[ARM] Select vaddva

[samtebbs]

Summary

This patch adds vaddva selection.

[dmgreen]

Can you add some i64 tests, just to show that it does something without breaking.

[dmgreen]

LGTM, thanks.

Its sometimes hard to tell if these functions exactly match up to that they are replacing, with all the extends going on. I believe this one is OK, but make sure it's tested.

llvm/test/CodeGen/Thumb2/mve-vaddv.ll

38

Can you add one for vaddv too, just before you commit?

[samtebbs]

llvm/test/CodeGen/Thumb2/mve-vaddv.ll

38

Yes, good spot.

llvm/lib/Target/ARM/ARMInstrMVE.td

 defm MVE_VADDVu8  : MVE_VADDV_A<"u8",  0b1, 0b00>;
 defm MVE_VADDVu16 : MVE_VADDV_A<"u16", 0b1, 0b01>;
 defm MVE_VADDVu32 : MVE_VADDV_A<"u32", 0b1, 0b10>;
 
 let Predicates = [HasMVEInt] in {
   def : Pat<(i32 (vecreduce_add (v4i32 MQPR:$src))), (i32 (MVE_VADDVu32no_acc $src))>;
   def : Pat<(i32 (vecreduce_add (v8i16 MQPR:$src))), (i32 (MVE_VADDVu16no_acc $src))>;
   def : Pat<(i32 (vecreduce_add (v16i8 MQPR:$src))), (i32 (MVE_VADDVu8no_acc $src))>;
+  def : Pat<(i32 (add (i32 (vecreduce_add (v4i32 MQPR:$src1))), (i32 tGPR:$src2))),
+            (i32 (MVE_VADDVu32acc $src2, $src1))>;
+  def : Pat<(i32 (add (i32 (vecreduce_add (v8i16 MQPR:$src1))), (i32 tGPR:$src2))),
+            (i32 (MVE_VADDVu16acc $src2, $src1))>;
+  def : Pat<(i32 (add (i32 (vecreduce_add (v16i8 MQPR:$src1))), (i32 tGPR:$src2))),
+            (i32 (MVE_VADDVu8acc $src2, $src1))>;
+
 }
 
 class MVE_VADDLV<string iname, string suffix, dag iops, string cstr,
                bit A, bit U, list<dag> pattern=[]>
   : MVE_rDest<(outs tGPREven:$RdaLo, tGPROdd:$RdaHi), iops, NoItinerary, iname,
               suffix, "$RdaLo, $RdaHi, $Qm", cstr, pattern> {
   bits<3> Qm;
   bits<4> RdaLo;

llvm/test/CodeGen/Thumb2/mve-vaddv.ll

 ; RUN: llc -mtriple=thumbv8.1m.main-arm-none-eabi -mattr=+mve.fp %s -o - | FileCheck %s
 
+declare i64 @llvm.experimental.vector.reduce.add.i64.v2i64(<2 x i64>)
 declare i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32>)
+declare i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16>)
+declare i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8>)
+
 define arm_aapcs_vfpcc i32 @vaddv_v4i32_i32(<4 x i32> %s1) {
 ; CHECK-LABEL: vaddv_v4i32_i32:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vaddv.u32 r0, q0
 ; CHECK-NEXT:    bx lr
 entry:
   %r = call i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %s1)
   ret i32 %r
 }
 
-declare i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16>)
 define arm_aapcs_vfpcc i16 @vaddv_v16i16_i16(<8 x i16> %s1) {
 ; CHECK-LABEL: vaddv_v16i16_i16:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vaddv.u16 r0, q0
 ; CHECK-NEXT:    bx lr
 entry:
   %r = call i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16> %s1)
   ret i16 %r
 }
 
-declare i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8>)
 define arm_aapcs_vfpcc i8 @vaddv_v16i8_i8(<16 x i8> %s1) {
 ; CHECK-LABEL: vaddv_v16i8_i8:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vaddv.u8 r0, q0
 ; CHECK-NEXT:    bx lr
 entry:
   %r = call i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8> %s1)
   ret i8 %r
 }
+
+define arm_aapcs_vfpcc i64 @vaddva_v2i64_i64(<2 x i64> %s1, i64 %x) {
+; CHECK-LABEL: vaddva_v2i64_i64:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    .save {r4, lr}
+; CHECK-NEXT:    push {r4, lr}
+; CHECK-NEXT:    vmov r2, s2
+; CHECK-NEXT:    vmov r3, s0
+; CHECK-NEXT:    vmov r12, s3
+; CHECK-NEXT:    vmov lr, s1
+; CHECK-NEXT:    adds r4, r3, r2
+; CHECK-NEXT:    adc.w r3, lr, r12
+; CHECK-NEXT:    mov r2, r4
+; CHECK-NEXT:    lsrl r2, r3, #32
+; CHECK-NEXT:    adds r0, r0, r4
+; CHECK-NEXT:    adcs r1, r2
+; CHECK-NEXT:    pop {r4, pc}
+entry:
+  %t = call i64 @llvm.experimental.vector.reduce.add.i64.v2i64(<2 x i64> %s1)
+  %r = add i64 %t, %x
+  ret i64 %r
+}
+
+define arm_aapcs_vfpcc i32 @vaddva_v4i32_i32(<4 x i32> %s1, i32 %x) {
+; CHECK-LABEL: vaddva_v4i32_i32:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddva.u32 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %t = call i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %s1)
+  %r = add i32 %t, %x
+  ret i32 %r
+}
+
+define arm_aapcs_vfpcc i16 @vaddva_v8i16_i16(<8 x i16> %s1, i16 %x) {
+; CHECK-LABEL: vaddva_v8i16_i16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddva.u16 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %t = call i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16> %s1)
+  %r = add i16 %t, %x
+  ret i16 %r
+}
+
+define arm_aapcs_vfpcc i8 @vaddva_v16i8_i8(<16 x i8> %s1, i8 %x) {
+; CHECK-LABEL: vaddva_v16i8_i8:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddva.u8 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %t = call i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8> %s1)
+  %r = add i8 %t, %x
+  ret i8 %r
+}