[AArch64] Add patterns for conversions using fixed-point scvtf

[momchil-velikov]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: Momchil Velikov (momchil-velikov)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/92922.diff

3 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+1-1)
• (modified) llvm/lib/Target/AArch64/AArch64InstrInfo.td (+25)
• (added) llvm/test/CodeGen/AArch64/fixed-point-conv-vec-pat.ll (+103)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index e31a27e9428e8..dd422214d45f4 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -14328,7 +14328,7 @@ SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
       unsigned Opc =
           (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
       return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
-                         DAG.getConstant(Cnt, DL, MVT::i32));
+                         DAG.getConstant(Cnt, DL, MVT::i32), Op->getFlags());
     }
 
     // Right shift register.  Note, there is not a shift right register
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
index a39e3b7be76dc..291f553776752 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrInfo.td
+++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -735,6 +735,12 @@ def AArch64rev64 : SDNode<"AArch64ISD::REV64", SDT_AArch64UnaryVec>;
 def AArch64ext : SDNode<"AArch64ISD::EXT", SDT_AArch64ExtVec>;
 
 def AArch64vashr : SDNode<"AArch64ISD::VASHR", SDT_AArch64vshift>;
+
+def AArch64vashr_exact : PatFrag<(ops          node:$lhs, node:$rhs),
+                                 (AArch64vashr node:$lhs, node:$rhs), [{
+  return N->getFlags().hasExact();
+}]>;
+
 def AArch64vlshr : SDNode<"AArch64ISD::VLSHR", SDT_AArch64vshift>;
 def AArch64vshl : SDNode<"AArch64ISD::VSHL", SDT_AArch64vshift>;
 def AArch64sqshli : SDNode<"AArch64ISD::SQSHL_I", SDT_AArch64vshift>;
@@ -7712,6 +7718,25 @@ defm SCVTF: SIMDVectorRShiftToFP<0, 0b11100, "scvtf",
 defm RSHRN   : SIMDVectorRShiftNarrowBHS<0, 0b10001, "rshrn", AArch64rshrn>;
 defm SHL     : SIMDVectorLShiftBHSD<0, 0b01010, "shl", AArch64vshl>;
 
