[X86] Teach combineToExtendBoolVectorInReg to create opportunities for using broadcast load instructions.

If we're inserting a scalar that is smaller than the element
size of the final VT, the value of the extra bits doesn't matter.

Previously we any_extended in the scalar domain before inserting.

This patch changes this to use a broadcast of the original
scalar type and then a bitcast to the final type. This might
enable the use of a broadcast load.

This recovers regressions from 07d68c24aa19483e44db4336b0935b00a5d69949
and 9fcd212e2f678fdbdf304399a1e58ca490dc54d1 without relying on
alignment of the load.

Differential Revision: https://reviews.llvm.org/D75835

5 files changed, 29 insertions, 19 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f351132..5c9313d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -44278,6 +44278,21 @@ combineToExtendBoolVectorInReg(SDNode *N, SelectionDAG &DAG,
 
     for (unsigned i = 0; i != Scale; ++i)
       ShuffleMask.append(EltSizeInBits, i);
+    Vec = DAG.getVectorShuffle(VT, DL, Vec, Vec, ShuffleMask);
+  } else if (Subtarget.hasAVX2() && NumElts < EltSizeInBits &&
+             (SclVT == MVT::i8 || SclVT == MVT::i16 || SclVT == MVT::i32)) {
+    // If we have register broadcast instructions, use the scalar size as the
+    // element type for the shuffle. Then cast to the wider element type. The
+    // widened bits won't be used, and this might allow the use of a broadcast
+    // load.
+    assert((EltSizeInBits % NumElts) == 0 && "Unexpected integer scale");
+    unsigned Scale = EltSizeInBits / NumElts;
+    EVT BroadcastVT =
+        EVT::getVectorVT(*DAG.getContext(), SclVT, NumElts * Scale);
+    Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, BroadcastVT, N00);
+    ShuffleMask.append(NumElts * Scale, 0);
+    Vec = DAG.getVectorShuffle(BroadcastVT, DL, Vec, Vec, ShuffleMask);
+    Vec = DAG.getBitcast(VT, Vec);
   } else {
     // For smaller scalar integers, we can simply any-extend it to the vector
     // element size (we don't care about the upper bits) and broadcast it to all
@@ -44285,8 +44300,8 @@ combineToExtendBoolVectorInReg(SDNode *N, SelectionDAG &DAG,
     SDValue Scl = DAG.getAnyExtOrTrunc(N00, DL, SVT);
     Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Scl);
     ShuffleMask.append(NumElts, 0);
+    Vec = DAG.getVectorShuffle(VT, DL, Vec, Vec, ShuffleMask);
   }
-  Vec = DAG.getVectorShuffle(VT, DL, Vec, Vec, ShuffleMask);
 
   // Now, mask the relevant bit in each element.
   SmallVector<SDValue, 32> Bits;
diff --git a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
index 85a560d..cb2af23 100644
--- a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
+++ b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-sext.ll
@@ -114,7 +114,7 @@ define <8 x i16> @ext_i8_8i16(i8 %a0) {
 ; AVX2-LABEL: ext_i8_8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
@@ -273,7 +273,7 @@ define <8 x i32> @ext_i8_8i32(i8 %a0) {
 ; AVX2-LABEL: ext_i8_8i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
@@ -456,9 +456,8 @@ define <8 x i64> @ext_i8_8i64(i8 %a0) {
 ;
 ; AVX2-LABEL: ext_i8_8i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX2-NEXT:    vmovq %rdi, %xmm0
-; AVX2-NEXT:    vpbroadcastq %xmm0, %ymm1
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm1
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm0 = [1,2,4,8]
 ; AVX2-NEXT:    vpand %ymm0, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpeqq %ymm0, %ymm2, %ymm0
@@ -525,7 +524,7 @@ define <16 x i32> @ext_i16_16i32(i16 %a0) {
 ; AVX2-LABEL: ext_i16_16i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm1
+; AVX2-NEXT:    vpbroadcastw %xmm0, %ymm1
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm0 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %ymm0, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpeqd %ymm0, %ymm2, %ymm0
diff --git a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
index d391090..ea52196 100644
--- a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
+++ b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool-zext.ll
@@ -141,7 +141,7 @@ define <8 x i16> @ext_i8_8i16(i8 %a0) {
 ; AVX2-LABEL: ext_i8_8i16:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
@@ -344,7 +344,7 @@ define <8 x i32> @ext_i8_8i32(i8 %a0) {
 ; AVX2-LABEL: ext_i8_8i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0
@@ -583,9 +583,8 @@ define <8 x i64> @ext_i8_8i64(i8 %a0) {
 ;
 ; AVX2-LABEL: ext_i8_8i64:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    # kill: def $edi killed $edi def $rdi
-; AVX2-NEXT:    vmovq %rdi, %xmm0
-; AVX2-NEXT:    vpbroadcastq %xmm0, %ymm1
+; AVX2-NEXT:    vmovd %edi, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %ymm1
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm0 = [1,2,4,8]
 ; AVX2-NEXT:    vpand %ymm0, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpeqq %ymm0, %ymm2, %ymm0
@@ -670,7 +669,7 @@ define <16 x i32> @ext_i16_16i32(i16 %a0) {
 ; AVX2-LABEL: ext_i16_16i32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm1
+; AVX2-NEXT:    vpbroadcastw %xmm0, %ymm1
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm0 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %ymm0, %ymm1, %ymm2
 ; AVX2-NEXT:    vpcmpeqd %ymm0, %ymm2, %ymm0
diff --git a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool.ll b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool.ll
index 9288335..ac2fd44 100644
--- a/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool.ll
+++ b/llvm/test/CodeGen/X86/bitcast-int-to-vector-bool.ll
@@ -115,7 +115,7 @@ define <8 x i1> @bitcast_i8_8i1(i8 zeroext %a0) {
 ; AVX2-LABEL: bitcast_i8_8i1:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovd %edi, %xmm0
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb %xmm0, %xmm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
diff --git a/llvm/test/CodeGen/X86/vector-sext.ll b/llvm/test/CodeGen/X86/vector-sext.ll
index 20c1be8..299a73f 100644
--- a/llvm/test/CodeGen/X86/vector-sext.ll
+++ b/llvm/test/CodeGen/X86/vector-sext.ll
@@ -2039,9 +2039,7 @@ define <8 x i16> @load_sext_8i1_to_8i16(<8 x i1> *%ptr) {
 ;
 ; AVX2-LABEL: load_sext_8i1_to_8i16:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    movzwl (%rdi), %eax
-; AVX2-NEXT:    vmovd %eax, %xmm0
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %xmm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} xmm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %xmm1, %xmm0, %xmm0
 ; AVX2-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
@@ -2261,8 +2259,7 @@ define <8 x i32> @load_sext_8i1_to_8i32(<8 x i1> *%ptr) {
 ;
 ; AVX2-LABEL: load_sext_8i1_to_8i32:
 ; AVX2:       # %bb.0: # %entry
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastd %xmm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %ymm0
 ; AVX2-NEXT:    vmovdqa {{.*#+}} ymm1 = [1,2,4,8,16,32,64,128]
 ; AVX2-NEXT:    vpand %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm0