[X86, AVX] improve insertion into zero element of 256-bit vector

[spatel]

Summary

This patch allows AVX blend instructions to handle insertion into the low element of a 256-bit vector for the appropriate data types.

For f32, instead of:

vblendps	$1, %xmm1, %xmm0, %xmm1 ## xmm1 = xmm1[0],xmm0[1,2,3]
vblendps	$15, %ymm1, %ymm0, %ymm0 ## ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]

we get:

vblendps	$1, %ymm1, %ymm0, %ymm0 ## ymm0 = ymm1[0],ymm0[1,2,3,4,5,6,7]

For f64, instead of:

vmovsd	%xmm1, %xmm0, %xmm1     ## xmm1 = xmm1[0],xmm0[1]
vblendpd	$3, %ymm1, %ymm0, %ymm0 ## ymm0 = ymm1[0,1],ymm0[2,3]

we get:

vblendpd	$1, %ymm1, %ymm0, %ymm0 ## ymm0 = ymm1[0],ymm0[1,2,3]

For the hardware-neglected integer data types, I left a TODO comment in the code and added regression tests for a follow-on patch.

[nadav]

This looks good to me.

[spatel]

Thanks, Nadav! Checked in at r233199.

lib/Target/X86/X86ISelLowering.cpp

SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
   if (!isa<ConstantSDNode>(N2))
     return SDValue();
   auto *N2C = cast<ConstantSDNode>(N2);
   unsigned IdxVal = N2C->getZExtValue();
 
   // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
   // into that, and then insert the subvector back into the result.
   if (VT.is256BitVector() || VT.is512BitVector()) {
+    // With a 256-bit vector, we can insert into the zero element efficiently
+    // using a blend if we have AVX or AVX2 and the right data type.
+    if (VT.is256BitVector() && IdxVal == 0) {
+      // TODO: It is worthwhile to cast integer to floating point and back
+      // and incur a domain crossing penalty if that's what we'll end up
+      // doing anyway after extracting to a 128-bit vector.
+      if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
+          (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
+        SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
+        N2 = DAG.getIntPtrConstant(1);
+        return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
+      }
+    }
+    
     // Get the desired 128-bit vector chunk.
     SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
 
     // Insert the element into the desired chunk.
     unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
     unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
 
     V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,

test/CodeGen/X86/avx-insertelt.ll

+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx  | FileCheck %s --check-prefix=ALL --check-prefix=AVX
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=ALL --check-prefix=AVX2
+
+define <8 x float> @insert_f32(<8 x float> %y, float %f, <8 x float> %x) {
+; ALL-LABEL: insert_f32:
+; ALL:         vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6,7]
+; ALL-NEXT:    retq
+  %i0 = insertelement <8 x float> %y, float %f, i32 0
+  ret <8 x float> %i0
+}
+
+define <4 x double> @insert_f64(<4 x double> %y, double %f, <4 x double> %x) {
+; ALL-LABEL: insert_f64:
+; ALL:         vblendpd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3]
+; ALL-NEXT:    retq
+  %i0 = insertelement <4 x double> %y, double %f, i32 0
+  ret <4 x double> %i0
+}
+
+define <32 x i8> @insert_i8(<32 x i8> %y, i8 %f, <32 x i8> %x) {
+; AVX-LABEL: insert_i8:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpinsrb $0, %edi, %xmm0, %xmm1
+; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX-NEXT:    retq
+;
+; AVX2-LABEL: insert_i8:
+; AVX2:       # BB#0:
+; AVX2-NEXT:    vpinsrb $0, %edi, %xmm0, %xmm1
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    retq
+  %i0 = insertelement <32 x i8> %y, i8 %f, i32 0
+  ret <32 x i8> %i0
+}
+
+define <16 x i16> @insert_i16(<16 x i16> %y, i16 %f, <16 x i16> %x) {
+; AVX-LABEL: insert_i16:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpinsrw $0, %edi, %xmm0, %xmm1
+; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX-NEXT:    retq
+;
+; AVX2-LABEL: insert_i16:
+; AVX2:       # BB#0:
+; AVX2-NEXT:    vpinsrw $0, %edi, %xmm0, %xmm1
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    retq
+  %i0 = insertelement <16 x i16> %y, i16 %f, i32 0
+  ret <16 x i16> %i0
+}
+
+define <8 x i32> @insert_i32(<8 x i32> %y, i32 %f, <8 x i32> %x) {
+; AVX-LABEL: insert_i32:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpinsrd $0, %edi, %xmm0, %xmm1
+; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX-NEXT:    retq
+;
+; AVX2-LABEL: insert_i32:
+; AVX2:       # BB#0:
+; AVX2-NEXT:    vmovd %edi, %xmm1
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6,7]
+; AVX2-NEXT:    retq
+  %i0 = insertelement <8 x i32> %y, i32 %f, i32 0
+  ret <8 x i32> %i0
+}
+
+define <4 x i64> @insert_i64(<4 x i64> %y, i64 %f, <4 x i64> %x) {
+; AVX-LABEL: insert_i64:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpinsrq $0, %rdi, %xmm0, %xmm1
+; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX-NEXT:    retq
+;
+; AVX2-LABEL: insert_i64:
+; AVX2:       # BB#0:
+; AVX2-NEXT:    vpinsrq $0, %rdi, %xmm0, %xmm1
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    retq
+  %i0 = insertelement <4 x i64> %y, i64 %f, i32 0
+  ret <4 x i64> %i0
+}
+