Subzero: Don't use key SSE instructions on potentially unaligned loads.

tests_lit/llvm2ice_tests/address-mode-opt.ll

 ; This file checks support for address mode optimization.
 
 ; RUN: %p2i --filetype=obj --disassemble -i %s --args -O2 \
 ; RUN:   | FileCheck %s
 ; RUN: %p2i --filetype=obj --disassemble -i %s --args -O2 -mattr=sse4.1 \
 ; RUN:   | FileCheck --check-prefix=SSE41 %s
 
 define float @load_arg_plus_200000(float* %arg) {
 entry:
   %arg.int = ptrtoint float* %arg to i32
@@ skipping 38 matching lines @@
 ; CHECK: movss xmm0,DWORD PTR [e{{..}}]
 }
 
 define <8 x i16> @load_mul_v8i16_mem(<8 x i16> %arg0, i32 %arg1_iptr) {
 entry:
   %addr_sub = sub i32 %arg1_iptr, 200000
   %addr_ptr = inttoptr i32 %addr_sub to <8 x i16>*
   %arg1 = load <8 x i16>* %addr_ptr, align 2
   %res_vec = mul <8 x i16> %arg0, %arg1
   ret <8 x i16> %res_vec
+; Address mode optimization is generally unsafe for SSE vector instructions.
 ; CHECK-LABEL: load_mul_v8i16_mem
-; CHECK: pmullw xmm{{.*}},XMMWORD PTR [e{{.*}}-0x30d40]
+; CHECK-NOT: pmullw xmm{{.*}},XMMWORD PTR [e{{..}}-0x30d40]
 }
 
 define <4 x i32> @load_mul_v4i32_mem(<4 x i32> %arg0, i32 %arg1_iptr) {
 entry:
   %addr_sub = sub i32 %arg1_iptr, 200000
   %addr_ptr = inttoptr i32 %addr_sub to <4 x i32>*
   %arg1 = load <4 x i32>* %addr_ptr, align 4
   %res = mul <4 x i32> %arg0, %arg1
   ret <4 x i32> %res
+; Address mode optimization is generally unsafe for SSE vector instructions.
 ; CHECK-LABEL: load_mul_v4i32_mem
-; CHECK: pmuludq xmm{{.*}},XMMWORD PTR [e{{.*}}-0x30d40]
+; CHECK-NOT: pmuludq xmm{{.*}},XMMWORD PTR [e{{..}}-0x30d40]
 ; CHECK: pmuludq
 ;
 ; SSE41-LABEL: load_mul_v4i32_mem
-; SSE41: pmulld xmm{{.*}},XMMWORD PTR [e{{.*}}-0x30d40]
+; SSE41-NOT: pmulld xmm{{.*}},XMMWORD PTR [e{{..}}-0x30d40]
 }
 
 define float @address_mode_opt_chaining(float* %arg) {
 entry:
   %arg.int = ptrtoint float* %arg to i32
   %addr1.int = add i32 12, %arg.int
   %addr2.int = sub i32 %addr1.int, 4
   %addr2.ptr = inttoptr i32 %addr2.int to float*
   %addr2.load = load float* %addr2.ptr, align 4
   ret float %addr2.load
@@ skipping 53 matching lines @@
 define float @address_mode_opt_sub_min_int(float* %arg) {
 entry:
   %arg.int = ptrtoint float* %arg to i32
   %addr1.int = sub i32 %arg.int, 2147483648
   %addr1.ptr = inttoptr i32 %addr1.int to float*
   %addr1.load = load float* %addr1.ptr, align 4
   ret float %addr1.load
 ; CHECK-LABEL: address_mode_opt_sub_min_int
 ; CHECK: movss xmm0,DWORD PTR [{{.*}}-0x80000000]
 }