llvm-project/llvm/test/CodeGen/AArch64/saturating-vec-smull.ll
Nashe Mncube de4b458aa5
[AArch64][Codegen]Transform saturating smull to sqdmulh (#143671)
This patch adds a pattern for recognizing saturating vector
smull. Prior to this patch these were performed using a
combination of smull+smull2+uzp+smin  like the following

```
        smull2  v5.2d, v1.4s, v2.4s
        smull   v1.2d, v1.2s, v2.2s
        uzp2    v1.4s, v1.4s, v5.4s
        smin    v1.4s, v1.4s, v0.4s
```

which now optimizes to

```
    sqdmulh v0.4s, v1.4s, v0.4s
```

This only operates on vectors containing int32 and int16 types
2025-07-16 09:50:04 +01:00

224 lines
8.0 KiB
LLVM

; 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 <2 x i16> @saturating_2xi16(<2 x i16> %a, <2 x i16> %b) {
; CHECK-LABEL: saturating_2xi16:
; CHECK: // %bb.0:
; CHECK-NEXT: shl v0.2s, v0.2s, #16
; CHECK-NEXT: shl v1.2s, v1.2s, #16
; CHECK-NEXT: sshr v0.2s, v0.2s, #16
; CHECK-NEXT: sshr v1.2s, v1.2s, #16
; CHECK-NEXT: sqdmulh v0.2s, v1.2s, v0.2s
; CHECK-NEXT: ret
%as = sext <2 x i16> %a to <2 x i32>
%bs = sext <2 x i16> %b to <2 x i32>
%m = mul <2 x i32> %bs, %as
%sh = ashr <2 x i32> %m, splat (i32 15)
%ma = tail call <2 x i32> @llvm.smin.v4i32(<2 x i32> %sh, <2 x i32> splat (i32 32767))
%t = trunc <2 x i32> %ma to <2 x i16>
ret <2 x i16> %t
}
define <4 x i16> @saturating_4xi16(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: saturating_4xi16:
; CHECK: // %bb.0:
; CHECK-NEXT: sqdmulh v0.4h, v1.4h, v0.4h
; CHECK-NEXT: ret
%as = sext <4 x i16> %a to <4 x i32>
%bs = sext <4 x i16> %b to <4 x i32>
%m = mul <4 x i32> %bs, %as
%sh = ashr <4 x i32> %m, splat (i32 15)
%ma = tail call <4 x i32> @llvm.smin.v4i32(<4 x i32> %sh, <4 x i32> splat (i32 32767))
%t = trunc <4 x i32> %ma to <4 x i16>
ret <4 x i16> %t
}
define <8 x i16> @saturating_8xi16(<8 x i16> %a, <8 x i16> %b) {
; CHECK-LABEL: saturating_8xi16:
; 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 <2 x i32> @saturating_2xi32(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: saturating_2xi32:
; CHECK: // %bb.0:
; CHECK-NEXT: sqdmulh v0.2s, v1.2s, v0.2s
; CHECK-NEXT: ret
%as = sext <2 x i32> %a to <2 x i64>
%bs = sext <2 x i32> %b to <2 x i64>
%m = mul <2 x i64> %bs, %as
%sh = ashr <2 x i64> %m, splat (i64 31)
%ma = tail call <2 x i64> @llvm.smin.v8i64(<2 x i64> %sh, <2 x i64> splat (i64 2147483647))
%t = trunc <2 x i64> %ma to <2 x i32>
ret <2 x i32> %t
}
define <4 x i32> @saturating_4xi32(<4 x i32> %a, <4 x i32> %b) {
; CHECK-LABEL: saturating_4xi32:
