AArch64/AArch64InstrInfo.td

6335

It is probably quite a minor point, but can you change this to a (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv8i16v V128:$op), ssub)). The EXTRACT_SUBREG is using the fact that the higher lanes will be implicitly zeroed.

[jaykang10]

llvm/lib/Target/AArch64/AArch64InstrInfo.td

6335

Let me update the pattern.

[jaykang10]

Following @dmgreen's comment, updated the pattern.

[dmgreen]

llvm/lib/Target/AArch64/AArch64InstrInfo.td

6340

This one too, as it returns a h reg.

[jaykang10]

llvm/lib/Target/AArch64/AArch64InstrInfo.td

6340

Sorry. Let me update the pattern.

[jaykang10]

Following @dmgreen's comment, updated pattern.

[dmgreen]

Thanks. LGTM

llvm/lib/Target/AArch64/AArch64InstrInfo.td

def : Pat<(v2i32 (AArch64uaddv (v2i32 (addlp (v4i16 V64:$op))))),
             (INSERT_SUBREG (v2i32 (IMPLICIT_DEF)), (!cast<Instruction>(Opc#"v4i16v") V64:$op), ssub)>;
   def : Pat<(v2i64 (AArch64uaddv (v2i64 (addlp (v4i32 V128:$op))))),
             (INSERT_SUBREG (v2i64 (IMPLICIT_DEF)), (!cast<Instruction>(Opc#"v4i32v") V128:$op), dsub)>;
 }
 
 defm : SIMDAcrossLaneLongPairIntrinsic<"UADDLV", AArch64uaddlp>;
 defm : SIMDAcrossLaneLongPairIntrinsic<"SADDLV", AArch64saddlp>;
 
+// Patterns for uaddlv(uaddlp(x)) ==> uaddlv
+def : Pat<(i64 (int_aarch64_neon_uaddlv (v4i32 (AArch64uaddlp (v8i16 V128:$op))))),
+          (i64 (EXTRACT_SUBREG
+            (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv8i16v V128:$op), ssub)),
+            dsub))>;
+
+def : Pat<(i32 (int_aarch64_neon_uaddlv (v8i16 (AArch64uaddlp (v16i8 V128:$op))))),
+          (i32 (EXTRACT_SUBREG
+            (v8i16 (SUBREG_TO_REG (i64 0), (UADDLVv16i8v V128:$op), hsub)),
+            ssub))>;
+
 // Patterns for across-vector intrinsics, that have a node equivalent, that
 // returns a vector (with only the low lane defined) instead of a scalar.
 // In effect, opNode is the same as (scalar_to_vector (IntNode)).
 multiclass SIMDAcrossLanesIntrinsic<string baseOpc,
                                     SDPatternOperator opNode> {
 // If a lane instruction caught the vector_extract around opNode, we can
 // directly match the latter to the instruction.
 def : Pat<(v8i8 (opNode V64:$Rn)),

llvm/test/CodeGen/AArch64/uaddlv-vaddlp-combine.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -mtriple aarch64-none-linux-gnu < %s | FileCheck %s
+
+define i32 @uaddlv_uaddlp_v8i16(<8 x i16> %0) {
+; CHECK-LABEL: uaddlv_uaddlp_v8i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uaddlv s0, v0.8h
+; CHECK-NEXT:    fmov x0, d0
+; CHECK-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; CHECK-NEXT:    ret
+  %2 = tail call <4 x i32> @llvm.aarch64.neon.uaddlp.v4i32.v8i16(<8 x i16> %0)
+  %3 = tail call i64 @llvm.aarch64.neon.uaddlv.i64.v4i32(<4 x i32> %2)
+  %4 = trunc i64 %3 to i32
+  ret i32 %4
+}
+
+define i16 @uaddlv_uaddlp_v16i8(<16 x i8> %0) {
+; CHECK-LABEL: uaddlv_uaddlp_v16i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uaddlv h0, v0.16b
+; CHECK-NEXT:    fmov w0, s0
+; CHECK-NEXT:    ret
+  %2 = tail call <8 x i16> @llvm.aarch64.neon.uaddlp.v8i16.v16i8(<16 x i8> %0)
+  %3 = tail call i32 @llvm.aarch64.neon.uaddlv.i32.v8i16(<8 x i16> %2)
+  %4 = trunc i32 %3 to i16
+  ret i16 %4
+}
+
+declare i64 @llvm.aarch64.neon.uaddlv.i64.v4i32(<4 x i32>)
+declare i32 @llvm.aarch64.neon.uaddlv.i32.v8i16(<8 x i16>)
+declare <4 x i32> @llvm.aarch64.neon.uaddlp.v4i32.v8i16(<8 x i16>)
+declare <8 x i16> @llvm.aarch64.neon.uaddlp.v8i16.v16i8(<16 x i8>)