[X86] Add DAG combine to combine a v8i32->v8i16 truncate with a packuswb that truncates v8i16->v8i8.

[craig.topper]

Summary

Under -x86-experimental-vector-widening-legalization, fp_to_uint/fp_to_sint with a smaller than 128 bit vector type results are custom type legalized by promoting the result to a 128 bit vector by promoting the elements, inserting an assertzext/assertsext, then truncating back to original type. The truncate will be further legalizdd to a pack shuffle. In the case of a v8i8 result type, we'll end up with a v8i16 fp_to_sint. This will need to be further legalized during vector op legalization by promoting to v8i32 and then truncating again. Under avx2 this produces good code with two pack instructions, but Under avx512 this will result in a truncate instruction and a packuswb instruction. But we should be able to get away with a single truncate instruction.

The other option is to promote all the way to vXi32 result type during the first type legalization. But in some experimentation that seemed to require more work to produce good code for other configurations.

[craig.topper]

Add a DAG combine to avoid this failing after r347593. Since we now have an AssertSExt and a stronger AssertZExt sandwiched around the truncate.

[craig.topper]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

9082

I wonder if we could just float all Asserts above truncates? And more aggressively merge adjacent asserts?

[RKSimon]

lib/Target/X86/X86ISelLowering.cpp

35383

Can the PACKSS equivalent occur as well?

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

35383

It doesn't occur in any of our lit tests so I'm not sure.

[RKSimon]

LGTM with one minor, cheers.

lib/Target/X86/X86ISelLowering.cpp

35383

OK - please can you add a TODO comment for now?

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
     SDLoc DL(N);
     EVT MinAssertVT = AssertVT.bitsLT(BigA_AssertVT) ? AssertVT : BigA_AssertVT;
     SDValue MinAssertVTVal = DAG.getValueType(MinAssertVT);
     SDValue NewAssert = DAG.getNode(Opcode, DL, BigA.getValueType(),
                                     BigA.getOperand(0), MinAssertVTVal);
     return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewAssert);
   }
 
+  // If we have (AssertZext (truncate (AssertSext X, iX)), iY) and Y is smaller
+  // than X. Just move the AssertZext in front of the truncate and drop the
+  // AssertSExt.
+  if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() &&
+      N0.getOperand(0).getOpcode() == ISD::AssertSext &&
+      Opcode == ISD::AssertZext) {
+    SDValue BigA = N0.getOperand(0);
+    EVT BigA_AssertVT = cast<VTSDNode>(BigA.getOperand(1))->getVT();
+    assert(BigA_AssertVT.bitsLE(N0.getValueType()) &&
+           "Asserting zero/sign-extended bits to a type larger than the "
+           "truncated destination does not provide information");
+
+    if (AssertVT.bitsLT(BigA_AssertVT)) {
+      SDLoc DL(N);
+      SDValue NewAssert = DAG.getNode(Opcode, DL, BigA.getValueType(),
+                                      BigA.getOperand(0), N1);
+      return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewAssert);
+    }
+  }
+
   return SDValue();
 }
 
 /// If the result of a wider load is shifted to right of N  bits and then
 /// truncated to a narrower type and where N is a multiple of number of bits of
 /// the narrower type, transform it to a narrower load from address + N / num of
 /// bits of new type. Also narrow the load if the result is masked with an AND
 /// to effectively produce a smaller type. If the result is to be extended, also

lib/Target/X86/X86ISelLowering.cpp

for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
         }
         Bits[Lane * NumDstEltsPerLane + Elt] = Val;
       }
     }
 
     return getConstVector(Bits, Undefs, VT.getSimpleVT(), DAG, SDLoc(N));
   }
 
+  // Try to combine a PACKUSWB implemented truncate with a regular truncate to
+  // create a larger truncate.
+  if (Subtarget.hasAVX512() && Opcode == X86ISD::PACKUS &&
+      N0.getOpcode() == ISD::TRUNCATE && N1.isUndef() && VT == MVT::v16i8 &&
+      N0.getOperand(0).getValueType() == MVT::v8i32) {
+
+    APInt ZeroMask = APInt::getHighBitsSet(16, 8);
+    if (DAG.MaskedValueIsZero(N0, ZeroMask)) {
+      if (Subtarget.hasVLX())
+        return DAG.getNode(X86ISD::VTRUNC, SDLoc(N), VT, N0.getOperand(0));
+
+      // Widen input to v16i32 so we can truncate that.
+      SDLoc dl(N);
+      SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v16i32,
+                                   N0.getOperand(0), DAG.getUNDEF(MVT::v8i32));
+      return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Concat);
+    }
+  }
+
   // Attempt to combine as shuffle.
   SDValue Op(N, 0);
   if (SDValue Res =
           combineX86ShufflesRecursively({Op}, 0, Op, {0}, {}, /*Depth*/ 1,
                                         /*HasVarMask*/ false,
                                         /*AllowVarMask*/ true, DAG, Subtarget))
     return Res;
 

test/CodeGen/X86/avx512-cvt-widen.ll

; VL-NEXT:    retq
   %res = fptoui <8 x double> %f to <8 x i16>
   ret <8 x i16> %res
 }
 
 define <8 x i8> @f64to8uc(<8 x double> %f) {
 ; NOVL-LABEL: f64to8uc:
 ; NOVL:       # %bb.0:
 ; NOVL-NEXT:    vcvttpd2dq %zmm0, %ymm0
-; NOVL-NEXT:    vpmovdw %zmm0, %ymm0
-; NOVL-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
+; NOVL-NEXT:    vpmovdb %zmm0, %xmm0
 ; NOVL-NEXT:    vzeroupper
 ; NOVL-NEXT:    retq
 ;
 ; VL-LABEL: f64to8uc:
 ; VL:       # %bb.0:
 ; VL-NEXT:    vcvttpd2dq %zmm0, %ymm0
-; VL-NEXT:    vpmovdw %ymm0, %xmm0
-; VL-NEXT:    vpackuswb %xmm0, %xmm0, %xmm0
+; VL-NEXT:    vpmovdb %ymm0, %xmm0
 ; VL-NEXT:    vzeroupper
 ; VL-NEXT:    retq
   %res = fptoui <8 x double> %f to <8 x i8>
   ret <8 x i8> %res
 }
 
 define <4 x i32> @f64to4ui(<4 x double> %a) nounwind {
 ; NOVL-LABEL: f64to4ui: