[ARM] Better patterns for fp <> predicate vectors

[dmgreen]

Summary

These are some better patterns for converting between predicates and floating points. Much like the extends, we select "1"/"-1" or "0" depending on the predicate value. Or we perform a compare against 0 to convert to a predicate.

[SjoerdMeijer]

LGTM

llvm/lib/Target/ARM/ARMISelLowering.cpp

for (auto VT : pTypes) {
     setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
     setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
     setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
     setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
     setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
     setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
     setOperationAction(ISD::SETCC, VT, Custom);
     setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
-    setOperationAction(ISD::FP_TO_SINT, VT, Expand);
-    setOperationAction(ISD::FP_TO_UINT, VT, Expand);
-    setOperationAction(ISD::SINT_TO_FP, VT, Expand);
-    setOperationAction(ISD::UINT_TO_FP, VT, Expand);
   }
 }
 
 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
                                      const ARMSubtarget &STI)
     : TargetLowering(TM), Subtarget(&STI) {
   RegInfo = Subtarget->getRegisterInfo();
   Itins = Subtarget->getInstrItineraryData();

llvm/lib/Target/ARM/ARMInstrMVE.td

let Predicates = [HasMVEInt] in {
 
   def : Pat<(v8f16 (vselect (v8i16 MQPR:$pred), (v8f16 MQPR:$v1), (v8f16 MQPR:$v2))),
             (v8f16 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
                               (MVE_VCMPi16 (v8i16 MQPR:$pred), (MVE_VMOVimmi16 0), 1)))>;
   def : Pat<(v4f32 (vselect (v4i32 MQPR:$pred), (v4f32 MQPR:$v1), (v4f32 MQPR:$v2))),
             (v4f32 (MVE_VPSEL MQPR:$v1, MQPR:$v2, 0,
                               (MVE_VCMPi32 (v4i32 MQPR:$pred), (MVE_VMOVimmi32 0), 1)))>;
 
+  // Pred <-> Int
   def : Pat<(v16i8 (zext  (v16i1 VCCR:$pred))),
             (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 1), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
   def : Pat<(v8i16 (zext  (v8i1  VCCR:$pred))),
             (v8i16 (MVE_VPSEL (MVE_VMOVimmi16 1), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
   def : Pat<(v4i32 (zext  (v4i1  VCCR:$pred))),
             (v4i32 (MVE_VPSEL (MVE_VMOVimmi32 1), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
 
   def : Pat<(v16i8 (sext  (v16i1 VCCR:$pred))),
             (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
   def : Pat<(v8i16 (sext  (v8i1  VCCR:$pred))),
             (v8i16 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
   def : Pat<(v4i32 (sext  (v4i1  VCCR:$pred))),
             (v4i32 (MVE_VPSEL (MVE_VMOVimmi8 255), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
 
   def : Pat<(v16i8 (anyext  (v16i1 VCCR:$pred))),
             (v16i8 (MVE_VPSEL (MVE_VMOVimmi8 1), (MVE_VMOVimmi8 0), 0, VCCR:$pred))>;
   def : Pat<(v8i16 (anyext  (v8i1  VCCR:$pred))),
             (v8i16 (MVE_VPSEL (MVE_VMOVimmi16 1), (MVE_VMOVimmi16 0), 0, VCCR:$pred))>;
   def : Pat<(v4i32 (anyext  (v4i1  VCCR:$pred))),
             (v4i32 (MVE_VPSEL (MVE_VMOVimmi32 1), (MVE_VMOVimmi32 0), 0, VCCR:$pred))>;
 }
 
