[X86] Add test showing the failure to fold FREEZE(MOVMSK(X)) -> MOVMSK(FREEZE(X))

Author: RKSimon

llvm/test/CodeGen/X86/combine-movmsk.ll

@@ -489,6 +489,34 @@ define i32 @or_pmovmskb_pmovmskb(<16 x i8> %a0, <8 x i16> %a1) {
   ret i32 %7
 }
 
+; TODO: FREEZE(MOVMSK(X)) -> MOVMSK(FREEZE(X))
+define i32 @movmskps_freeze(<4 x i32> %a0) {
+; SSE-LABEL: movmskps_freeze:
+; SSE:       # %bb.0:
+; SSE-NEXT:    movmskps %xmm0, %eax
+; SSE-NEXT:    andl $15, %eax
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: movmskps_freeze:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vmovmskps %xmm0, %eax
+; AVX-NEXT:    andl $15, %eax
+; AVX-NEXT:    retq
+;
+; ADL-LABEL: movmskps_freeze:
+; ADL:       # %bb.0:
+; ADL-NEXT:    vmovmskps %xmm0, %eax
+; ADL-NEXT:    andl $15, %eax
+; ADL-NEXT:    retq
+  %1 = icmp slt <4 x i32> %a0, zeroinitializer
+  %2 = sext <4 x i1> %1 to <4 x i32>
+  %3 = bitcast <4 x i32> %2 to <4 x float>
+  %4 = tail call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %3)
+  %5 = freeze i32 %4
+  %6 = and i32 %5, 15
+  ret i32 %6
+}
+
 ; We can't fold to ptest if we're not checking every pcmpeq result
 define i32 @movmskps_ptest_numelts_mismatch(<16 x i8> %a0) {
 ; SSE-LABEL: movmskps_ptest_numelts_mismatch: