[X86] Allow pack instructions to be used for 512->256 truncates when -mprefer-vector-width=256 is causing 512-bit vectors to be split

[craig.topper]

Summary

If we're splitting the 512-bit vector anyway and we have zero/sign bits, then we might as well use pack instructions to concat and truncate at once.

[RKSimon]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

 }
 
 /// This function transforms vector truncation of 'extended sign-bits' or
 /// 'extended zero-bits' values.
 /// vXi16/vXi32/vXi64 to vXi8/vXi16/vXi32 into X86ISD::PACKSS/PACKUS operations.
 static SDValue combineVectorSignBitsTruncation(SDNode *N, const SDLoc &DL,
                                                SelectionDAG &DAG,
                                                const X86Subtarget &Subtarget) {
-  // Requires SSE2 but AVX512 has fast truncate.
-  if (!Subtarget.hasSSE2() || Subtarget.hasAVX512())
+  // Requires SSE2.
+  if (!Subtarget.hasSSE2())
     return SDValue();
 
   if (!N->getValueType(0).isVector() || !N->getValueType(0).isSimple())
     return SDValue();
 
   SDValue In = N->getOperand(0);
   if (!In.getValueType().isSimple())
     return SDValue();
 
   MVT VT = N->getValueType(0).getSimpleVT();
   MVT SVT = VT.getScalarType();
 
   MVT InVT = In.getValueType().getSimpleVT();
   MVT InSVT = InVT.getScalarType();
 
   // Check we have a truncation suited for PACKSS/PACKUS.
   if (!VT.is128BitVector() && !VT.is256BitVector())
     return SDValue();
   if (SVT != MVT::i8 && SVT != MVT::i16 && SVT != MVT::i32)
     return SDValue();
   if (InSVT != MVT::i16 && InSVT != MVT::i32 && InSVT != MVT::i64)
     return SDValue();
 
+  // AVX512 has fast truncate, but if the input is already going to be split,
+  // there's no harm in trying pack.
+  if (Subtarget.hasAVX512() &&
+      !(!Subtarget.useAVX512Regs() && VT.is256BitVector() &&
+        InVT.is512BitVector()))
+    return SDValue();
+
   unsigned NumPackedSignBits = std::min<unsigned>(SVT.getSizeInBits(), 16);
   unsigned NumPackedZeroBits = Subtarget.hasSSE41() ? NumPackedSignBits : 8;
 
   // Use PACKUS if the input has zero-bits that extend all the way to the
   // packed/truncated value. e.g. masks, zext_in_reg, etc.
   KnownBits Known = DAG.computeKnownBits(In);
   unsigned NumLeadingZeroBits = Known.countMinLeadingZeros();
   if (NumLeadingZeroBits >= (InSVT.getSizeInBits() - NumPackedZeroBits))

llvm/trunk/test/CodeGen/X86/min-legal-vector-width.ll

; CHECK-NEXT:    retq
   ret <8 x i16> %b
 }
 
 define <8 x i32> @trunc_v8i64_v8i32_zeroes(<8 x i64>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v8i64_v8i32_zeroes:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsrlq $48, 32(%rdi), %ymm0
 ; CHECK-NEXT:    vpsrlq $48, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovqd %ymm1, %xmm1
-; CHECK-NEXT:    vpmovqd %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpackusdw %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-NEXT:    retq
   %a = load <8 x i64>, <8 x i64>* %x
   %b = lshr <8 x i64> %a, <i64 48, i64 48, i64 48, i64 48, i64 48, i64 48, i64 48, i64 48>
   %c = trunc <8 x i64> %b to <8 x i32>
   ret <8 x i32> %c
 }
 
 define <16 x i16> @trunc_v16i32_v16i16_zeroes(<16 x i32>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v16i32_v16i16_zeroes:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpsrld $16, 32(%rdi), %ymm0
-; CHECK-NEXT:    vpsrld $16, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovdw %ymm1, %xmm1
-; CHECK-NEXT:    vpmovdw %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vmovdqa (%rdi), %ymm1
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm0 = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31]
+; CHECK-NEXT:    vpermi2w 32(%rdi), %ymm1, %ymm0
 ; CHECK-NEXT:    retq
   %a = load <16 x i32>, <16 x i32>* %x
   %b = lshr <16 x i32> %a, <i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16>
   %c = trunc <16 x i32> %b to <16 x i16>
   ret <16 x i16> %c
 }
 
 define <32 x i8> @trunc_v32i16_v32i8_zeroes(<32 x i16>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v32i16_v32i8_zeroes:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsrlw $8, 32(%rdi), %ymm0
 ; CHECK-NEXT:    vpsrlw $8, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovwb %ymm1, %xmm1
-; CHECK-NEXT:    vpmovwb %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpackuswb %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-NEXT:    retq
   %a = load <32 x i16>, <32 x i16>* %x
   %b = lshr <32 x i16> %a, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
   %c = trunc <32 x i16> %b to <32 x i8>
   ret <32 x i8> %c
 }
 
 define <8 x i32> @trunc_v8i64_v8i32_sign(<8 x i64>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v8i64_v8i32_sign:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsraq $48, 32(%rdi), %ymm0
 ; CHECK-NEXT:    vpsraq $48, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovqd %ymm1, %xmm1
-; CHECK-NEXT:    vpmovqd %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpackssdw %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-NEXT:    retq
   %a = load <8 x i64>, <8 x i64>* %x
   %b = ashr <8 x i64> %a, <i64 48, i64 48, i64 48, i64 48, i64 48, i64 48, i64 48, i64 48>
   %c = trunc <8 x i64> %b to <8 x i32>
   ret <8 x i32> %c
 }
 
 define <16 x i16> @trunc_v16i32_v16i16_sign(<16 x i32>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v16i32_v16i16_sign:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsrad $16, 32(%rdi), %ymm0
 ; CHECK-NEXT:    vpsrad $16, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovdw %ymm1, %xmm1
-; CHECK-NEXT:    vpmovdw %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpackssdw %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-NEXT:    retq
   %a = load <16 x i32>, <16 x i32>* %x
   %b = ashr <16 x i32> %a, <i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16>
   %c = trunc <16 x i32> %b to <16 x i16>
   ret <16 x i16> %c
 }
 
 define <32 x i8> @trunc_v32i16_v32i8_sign(<32 x i16>* %x) nounwind "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_v32i16_v32i8_sign:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vpsraw $8, 32(%rdi), %ymm0
 ; CHECK-NEXT:    vpsraw $8, (%rdi), %ymm1
-; CHECK-NEXT:    vpmovwb %ymm1, %xmm1
-; CHECK-NEXT:    vpmovwb %ymm0, %xmm0
-; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpacksswb %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
 ; CHECK-NEXT:    retq
   %a = load <32 x i16>, <32 x i16>* %x
   %b = ashr <32 x i16> %a, <i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8, i16 8>
   %c = trunc <32 x i16> %b to <32 x i8>
   ret <32 x i8> %c
 }