+let Predicates = [HasMVEFloat] in {
+  // Pred <-> Float
+  // 112 is 1.0 in float
+  def : Pat<(v4f32 (uint_to_fp (v4i1 VCCR:$pred))),
+            (v4f32 (MVE_VPSEL (v4f32 (MVE_VMOVimmf32 112)), (v4f32 (MVE_VMOVimmi32 0)), 0, VCCR:$pred))>;
+  // 2620 in 1.0 in half
+  def : Pat<(v8f16 (uint_to_fp (v8i1 VCCR:$pred))),
+            (v8f16 (MVE_VPSEL (v8f16 (MVE_VMOVimmi16 2620)), (v8f16 (MVE_VMOVimmi16 0)), 0, VCCR:$pred))>;
+  // 240 is -1.0 in float
+  def : Pat<(v4f32 (sint_to_fp (v4i1 VCCR:$pred))),
+            (v4f32 (MVE_VPSEL (v4f32 (MVE_VMOVimmf32 240)), (v4f32 (MVE_VMOVimmi32 0)), 0, VCCR:$pred))>;
+  // 2748 is -1.0 in half
+  def : Pat<(v8f16 (sint_to_fp (v8i1 VCCR:$pred))),
+            (v8f16 (MVE_VPSEL (v8f16 (MVE_VMOVimmi16 2748)), (v8f16 (MVE_VMOVimmi16 0)), 0, VCCR:$pred))>;
+
+  def : Pat<(v4i1 (fp_to_uint (v4f32 MQPR:$v1))),
+            (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, 1))>;
+  def : Pat<(v8i1 (fp_to_uint (v8f16 MQPR:$v1))),
+            (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), ZR, 1))>;
+  def : Pat<(v4i1 (fp_to_sint (v4f32 MQPR:$v1))),
+            (v4i1 (MVE_VCMPf32r (v4f32 MQPR:$v1), ZR, 1))>;
+  def : Pat<(v8i1 (fp_to_sint (v8f16 MQPR:$v1))),
+            (v8i1 (MVE_VCMPf16r (v8f16 MQPR:$v1), ZR, 1))>;
+}
+
 def MVE_VPNOT : MVE_p<(outs), (ins), NoItinerary,
                       "vpnot", "", "", vpred_n, "", []> {
   let Inst{31-0} = 0b11111110001100010000111101001101;
   let Unpredictable{19-17} = 0b111;
   let Unpredictable{12} = 0b1;
   let Unpredictable{7} = 0b1;
   let Unpredictable{5} = 0b1;
   let Defs = [P0];

llvm/test/CodeGen/Thumb2/mve-pred-ext.ll

entry:
   %0 = zext <2 x i1> %c to <2 x i64>
   ret <2 x i64> %0
 }
 
 
 define arm_aapcs_vfpcc <4 x float> @uitofp_v4i1_v4f32(<4 x i32> %src) {
 ; CHECK-LABEL: uitofp_v4i1_v4f32:
 ; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i32 q1, #0x0
+; CHECK-NEXT:    vmov.f32 q2, #1.000000e+00
 ; CHECK-NEXT:    vcmp.s32 gt, q0, zr
-; CHECK-NEXT:    vmrs r0, p0
-; CHECK-NEXT:    ubfx r1, r0, #8, #1
-; CHECK-NEXT:    ubfx r2, r0, #12, #1
-; CHECK-NEXT:    vmov s0, r2
-; CHECK-NEXT:    vmov s4, r1
-; CHECK-NEXT:    vcvt.f32.u32 s3, s0
-; CHECK-NEXT:    ubfx r2, r0, #4, #1
-; CHECK-NEXT:    vcvt.f32.u32 s2, s4
-; CHECK-NEXT:    and r0, r0, #1
-; CHECK-NEXT:    vmov s4, r2
-; CHECK-NEXT:    vcvt.f32.u32 s1, s4
-; CHECK-NEXT:    vmov s4, r0
-; CHECK-NEXT:    vcvt.f32.u32 s0, s4
+; CHECK-NEXT:    vpsel q0, q2, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %c = icmp sgt <4 x i32> %src, zeroinitializer
   %0 = uitofp <4 x i1> %c to <4 x float>
   ret <4 x float> %0
 }
 
 define arm_aapcs_vfpcc <4 x float> @sitofp_v4i1_v4f32(<4 x i32> %src) {
 ; CHECK-LABEL: sitofp_v4i1_v4f32:
 ; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i32 q1, #0x0
+; CHECK-NEXT:    vmov.f32 q2, #-1.000000e+00
 ; CHECK-NEXT:    vcmp.s32 gt, q0, zr
-; CHECK-NEXT:    vmrs r0, p0
-; CHECK-NEXT:    and r1, r0, #1
-; CHECK-NEXT:    ubfx r2, r0, #8, #1
-; CHECK-NEXT:    ubfx r3, r0, #4, #1
-; CHECK-NEXT:    ubfx r0, r0, #12, #1
-; CHECK-NEXT:    rsbs r2, r2, #0
-; CHECK-NEXT:    rsbs r0, r0, #0
-; CHECK-NEXT:    vmov s4, r2
-; CHECK-NEXT:    vmov s0, r0
-; CHECK-NEXT:    rsbs r0, r3, #0
-; CHECK-NEXT:    vcvt.f32.s32 s3, s0
-; CHECK-NEXT:    vcvt.f32.s32 s2, s4
-; CHECK-NEXT:    vmov s4, r0
-; CHECK-NEXT:    rsbs r0, r1, #0
-; CHECK-NEXT:    vcvt.f32.s32 s1, s4
-; CHECK-NEXT:    vmov s4, r0
-; CHECK-NEXT:    vcvt.f32.s32 s0, s4
+; CHECK-NEXT:    vpsel q0, q2, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %c = icmp sgt <4 x i32> %src, zeroinitializer
   %0 = sitofp <4 x i1> %c to <4 x float>
   ret <4 x float> %0
 }
 
 define arm_aapcs_vfpcc <4 x float> @fptoui_v4i1_v4f32(<4 x float> %src) {
 ; CHECK-LABEL: fptoui_v4i1_v4f32:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vcvt.s32.f32 s4, s0
-; CHECK-NEXT:    movs r0, #0
-; CHECK-NEXT:    vmov.f32 q2, #1.000000e+00
-; CHECK-NEXT:    vmov r1, s4
-; CHECK-NEXT:    vcvt.s32.f32 s4, s1
-; CHECK-NEXT:    rsbs r1, r1, #0
-; CHECK-NEXT:    bfi r0, r1, #0, #4
-; CHECK-NEXT:    vmov r1, s4
-; CHECK-NEXT:    vcvt.s32.f32 s4, s2
-; CHECK-NEXT:    vcvt.s32.f32 s0, s3
-; CHECK-NEXT:    rsbs r1, r1, #0
-; CHECK-NEXT:    bfi r0, r1, #4, #4
-; CHECK-NEXT:    vmov r1, s4
 ; CHECK-NEXT:    vmov.i32 q1, #0x0
-; CHECK-NEXT:    rsbs r1, r1, #0
-; CHECK-NEXT:    bfi r0, r1, #8, #4
-; CHECK-NEXT:    vmov r1, s0
-; CHECK-NEXT:    rsbs r1, r1, #0
-; CHECK-NEXT:    bfi r0, r1, #12, #4
-; CHECK-NEXT:    vmsr p0, r0
+; CHECK-NEXT:    vmov.f32 q2, #1.000000e+00
+; CHECK-NEXT:    vcmp.f32 ne, q0, zr
 ; CHECK-NEXT:    vpsel q0, q2, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = fptoui <4 x float> %src to <4 x i1>
   %s = select <4 x i1> %0, <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, <4 x float> zeroinitializer
   ret <4 x float> %s
 }
 
 define arm_aapcs_vfpcc <4 x float> @fptosi_v4i1_v4f32(<4 x float> %src) {
 ; CHECK-LABEL: fptosi_v4i1_v4f32:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vcvt.s32.f32 s4, s0
-; CHECK-NEXT:    movs r0, #0
-; CHECK-NEXT:    vmov.f32 q2, #1.000000e+00
-; CHECK-NEXT:    vmov r1, s4
-; CHECK-NEXT:    vcvt.s32.f32 s4, s1
-; CHECK-NEXT:    bfi r0, r1, #0, #4
-; CHECK-NEXT:    vmov r1, s4
-; CHECK-NEXT:    vcvt.s32.f32 s4, s2
-; CHECK-NEXT:    bfi r0, r1, #4, #4
-; CHECK-NEXT:    vcvt.s32.f32 s0, s3
-; CHECK-NEXT:    vmov r1, s4
 ; CHECK-NEXT:    vmov.i32 q1, #0x0
-; CHECK-NEXT:    bfi r0, r1, #8, #4
-; CHECK-NEXT:    vmov r1, s0
-; CHECK-NEXT:    bfi r0, r1, #12, #4
-; CHECK-NEXT:    vmsr p0, r0
+; CHECK-NEXT:    vmov.f32 q2, #1.000000e+00
+; CHECK-NEXT:    vcmp.f32 ne, q0, zr
 ; CHECK-NEXT:    vpsel q0, q2, q1
 ; CHECK-NEXT:    bx lr
 entry:
   %0 = fptosi <4 x float> %src to <4 x i1>
   %s = select <4 x i1> %0, <4 x float> <float 1.0, float 1.0, float 1.0, float 1.0>, <4 x float> zeroinitializer
   ret <4 x float> %s
 }
 
+
+
+define arm_aapcs_vfpcc <8 x half> @uitofp_v8i1_v8f16(<8 x i16> %src) {
+; CHECK-LABEL: uitofp_v8i1_v8f16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i16 q1, #0x0
+; CHECK-NEXT:    vmov.i16 q2, #0x3c00
+; CHECK-NEXT:    vcmp.s16 gt, q0, zr
+; CHECK-NEXT:    vpsel q0, q2, q1
+; CHECK-NEXT:    bx lr
+entry:
+  %c = icmp sgt <8 x i16> %src, zeroinitializer
+  %0 = uitofp <8 x i1> %c to <8 x half>
+  ret <8 x half> %0
+}
+
+define arm_aapcs_vfpcc <8 x half> @sitofp_v8i1_v8f16(<8 x i16> %src) {
+; CHECK-LABEL: sitofp_v8i1_v8f16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i16 q1, #0x0
+; CHECK-NEXT:    vmov.i16 q2, #0xbc00
+; CHECK-NEXT:    vcmp.s16 gt, q0, zr
+; CHECK-NEXT:    vpsel q0, q2, q1
+; CHECK-NEXT:    bx lr
+entry:
+  %c = icmp sgt <8 x i16> %src, zeroinitializer
+  %0 = sitofp <8 x i1> %c to <8 x half>
+  ret <8 x half> %0
+}
+
+define arm_aapcs_vfpcc <8 x half> @fptoui_v8i1_v8f16(<8 x half> %src) {
+; CHECK-LABEL: fptoui_v8i1_v8f16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i32 q1, #0x0
+; CHECK-NEXT:    vmov.i16 q2, #0x3c00
+; CHECK-NEXT:    vcmp.f16 ne, q0, zr
+; CHECK-NEXT:    vpsel q0, q2, q1
+; CHECK-NEXT:    bx lr
+entry:
+  %0 = fptoui <8 x half> %src to <8 x i1>
+  %s = select <8 x i1> %0, <8 x half> <half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0>, <8 x half> zeroinitializer
+  ret <8 x half> %s
+}
+
+define arm_aapcs_vfpcc <8 x half> @fptosi_v8i1_v8f16(<8 x half> %src) {
+; CHECK-LABEL: fptosi_v8i1_v8f16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vmov.i32 q1, #0x0
+; CHECK-NEXT:    vmov.i16 q2, #0x3c00
+; CHECK-NEXT:    vcmp.f16 ne, q0, zr
+; CHECK-NEXT:    vpsel q0, q2, q1
+; CHECK-NEXT:    bx lr
+entry:
+  %0 = fptosi <8 x half> %src to <8 x i1>
+  %s = select <8 x i1> %0, <8 x half> <half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0, half 1.0>, <8 x half> zeroinitializer
+  ret <8 x half> %s
+}