[AArch64] Extend int-to-fp load optimization to support f16 (llvm#168076)

Author: guy-david

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -20146,7 +20146,9 @@ static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
     return Res;
 
   EVT VT = N->getValueType(0);
-  if (VT != MVT::f32 && VT != MVT::f64)
+  if (VT != MVT::f16 && VT != MVT::f32 && VT != MVT::f64)
+    return SDValue();
+  if (VT == MVT::f16 && !Subtarget->hasFullFP16())
     return SDValue();
 
   // Only optimize when the source and destination types have the same width.

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -7014,6 +7014,19 @@ multiclass UIntToFPROLoadPat<ValueType DstTy, ValueType SrcTy,
                                  sub))>;
 }
 
+let Predicates = [HasNEONandIsSME2p2StreamingSafe, HasFullFP16] in {
+defm : UIntToFPROLoadPat<f16, i32, zextloadi8,
+                         UCVTFv1i16, ro8, LDRBroW, LDRBroX, bsub>;
+def : Pat <(f16 (uint_to_fp (i32
+               (zextloadi8 (am_indexed8 GPR64sp:$Rn, uimm12s1:$offset))))),
+           (UCVTFv1i16 (INSERT_SUBREG (f16 (IMPLICIT_DEF)),
+                          (LDRBui GPR64sp:$Rn, uimm12s1:$offset), bsub))>;
+def : Pat <(f16 (uint_to_fp (i32
+                     (zextloadi8 (am_unscaled8 GPR64sp:$Rn, simm9:$offset))))),
+           (UCVTFv1i16 (INSERT_SUBREG (f16 (IMPLICIT_DEF)),
+                          (LDURBi GPR64sp:$Rn, simm9:$offset), bsub))>;
+}
+
 defm : UIntToFPROLoadPat<f32, i32, zextloadi8,
                          UCVTFv1i32, ro8, LDRBroW, LDRBroX, bsub>;
 def : Pat <(f32 (uint_to_fp (i32

llvm/test/CodeGen/AArch64/int-to-fp-no-neon.ll

@@ -43,11 +43,17 @@ entry:
 }
 
 define half @ui8_to_half(ptr %i, ptr %f) {
-; CHECK-LABEL: ui8_to_half:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldrb w8, [x0]
-; CHECK-NEXT:    ucvtf h0, w8
-; CHECK-NEXT:    ret
+; NEON-ENABLED-LABEL: ui8_to_half:
+; NEON-ENABLED:       // %bb.0: // %entry
+; NEON-ENABLED-NEXT:    ldr b0, [x0]
+; NEON-ENABLED-NEXT:    ucvtf h0, h0
+; NEON-ENABLED-NEXT:    ret
+;
+; NEON-DISABLED-LABEL: ui8_to_half:
+; NEON-DISABLED:       // %bb.0: // %entry
+; NEON-DISABLED-NEXT:    ldrb w8, [x0]
+; NEON-DISABLED-NEXT:    ucvtf h0, w8
+; NEON-DISABLED-NEXT:    ret
 entry:
   %ld = load i8, ptr %i, align 1
   %conv = uitofp i8 %ld to half
@@ -91,11 +97,17 @@ entry:
 }
 
 define half @ui16_to_half(ptr %i, ptr %f) {
-; CHECK-LABEL: ui16_to_half:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldrh w8, [x0]
-; CHECK-NEXT:    ucvtf h0, w8
-; CHECK-NEXT:    ret
+; NEON-ENABLED-LABEL: ui16_to_half:
+; NEON-ENABLED:       // %bb.0: // %entry
+; NEON-ENABLED-NEXT:    ldr h0, [x0]
+; NEON-ENABLED-NEXT:    ucvtf h0, h0
+; NEON-ENABLED-NEXT:    ret
+;
+; NEON-DISABLED-LABEL: ui16_to_half:
+; NEON-DISABLED:       // %bb.0: // %entry
+; NEON-DISABLED-NEXT:    ldrh w8, [x0]
+; NEON-DISABLED-NEXT:    ucvtf h0, w8
+; NEON-DISABLED-NEXT:    ret
 entry:
   %ld = load i16, ptr %i, align 1
   %conv = uitofp i16 %ld to half
@@ -277,11 +289,17 @@ entry:
 }
 
 define half @si16_to_half(ptr %i, ptr %f) {
-; CHECK-LABEL: si16_to_half:
-; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldrsh w8, [x0]
-; CHECK-NEXT:    scvtf h0, w8
-; CHECK-NEXT:    ret
+; NEON-ENABLED-LABEL: si16_to_half:
+; NEON-ENABLED:       // %bb.0: // %entry
+; NEON-ENABLED-NEXT:    ldr h0, [x0]
+; NEON-ENABLED-NEXT:    scvtf h0, h0
+; NEON-ENABLED-NEXT:    ret
+;
+; NEON-DISABLED-LABEL: si16_to_half:
+; NEON-DISABLED:       // %bb.0: // %entry
+; NEON-DISABLED-NEXT:    ldrsh w8, [x0]
+; NEON-DISABLED-NEXT:    scvtf h0, w8
+; NEON-DISABLED-NEXT:    ret
 entry:
   %ld = load i16, ptr %i, align 1
   %conv = sitofp i16 %ld to half

llvm/test/CodeGen/AArch64/itofp.ll

@@ -1090,11 +1090,17 @@ define half @stofp_load_i16_f16(ptr %p) {
 ; CHECK-NOFP16-NEXT:    fcvt h0, s0
 ; CHECK-NOFP16-NEXT:    ret
 ;
-; CHECK-FP16-LABEL: stofp_load_i16_f16:
-; CHECK-FP16:       // %bb.0: // %entry
-; CHECK-FP16-NEXT:    ldrsh w8, [x0]
-; CHECK-FP16-NEXT:    scvtf h0, w8
-; CHECK-FP16-NEXT:    ret
+; CHECK-FP16-SD-LABEL: stofp_load_i16_f16:
+; CHECK-FP16-SD:       // %bb.0: // %entry
+; CHECK-FP16-SD-NEXT:    ldr h0, [x0]
+; CHECK-FP16-SD-NEXT:    scvtf h0, h0
+; CHECK-FP16-SD-NEXT:    ret
+;
+; CHECK-FP16-GI-LABEL: stofp_load_i16_f16:
+; CHECK-FP16-GI:       // %bb.0: // %entry
+; CHECK-FP16-GI-NEXT:    ldrsh w8, [x0]
+; CHECK-FP16-GI-NEXT:    scvtf h0, w8
+; CHECK-FP16-GI-NEXT:    ret
 entry:
   %a = load i16, ptr %p
   %c = sitofp i16 %a to half
@@ -1109,11 +1115,17 @@ define half @utofp_load_i16_f16(ptr %p) {
 ; CHECK-NOFP16-NEXT:    fcvt h0, s0
 ; CHECK-NOFP16-NEXT:    ret
 ;
-; CHECK-FP16-LABEL: utofp_load_i16_f16:
-; CHECK-FP16:       // %bb.0: // %entry
-; CHECK-FP16-NEXT:    ldrh w8, [x0]
-; CHECK-FP16-NEXT:    ucvtf h0, w8
-; CHECK-FP16-NEXT:    ret
+; CHECK-FP16-SD-LABEL: utofp_load_i16_f16:
+; CHECK-FP16-SD:       // %bb.0: // %entry
+; CHECK-FP16-SD-NEXT:    ldr h0, [x0]
+; CHECK-FP16-SD-NEXT:    ucvtf h0, h0
+; CHECK-FP16-SD-NEXT:    ret
+;
+; CHECK-FP16-GI-LABEL: utofp_load_i16_f16:
+; CHECK-FP16-GI:       // %bb.0: // %entry
+; CHECK-FP16-GI-NEXT:    ldrh w8, [x0]
+; CHECK-FP16-GI-NEXT:    ucvtf h0, w8
+; CHECK-FP16-GI-NEXT:    ret
 entry:
   %a = load i16, ptr %p
   %c = uitofp i16 %a to half
@@ -1149,8 +1161,8 @@ define half @utofp_load_i8_f16(ptr %p) {
 ;
 ; CHECK-FP16-LABEL: utofp_load_i8_f16:
 ; CHECK-FP16:       // %bb.0: // %entry
-; CHECK-FP16-NEXT:    ldrb w8, [x0]
-; CHECK-FP16-NEXT:    ucvtf h0, w8
+; CHECK-FP16-NEXT:    ldr b0, [x0]
+; CHECK-FP16-NEXT:    ucvtf h0, h0
 ; CHECK-FP16-NEXT:    ret
 entry:
   %a = load i8, ptr %p