[PATCH] D68763: [X86] Use packusdw+vpmovuswb to implement v16i32->V16i8 that clamps signed inputs to be between 0 and 255 when zmm registers are disabled on SKX.

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.
craig.topper added reviewers: RKSimon, spatel.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.

If we've disable zmm registers, the v16i32 will need to be split. This split will propagate through min/max the truncate. This creates two sequences that need to be concatenated back to v16i8. We can instead use packusdw to do part of the clamping, truncating, and concatenating all at once. Then we can use a vpmovuswb to finish off the clamp.


https://reviews.llvm.org/D68763

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/min-legal-vector-width.ll


Index: llvm/test/CodeGen/X86/min-legal-vector-width.ll
===================================================================
--- llvm/test/CodeGen/X86/min-legal-vector-width.ll
+++ llvm/test/CodeGen/X86/min-legal-vector-width.ll
@@ -1083,12 +1083,10 @@
 define <16 x i8> @trunc_packus_v16i32_v16i8(<16 x i32>* %p, <16 x i8>* %q) "min-legal-vector-width"="256" {
 ; CHECK-LABEL: trunc_packus_v16i32_v16i8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; CHECK-NEXT:    vpmaxsd 32(%rdi), %ymm0, %ymm1
-; CHECK-NEXT:    vpmovusdb %ymm1, %xmm1
-; CHECK-NEXT:    vpmaxsd (%rdi), %ymm0, %ymm0
-; CHECK-NEXT:    vpmovusdb %ymm0, %xmm0
-; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
+; CHECK-NEXT:    vmovdqa (%rdi), %ymm0
+; CHECK-NEXT:    vpackusdw 32(%rdi), %ymm0, %ymm0
+; CHECK-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; CHECK-NEXT:    vpmovuswb %ymm0, %xmm0
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %a = load <16 x i32>, <16 x i32>* %p
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -39841,6 +39841,21 @@
     if (auto USatVal = detectUSatPattern(In, VT, DAG, DL))
       return DAG.getNode(X86ISD::VTRUNCUS, DL, VT, USatVal);
   }
+
+  // If we're clamping a signed 32-bit vector to 0-255 and the 32-bit vector is
+  // split across two registers. We can use a packusdw+perm to clamp to 0-65535
+  // and concatenate at the same time. Then we can use a final vpmovuswb to
+  // clip to 0-255.
+  if (Subtarget.hasBWI() && !Subtarget.useAVX512Regs() &&
+      InVT == MVT::v16i32 && VT == MVT::v16i8) {
+    if (auto USatVal = detectSSatPattern(In, VT, true)) {
+      // Emit a VPACKUSDW+VPERMQ followed by a VPMOVUSWB.
+      SDValue Mid = truncateVectorWithPACK(X86ISD::PACKUS, MVT::v16i16, USatVal,
+                                           DL, DAG, Subtarget);
+      return DAG.getNode(X86ISD::VTRUNCUS, DL, VT, Mid);
+    }
+  }
+
   if (VT.isVector() && isPowerOf2_32(VT.getVectorNumElements()) &&
       !(Subtarget.hasAVX512() && InSVT == MVT::i32) &&
       !(Subtarget.hasBWI() && InSVT == MVT::i16) &&


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D68763.224258.patch
Type: text/x-patch
Size: 2231 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20191010/199f2cbc/attachment.bin>