[X86] Add tests showing failure to fold mul(abs(x),abs(x)) -> mul(x,x…

…) (PR39476)
llvm · May 3, 2020 · ff5094c · ff5094c
1 parent 667f558
commit ff5094c
Showing 1 changed file with 71 additions and 0 deletions.
diff --git a/llvm/test/CodeGen/X86/combine-mul.ll b/llvm/test/CodeGen/X86/combine-mul.ll
@@ -283,6 +283,77 @@ define <4 x i32> @combine_vec_mul_add(<4 x i32> %x) {
   ret <4 x i32> %2
 }
 
+; TODO fold mul(abs(x),abs(x)) -> mul(x,x)
+
+define i31 @combine_mul_abs_i31(i31 %0) {
+; SSE-LABEL: combine_mul_abs_i31:
+; SSE:       # %bb.0:
+; SSE-NEXT:    addl %edi, %edi
+; SSE-NEXT:    sarl %edi
+; SSE-NEXT:    movl %edi, %eax
+; SSE-NEXT:    negl %eax
+; SSE-NEXT:    cmovll %edi, %eax
+; SSE-NEXT:    imull %eax, %eax
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: combine_mul_abs_i31:
+; AVX:       # %bb.0:
+; AVX-NEXT:    addl %edi, %edi
+; AVX-NEXT:    sarl %edi
+; AVX-NEXT:    movl %edi, %eax
+; AVX-NEXT:    negl %eax
+; AVX-NEXT:    cmovll %edi, %eax
+; AVX-NEXT:    imull %eax, %eax
+; AVX-NEXT:    retq
+  %c = icmp slt i31 %0, 0
+  %s = sub nsw i31 0, %0
+  %r = select i1 %c, i31 %s, i31 %0
+  %m = mul i31 %r, %r
+  ret i31 %m
+}
+
+define i32 @combine_mul_abs_i32(i32 %0) {
+; SSE-LABEL: combine_mul_abs_i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    movl %edi, %eax
+; SSE-NEXT:    negl %eax
+; SSE-NEXT:    cmovll %edi, %eax
+; SSE-NEXT:    imull %eax, %eax
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: combine_mul_abs_i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    movl %edi, %eax
+; AVX-NEXT:    negl %eax
+; AVX-NEXT:    cmovll %edi, %eax
+; AVX-NEXT:    imull %eax, %eax
+; AVX-NEXT:    retq
+  %c = icmp slt i32 %0, 0
+  %s = sub nsw i32 0, %0
+  %r = select i1 %c, i32 %s, i32 %0
+  %m = mul i32 %r, %r
+  ret i32 %m
+}
+
+define <4 x i32> @combine_mul_abs_v4i32(<4 x i32> %0) {
+; SSE-LABEL: combine_mul_abs_v4i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    pabsd %xmm0, %xmm0
+; SSE-NEXT:    pmulld %xmm0, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: combine_mul_abs_v4i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpabsd %xmm0, %xmm0
+; AVX-NEXT:    vpmulld %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %c = icmp slt <4 x i32> %0, zeroinitializer
+  %s = sub nsw <4 x i32> zeroinitializer, %0
+  %r = select <4 x i1> %c, <4 x i32> %s, <4 x i32> %0
+  %m = mul <4 x i32> %r, %r
+  ret <4 x i32> %m
+}
+
 ; This would infinite loop because DAGCombiner wants to turn this into a shift,
 ; but x86 lowering wants to avoid non-uniform vector shift amounts.