diff --git a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 31eb6aed31c..adc433c4b70 100644
--- a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -768,6 +768,7 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
         // TODO: Could compute known zero/one bits based on the input.
         break;
       }
+      case Intrinsic::x86_mmx_pmovmskb:
       case Intrinsic::x86_sse_movmsk_ps:
       case Intrinsic::x86_sse2_movmsk_pd:
       case Intrinsic::x86_sse2_pmovmskb_128:
@@ -776,9 +777,14 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
       case Intrinsic::x86_avx2_pmovmskb: {
         // MOVMSK copies the vector elements' sign bits to the low bits
         // and zeros the high bits.
-        auto Arg = II->getArgOperand(0);
-        auto ArgType = cast<VectorType>(Arg->getType());
-        unsigned ArgWidth = ArgType->getNumElements();
+        unsigned ArgWidth;
+        if (II->getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) {
+          ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>.
+        } else {
+          auto Arg = II->getArgOperand(0);
+          auto ArgType = cast<VectorType>(Arg->getType());
+          ArgWidth = ArgType->getNumElements();
+        }
 
         // If we don't need any of low bits then return zero,
         // we know that DemandedMask is non-zero already.
diff --git a/test/Transforms/InstCombine/x86-movmsk.ll b/test/Transforms/InstCombine/x86-movmsk.ll
index 767899432b0..3b644cba8a2 100644
--- a/test/Transforms/InstCombine/x86-movmsk.ll
+++ b/test/Transforms/InstCombine/x86-movmsk.ll
@@ -7,6 +7,16 @@ target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 ; DemandedBits - MOVMSK zeros the upper bits of the result.
 ;
 
+define i32 @test_upper_x86_mmx_pmovmskb(x86_mmx %a0) {
+; CHECK-LABEL: @test_upper_x86_mmx_pmovmskb(
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %a0)
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %a0)
+  %2 = and i32 %1, 255
+  ret i32 %2
+}
+
 define i32 @test_upper_x86_sse_movmsk_ps(<4 x float> %a0) {
 ; CHECK-LABEL: @test_upper_x86_sse_movmsk_ps(
 ; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %a0)
@@ -63,6 +73,15 @@ define i32 @test_upper_x86_avx_movmsk_pd_256(<4 x double> %a0) {
 ; DemandedBits - If we don't use the lower bits then we just return zero.
 ;
 
+define i32 @test_lower_x86_mmx_pmovmskb(x86_mmx %a0) {
+; CHECK-LABEL: @test_lower_x86_mmx_pmovmskb(
+; CHECK-NEXT:    ret i32 0
+;
+  %1 = call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %a0)
+  %2 = and i32 %1, -256
+  ret i32 %2
+}
+
 define i32 @test_lower_x86_sse_movmsk_ps(<4 x float> %a0) {
 ; CHECK-LABEL: @test_lower_x86_sse_movmsk_ps(
 ; CHECK-NEXT:    ret i32 0
@@ -110,6 +129,7 @@ define i32 @test_lower_x86_avx_movmsk_pd_256(<4 x double> %a0) {
 
 ; llvm.x86.avx2.pmovmskb uses the whole of the 32-bit register.
 
+declare i32 @llvm.x86.mmx.pmovmskb(x86_mmx)
 
 declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>)
 declare i32 @llvm.x86.sse2.movmsk.pd(<2 x double>)