Subzero: Add support for SSE4.1 instructions.

tests_lit/llvm2ice_tests/vector-arith.ll

 ; This test checks support for vector arithmetic.
 
 ; RUN: %llvm2ice -O2 --verbose none %s | FileCheck %s
 ; RUN: %llvm2ice -Om1 --verbose none %s | FileCheck %s
+; RUN: %llvm2ice -O2 -mattr=sse4.1 --verbose none %s \
+; RUN:                | FileCheck %s --check-prefix=SSE41
+; RUN: %llvm2ice -Om1 -mattr=sse4.1 --verbose none %s \
+; RUN:                | FileCheck %s --check-prefix=SSE41
 ; RUN: %llvm2ice -O2 --verbose none %s \
 ; RUN:               | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj
 ; RUN: %llvm2ice -Om1 --verbose none %s \
 ; RUN:               | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj
+; RUN: %llvm2ice -O2 -mattr=sse4.1 --verbose none %s \
+; RUN:               | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj
+; RUN: %llvm2ice -Om1 -mattr=sse4.1 --verbose none %s \
+; RUN:               | llvm-mc -arch=x86 -x86-asm-syntax=intel -filetype=obj
 ; RUN: %llvm2ice --verbose none %s | FileCheck --check-prefix=ERRORS %s
 ; RUN: %llvm2iceinsts %s | %szdiff %s | FileCheck --check-prefix=DUMP %s
 ; RUN: %llvm2iceinsts --pnacl %s | %szdiff %s \
 ; RUN:                           | FileCheck --check-prefix=DUMP %s
 
 define <4 x float> @test_fadd(<4 x float> %arg0, <4 x float> %arg1) {
 entry:
   %res = fadd <4 x float> %arg0, %arg1
   ret <4 x float> %res
 ; CHECK-LABEL: test_fadd:
@@ skipping 280 matching lines @@
 ; CHECK: psubd
 }
 
 define <4 x i32> @test_mul_v4i32(<4 x i32> %arg0, <4 x i32> %arg1) {
 entry:
   %res = mul <4 x i32> %arg0, %arg1
   ret <4 x i32> %res
 ; CHECK-LABEL: test_mul_v4i32:
 ; CHECK: pmuludq
 ; CHECK: pmuludq
+;
+; SSE41-LABEL: test_mul_v4i32:
+; SSE41: pmulld
 }
 
 define <4 x i32> @test_shl_v4i32(<4 x i32> %arg0, <4 x i32> %arg1) {
 entry:
   %res = shl <4 x i32> %arg0, %arg1
   ret <4 x i32> %res
 ; CHECK-LABEL: test_shl_v4i32:
 ; CHECK: Sz_shl_v4i32
+
+; This line is to ensure that pmulld is generated in test_mul_v4i32 above.
+; SSE41-LABEL: test_shl_v4i32:
 }
 
 define <4 x i32> @test_lshr_v4i32(<4 x i32> %arg0, <4 x i32> %arg1) {
 entry:
   %res = lshr <4 x i32> %arg0, %arg1
   ret <4 x i32> %res
 ; CHECK-LABEL: test_lshr_v4i32:
 ; CHECK: Sz_lshr_v4i32
 }
 
@@ skipping 32 matching lines @@
 define <4 x i32> @test_srem_v4i32(<4 x i32> %arg0, <4 x i32> %arg1) {
 entry:
   %res = srem <4 x i32> %arg0, %arg1
   ret <4 x i32> %res
 ; CHECK-LABEL: test_srem_v4i32:
 ; CHECK: Sz_srem_v4i32
 }
 
 ; ERRORS-NOT: ICE translation error
 ; DUMP-NOT: SZ