; 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.v4i64(<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 i32> @saturating_8xi32(<8 x i32> %a, <8 x i32> %b) {
; CHECK-LABEL: saturating_8xi32:
; CHECK: // %bb.0:
; CHECK-NEXT: sqdmulh v1.4s, v3.4s, v1.4s
; CHECK-NEXT: sqdmulh v0.4s, v2.4s, v0.4s
; CHECK-NEXT: ret
%as = sext <8 x i32> %a to <8 x i64>
%bs = sext <8 x i32> %b to <8 x i64>
%m = mul <8 x i64> %bs, %as
%sh = ashr <8 x i64> %m, splat (i64 31)
%ma = tail call <8 x i64> @llvm.smin.v8i64(<8 x i64> %sh, <8 x i64> splat (i64 2147483647))
%t = trunc <8 x i64> %ma to <8 x i32>
ret <8 x i32> %t
}
define <2 x i64> @saturating_2xi32_2xi64(<2 x i32> %a, <2 x i32> %b) {
; CHECK-LABEL: saturating_2xi32_2xi64:
; CHECK: // %bb.0:
; CHECK-NEXT: sqdmulh v0.2s, v1.2s, v0.2s
; CHECK-NEXT: sshll v0.2d, v0.2s, #0
; CHECK-NEXT: ret
%as = sext <2 x i32> %a to <2 x i64>
%bs = sext <2 x i32> %b to <2 x i64>
%m = mul <2 x i64> %bs, %as
%sh = ashr <2 x i64> %m, splat (i64 31)
%ma = tail call <2 x i64> @llvm.smin.v8i64(<2 x i64> %sh, <2 x i64> splat (i64 2147483647))
ret <2 x i64> %ma
}
define <6 x i16> @saturating_6xi16(<6 x i16> %a, <6 x i16> %b) {
; CHECK-LABEL: saturating_6xi16:
; CHECK: // %bb.0:
; CHECK-NEXT: smull2 v3.4s, v1.8h, v0.8h
; CHECK-NEXT: movi v2.4s, #127, msl #8
; CHECK-NEXT: sqdmulh v0.4h, v1.4h, v0.4h
; CHECK-NEXT: sshr v3.4s, v3.4s, #15
; CHECK-NEXT: smin v2.4s, v3.4s, v2.4s
; CHECK-NEXT: xtn2 v0.8h, v2.4s
; CHECK-NEXT: ret
%as = sext <6 x i16> %a to <6 x i32>
%bs = sext <6 x i16> %b to <6 x i32>
%m = mul <6 x i32> %bs, %as
%sh = ashr <6 x i32> %m, splat (i32 15)
%ma = tail call <6 x i32> @llvm.smin.v6i32(<6 x i32> %sh, <6 x i32> splat (i32 32767))
%t = trunc <6 x i32> %ma to <6 x i16>
ret <6 x i16> %t
}
define <4 x i16> @unsupported_saturation_value_v4i16(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: unsupported_saturation_value_v4i16:
; CHECK: // %bb.0:
; CHECK-NEXT: smull v0.4s, v1.4h, v0.4h
; CHECK-NEXT: movi v1.4s, #42
; CHECK-NEXT: sshr v0.4s, v0.4s, #15
; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: ret
%as = sext <4 x i16> %a to <4 x i32>
%bs = sext <4 x i16> %b to <4 x i32>
%m = mul <4 x i32> %bs, %as
%sh = ashr <4 x i32> %m, splat (i32 15)
%ma = tail call <4 x i32> @llvm.smin.v4i32(<4 x i32> %sh, <4 x i32> splat (i32 42))
%t = trunc <4 x i32> %ma to <4 x i16>
ret <4 x i16> %t
}
define <4 x i16> @unsupported_shift_value_v4i16(<4 x i16> %a, <4 x i16> %b) {
; CHECK-LABEL: unsupported_shift_value_v4i16:
; CHECK: // %bb.0:
; CHECK-NEXT: smull v0.4s, v1.4h, v0.4h
; CHECK-NEXT: movi v1.4s, #127, msl #8
; CHECK-NEXT: sshr v0.4s, v0.4s, #3
; CHECK-NEXT: smin v0.4s, v0.4s, v1.4s
; CHECK-NEXT: xtn v0.4h, v0.4s
; CHECK-NEXT: ret
%as = sext <4 x i16> %a to <4 x i32>
%bs = sext <4 x i16> %b to <4 x i32>
%m = mul <4 x i32> %bs, %as
%sh = ashr <4 x i32> %m, splat (i32 3)
%ma = tail call <4 x i32> @llvm.smin.v4i32(<4 x i32> %sh, <4 x i32> splat (i32 32767))
%t = trunc <4 x i32> %ma to <4 x i16>
ret <4 x i16> %t
}
define <2 x i16> @extend_to_illegal_type(<2 x i16> %a, <2 x i16> %b) {
; CHECK-LABEL: extend_to_illegal_type:
; CHECK: // %bb.0:
; CHECK-NEXT: shl v0.2s, v0.2s, #16
; CHECK-NEXT: shl v1.2s, v1.2s, #16
; CHECK-NEXT: sshr v0.2s, v0.2s, #16
; CHECK-NEXT: sshr v1.2s, v1.2s, #16
; CHECK-NEXT: sqdmulh v0.2s, v1.2s, v0.2s
; CHECK-NEXT: ret
%as = sext <2 x i16> %a to <2 x i48>
%bs = sext <2 x i16> %b to <2 x i48>
%m = mul <2 x i48> %bs, %as
%sh = ashr <2 x i48> %m, splat (i48 15)
%ma = tail call <2 x i48> @llvm.smin.v4i32(<2 x i48> %sh, <2 x i48> splat (i48 32767))
%t = trunc <2 x i48> %ma to <2 x i16>
ret <2 x i16> %t
}
define <2 x i11> @illegal_source(<2 x i11> %a, <2 x i11> %b) {
; CHECK-LABEL: illegal_source:
; CHECK: // %bb.0:
; CHECK-NEXT: shl v0.2s, v0.2s, #21
; CHECK-NEXT: shl v1.2s, v1.2s, #21
; CHECK-NEXT: sshr v0.2s, v0.2s, #21
; CHECK-NEXT: sshr v1.2s, v1.2s, #21
; CHECK-NEXT: mul v0.2s, v1.2s, v0.2s
; CHECK-NEXT: movi v1.2s, #127, msl #8
; CHECK-NEXT: sshr v0.2s, v0.2s, #15
; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: ret
%as = sext <2 x i11> %a to <2 x i32>
%bs = sext <2 x i11> %b to <2 x i32>
%m = mul <2 x i32> %bs, %as
%sh = ashr <2 x i32> %m, splat (i32 15)
%ma = tail call <2 x i32> @llvm.smin.v2i32(<2 x i32> %sh, <2 x i32> splat (i32 32767))
%t = trunc <2 x i32> %ma to <2 x i11>
ret <2 x i11> %t
}
define <1 x i16> @saturating_1xi16(<1 x i16> %a, <1 x i16> %b) {
; CHECK-LABEL: saturating_1xi16:
; CHECK: // %bb.0:
; CHECK-NEXT: zip1 v0.4h, v0.4h, v0.4h
; CHECK-NEXT: zip1 v1.4h, v1.4h, v0.4h
; CHECK-NEXT: shl v0.2s, v0.2s, #16
; CHECK-NEXT: sshr v0.2s, v0.2s, #16
; CHECK-NEXT: shl v1.2s, v1.2s, #16
; CHECK-NEXT: sshr v1.2s, v1.2s, #16
; CHECK-NEXT: mul v0.2s, v1.2s, v0.2s
; CHECK-NEXT: movi v1.2s, #127, msl #8
; CHECK-NEXT: sshr v0.2s, v0.2s, #15
; CHECK-NEXT: smin v0.2s, v0.2s, v1.2s
; CHECK-NEXT: uzp1 v0.4h, v0.4h, v0.4h
; CHECK-NEXT: ret
%as = sext <1 x i16> %a to <1 x i32>
%bs = sext <1 x i16> %b to <1 x i32>
%m = mul <1 x i32> %bs, %as
%sh = ashr <1 x i32> %m, splat (i32 15)
%ma = tail call <1 x i32> @llvm.smin.v1i32(<1 x i32> %sh, <1 x i32> splat (i32 32767))
%t = trunc <1 x i32> %ma to <1 x i16>
ret <1 x i16> %t
}