+let Predicates = [HasNEON] in {
+def : Pat<(v2f32 (sint_to_fp (v2i32 (AArch64vashr_exact v2i32:$Vn, i32:$shift)))),
+          (SCVTFv2i32_shift $Vn, vecshiftR32:$shift)>;
+
+def : Pat<(v4f32 (sint_to_fp (v4i32 (AArch64vashr_exact v4i32:$Vn, i32:$shift)))),
+          (SCVTFv4i32_shift $Vn, vecshiftR32:$shift)>;
+
+def : Pat<(v2f64 (sint_to_fp (v2i64 (AArch64vashr_exact v2i64:$Vn, i32:$shift)))),
+          (SCVTFv2i64_shift $Vn, vecshiftR64:$shift)>;
+}
+
+let Predicates = [HasNEON, HasFullFP16] in {
+def : Pat<(v4f16 (sint_to_fp (v4i16 (AArch64vashr_exact v4i16:$Vn, i32:$shift)))),
+          (SCVTFv4i16_shift $Vn, vecshiftR16:$shift)>;
+
+def : Pat<(v8f16 (sint_to_fp (v8i16 (AArch64vashr_exact v8i16:$Vn, i32:$shift)))),
+          (SCVTFv8i16_shift $Vn, vecshiftR16:$shift)>;
+}
+
 // X << 1 ==> X + X
 class SHLToADDPat<ValueType ty, RegisterClass regtype>
   : Pat<(ty (AArch64vshl (ty regtype:$Rn), (i32 1))),
diff --git a/llvm/test/CodeGen/AArch64/fixed-point-conv-vec-pat.ll b/llvm/test/CodeGen/AArch64/fixed-point-conv-vec-pat.ll
new file mode 100644
index 0000000000000..7141b5b03a1ac
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/fixed-point-conv-vec-pat.ll
@@ -0,0 +1,103 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s | FileCheck %s
+
+target triple = "aarch64"
+
+; First some corner cases
+define <4 x float> @f_v4_s0(<4 x i32> %u) {
+; CHECK-LABEL: f_v4_s0:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.4s, v0.4s
+; CHECK-NEXT:    ret
+  %s = ashr exact <4 x i32> %u, <i32 0, i32 0, i32 0, i32 0>
+  %v = sitofp <4 x i32> %s to <4 x float>
+  ret <4 x float> %v
+}
+
+define <4 x float> @f_v4_s1(<4 x i32> %u) {
+; CHECK-LABEL: f_v4_s1:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.4s, v0.4s, #1
+; CHECK-NEXT:    ret
+  %s = ashr exact <4 x i32> %u, <i32 1, i32 1, i32 1, i32 1>
+  %v = sitofp <4 x i32> %s to <4 x float>
+  ret <4 x float> %v
+}
+
+define <4 x float> @f_v4_s24_inexact(<4 x i32> %u) {
+; CHECK-LABEL: f_v4_s24_inexact:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sshr v0.4s, v0.4s, #24
+; CHECK-NEXT:    scvtf v0.4s, v0.4s
+; CHECK-NEXT:    ret
+  %s = ashr <4 x i32> %u, <i32 24, i32 24, i32 24, i32 24>
+  %v = sitofp <4 x i32> %s to <4 x float>
+  ret <4 x float> %v
+}
+
+define <4 x float> @f_v4_s32(<4 x i32> %u) {
+; CHECK-LABEL: f_v4_s32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    movi v0.2d, #0000000000000000
+; CHECK-NEXT:    ret
+  %s = ashr <4 x i32> %u, <i32 32, i32 32, i32 32, i32 32>
+  %v = sitofp <4 x i32> %s to <4 x float>
+  ret <4 x float> %v
+}
+
+; Common cases for conversion from signed integer to floating point types
+define <2 x float> @f_v2_s24(<2 x i32> %u) {
+; CHECK-LABEL: f_v2_s24:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.2s, v0.2s, #24
+; CHECK-NEXT:    ret
+  %s = ashr exact <2 x i32> %u, <i32 24, i32 24>
+  %v = sitofp <2 x i32> %s to <2 x float>
+  ret <2 x float> %v
+}
+
+define <4 x float> @f_v4_s24(<4 x i32> %u) {
+; CHECK-LABEL: f_v4_s24:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.4s, v0.4s, #24
+; CHECK-NEXT:    ret
+  %s = ashr exact <4 x i32> %u, <i32 24, i32 24, i32 24, i32 24>
+  %v = sitofp <4 x i32> %s to <4 x float>
+  ret <4 x float> %v
+}
+
+; Check legalisation to <2 x f64> does not get in the way
+define <8 x double> @d_v8_s64(<8 x i64> %u) {
+; CHECK-LABEL: d_v8_s64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.2d, v0.2d, #56
+; CHECK-NEXT:    scvtf v1.2d, v1.2d, #56
+; CHECK-NEXT:    scvtf v2.2d, v2.2d, #56
+; CHECK-NEXT:    scvtf v3.2d, v3.2d, #56
+; CHECK-NEXT:    ret
+  %s = ashr exact <8 x i64> %u, <i64 56, i64 56, i64 56, i64 56, i64 56, i64 56, i64 56, i64 56>
+  %v = sitofp <8 x i64> %s to <8 x double>
+  ret <8 x double> %v
+}
+
+define <4 x half> @h_v4_s8(<4 x i16> %u) #0 {
+; CHECK-LABEL: h_v4_s8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.4h, v0.4h, #8
+; CHECK-NEXT:    ret
+  %s = ashr exact <4 x i16> %u, <i16 8, i16 8, i16 8, i16 8>
+  %v = sitofp <4 x i16> %s to <4 x half>
+  ret <4 x half> %v
+}
+
+define <8 x half> @h_v8_s8(<8 x i16> %u) #0 {
+; CHECK-LABEL: h_v8_s8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    scvtf v0.8h, v0.8h, #8
+; CHECK-NEXT:    ret
+  %s = ashr exact <8 x i16> %u, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
+  %v = sitofp <8 x i16> %s to <8 x half>
+  ret <8 x half> %v
+}
+
+attributes #0 = { "target-features"="+fullfp16"}

[davemgreen]

Thanks - This looks great, and I couldn't figure out any way in which it wouldn't match the version with a fmul/fdiv like the pseudo code. LGTM