[SelectionDAG] Teach computeKnownBits some improvements to ISD::SRL with a non-splat constant shift amount.

If we have a non-splat constant shift amount, the minimum shift amount can be used to infer the number of zero upper bits of the result. There's probably a lot more that we can do here, but this
fixes a case where I wanted to infer the sign bit as zero when all the shift amounts are non-zero.

llvm-svn: 319639

2 files changed, 25 insertions, 6 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 0b59af2fa10..bcc972e0b4f 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2478,6 +2478,25 @@ void SelectionDAG::computeKnownBits(SDValue Op, KnownBits &Known,
       Known.One.lshrInPlace(Shift);
       // High bits are known zero.
       Known.Zero.setHighBits(Shift);
+    } else if (auto *BV = dyn_cast<BuildVectorSDNode>(Op.getOperand(1))) {
+      // If the shift amount is a vector of constants see if we can bound
+      // the number of upper zero bits.
+      unsigned ShiftAmountMin = BitWidth;
+      for (unsigned i = 0; i != BV->getNumOperands(); ++i) {
+        if (auto *C = dyn_cast<ConstantSDNode>(BV->getOperand(i))) {
+          const APInt &ShAmt = C->getAPIntValue();
+          if (ShAmt.ult(BitWidth)) {
+            ShiftAmountMin = std::min<unsigned>(ShiftAmountMin,
+                                                ShAmt.getZExtValue());
+            continue;
+          }
+        }
+        // Don't know anything.
+        ShiftAmountMin = 0;
+        break;
+      }
+
+      Known.Zero.setHighBits(ShiftAmountMin);
     }
     break;
   case ISD::SRA:
diff --git a/llvm/test/CodeGen/X86/combine-srl.ll b/llvm/test/CodeGen/X86/combine-srl.ll
index c5f03dbd5a3..9be85422731 100644
--- a/llvm/test/CodeGen/X86/combine-srl.ll
+++ b/llvm/test/CodeGen/X86/combine-srl.ll
@@ -257,13 +257,13 @@ define <4 x i32> @combine_vec_lshr_trunc_lshr_zero1(<4 x i64> %x) {
 ; SSE-NEXT:    psrlq $49, %xmm2
 ; SSE-NEXT:    psrlq $48, %xmm0
 ; SSE-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
-; SSE-NEXT:    movaps %xmm0, %xmm1
+; SSE-NEXT:    packusdw %xmm1, %xmm0
+; SSE-NEXT:    movdqa %xmm0, %xmm1
 ; SSE-NEXT:    psrld $27, %xmm1
-; SSE-NEXT:    movaps %xmm0, %xmm2
+; SSE-NEXT:    movdqa %xmm0, %xmm2
 ; SSE-NEXT:    psrld $25, %xmm2
 ; SSE-NEXT:    pblendw {{.*#+}} xmm2 = xmm2[0,1,2,3],xmm1[4,5,6,7]
-; SSE-NEXT:    movaps %xmm0, %xmm1
+; SSE-NEXT:    movdqa %xmm0, %xmm1
 ; SSE-NEXT:    psrld $26, %xmm1
 ; SSE-NEXT:    psrld $24, %xmm0
 ; SSE-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm1[4,5,6,7]
@@ -273,8 +273,8 @@ define <4 x i32> @combine_vec_lshr_trunc_lshr_zero1(<4 x i64> %x) {
 ; AVX-LABEL: combine_vec_lshr_trunc_lshr_zero1:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vpsrlvq {{.*}}(%rip), %ymm0, %ymm0
-; AVX-NEXT:    vpshufd {{.*#+}} ymm0 = ymm0[0,2,2,3,4,6,6,7]
-; AVX-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,2,3]
+; AVX-NEXT:    vextracti128 $1, %ymm0, %xmm1
+; AVX-NEXT:    vpackusdw %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    vpsrlvd {{.*}}(%rip), %xmm0, %xmm0
 ; AVX-NEXT:    vzeroupper
 ; AVX-NEXT:    retq