Added a template for building target specific memory node in DAG.

I added API for creation a target specific memory node in DAG. Today, all memory nodes are common for all targets and their constructors are located in SelectionDAG.cpp.
There are some cases in X86 where we need to create a special node - truncation-with-saturation store, float-to-half-store. 
In the current patch I added truncation-with-saturation nodes and I'm using them for intrinsics. In the future I plan to implement DAG lowering for truncation-with-saturation pattern.

Differential Revision: https://reviews.llvm.org/D27899

llvm-svn: 290250

4 files changed, 42 insertions, 44 deletions

diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics.ll b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
index 6eedd264ada..2920125cbcf 100644
--- a/llvm/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
@@ -2912,8 +2912,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qb.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qb_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qb_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqb %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovsqb %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovsqb %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovs.qb.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -2946,8 +2946,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qb.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qb_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qb_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqb %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovusqb %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovusqb %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovus.qb.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -3014,8 +3014,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qw.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qw_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qw_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqw %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovsqw %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovsqw %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovs.qw.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -3048,8 +3048,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qw.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qw_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qw_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqw %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovusqw %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovusqw %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovus.qw.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -3116,8 +3116,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qd.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qd_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qd_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqd %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovsqd %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovsqd %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovs.qd.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -3150,8 +3150,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qd.mem.512(i8* %ptr, <8 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qd_mem_512(i8* %ptr, <8 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qd_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqd %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovusqd %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovusqd %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovus.qd.mem.512(i8* %ptr, <8 x i64> %x1, i8 -1)
@@ -3218,8 +3218,8 @@ declare void @llvm.x86.avx512.mask.pmovs.db.mem.512(i8* %ptr, <16 x i32>, i16)
 define void @test_int_x86_avx512_mask_pmovs_db_mem_512(i8* %ptr, <16 x i32> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_db_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdb %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovsdb %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovsdb %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovs.db.mem.512(i8* %ptr, <16 x i32> %x1, i16 -1)
@@ -3252,8 +3252,8 @@ declare void @llvm.x86.avx512.mask.pmovus.db.mem.512(i8* %ptr, <16 x i32>, i16)
 define void @test_int_x86_avx512_mask_pmovus_db_mem_512(i8* %ptr, <16 x i32> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_db_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdb %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovusdb %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovusdb %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovus.db.mem.512(i8* %ptr, <16 x i32> %x1, i16 -1)
@@ -3320,8 +3320,8 @@ declare void @llvm.x86.avx512.mask.pmovs.dw.mem.512(i8* %ptr, <16 x i32>, i16)
 define void @test_int_x86_avx512_mask_pmovs_dw_mem_512(i8* %ptr, <16 x i32> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_dw_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdw %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovsdw %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovsdw %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovs.dw.mem.512(i8* %ptr, <16 x i32> %x1, i16 -1)
@@ -3354,8 +3354,8 @@ declare void @llvm.x86.avx512.mask.pmovus.dw.mem.512(i8* %ptr, <16 x i32>, i16)
 define void @test_int_x86_avx512_mask_pmovus_dw_mem_512(i8* %ptr, <16 x i32> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_dw_mem_512:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdw %zmm0, (%rdi)
 ; CHECK-NEXT:    kmovw %esi, %k1
+; CHECK-NEXT:    vpmovusdw %zmm0, (%rdi)
 ; CHECK-NEXT:    vpmovusdw %zmm0, (%rdi) {%k1}
 ; CHECK-NEXT:    retq
     call void @llvm.x86.avx512.mask.pmovus.dw.mem.512(i8* %ptr, <16 x i32> %x1, i16 -1)
diff --git a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
index 05b59aff9b5..d9a9a2d655b 100644
--- a/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bw-intrinsics.ll
@@ -2010,18 +2010,17 @@ declare void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16>, i32)
 define void @test_int_x86_avx512_mask_pmovs_wb_mem_512(i8* %ptr, <32 x i16> %x1, i32 %x2) {
 ; AVX512BW-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_512:
 ; AVX512BW:       ## BB#0:
-; AVX512BW-NEXT:    vpmovswb %zmm0, (%rdi)
 ; AVX512BW-NEXT:    kmovd %esi, %k1
+; AVX512BW-NEXT:    vpmovswb %zmm0, (%rdi)
 ; AVX512BW-NEXT:    vpmovswb %zmm0, (%rdi) {%k1}
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512F-32-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_512:
 ; AVX512F-32:       # BB#0:
+; AVX512F-32-NEXT:    kmovd {{[0-9]+}}(%esp), %k1
 ; AVX512F-32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; AVX512F-32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
-; AVX512F-32-NEXT:    vpmovswb %zmm0, (%ecx)
-; AVX512F-32-NEXT:    kmovd %eax, %k1
-; AVX512F-32-NEXT:    vpmovswb %zmm0, (%ecx) {%k1}
+; AVX512F-32-NEXT:    vpmovswb %zmm0, (%eax)
+; AVX512F-32-NEXT:    vpmovswb %zmm0, (%eax) {%k1}
 ; AVX512F-32-NEXT:    retl
     call void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 -1)
     call void @llvm.x86.avx512.mask.pmovs.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 %x2)
@@ -2063,18 +2062,17 @@ declare void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16>, i32)
 define void @test_int_x86_avx512_mask_pmovus_wb_mem_512(i8* %ptr, <32 x i16> %x1, i32 %x2) {
 ; AVX512BW-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_512:
 ; AVX512BW:       ## BB#0:
-; AVX512BW-NEXT:    vpmovuswb %zmm0, (%rdi)
 ; AVX512BW-NEXT:    kmovd %esi, %k1
+; AVX512BW-NEXT:    vpmovuswb %zmm0, (%rdi)
 ; AVX512BW-NEXT:    vpmovuswb %zmm0, (%rdi) {%k1}
 ; AVX512BW-NEXT:    retq
 ;
 ; AVX512F-32-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_512:
 ; AVX512F-32:       # BB#0:
+; AVX512F-32-NEXT:    kmovd {{[0-9]+}}(%esp), %k1
 ; AVX512F-32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; AVX512F-32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
-; AVX512F-32-NEXT:    vpmovuswb %zmm0, (%ecx)
-; AVX512F-32-NEXT:    kmovd %eax, %k1
-; AVX512F-32-NEXT:    vpmovuswb %zmm0, (%ecx) {%k1}
+; AVX512F-32-NEXT:    vpmovuswb %zmm0, (%eax)
+; AVX512F-32-NEXT:    vpmovuswb %zmm0, (%eax) {%k1}
 ; AVX512F-32-NEXT:    retl
     call void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 -1)
     call void @llvm.x86.avx512.mask.pmovus.wb.mem.512(i8* %ptr, <32 x i16> %x1, i32 %x2)
diff --git a/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
index 5eacf2279f2..3a56de61e5e 100644
--- a/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bwvl-intrinsics.ll
@@ -4090,8 +4090,8 @@ declare void @llvm.x86.avx512.mask.pmovs.wb.mem.128(i8* %ptr, <8 x i16>, i8)
 define void @test_int_x86_avx512_mask_pmovs_wb_mem_128(i8* %ptr, <8 x i16> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovswb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x20,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovswb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x20,0x07]
 ; CHECK-NEXT:    vpmovswb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x20,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.wb.mem.128(i8* %ptr, <8 x i16> %x1, i8 -1)
@@ -4124,8 +4124,8 @@ declare void @llvm.x86.avx512.mask.pmovus.wb.mem.128(i8* %ptr, <8 x i16>, i8)
 define void @test_int_x86_avx512_mask_pmovus_wb_mem_128(i8* %ptr, <8 x i16> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovuswb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x10,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovuswb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x10,0x07]
 ; CHECK-NEXT:    vpmovuswb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x10,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.wb.mem.128(i8* %ptr, <8 x i16> %x1, i8 -1)
@@ -4192,8 +4192,8 @@ declare void @llvm.x86.avx512.mask.pmovs.wb.mem.256(i8* %ptr, <16 x i16>, i16)
 define void @test_int_x86_avx512_mask_pmovs_wb_mem_256(i8* %ptr, <16 x i16> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_wb_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovswb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x20,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovswb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x20,0x07]
 ; CHECK-NEXT:    vpmovswb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x20,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.wb.mem.256(i8* %ptr, <16 x i16> %x1, i16 -1)
@@ -4226,8 +4226,8 @@ declare void @llvm.x86.avx512.mask.pmovus.wb.mem.256(i8* %ptr, <16 x i16>, i16)
 define void @test_int_x86_avx512_mask_pmovus_wb_mem_256(i8* %ptr, <16 x i16> %x1, i16 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_wb_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovuswb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x10,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovuswb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x10,0x07]
 ; CHECK-NEXT:    vpmovuswb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x10,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.wb.mem.256(i8* %ptr, <16 x i16> %x1, i16 -1)
diff --git a/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
index a99c8641d9c..90b9fc2192f 100644
--- a/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512vl-intrinsics.ll
@@ -2040,8 +2040,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qb.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qb_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qb_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x22,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x22,0x07]
 ; CHECK-NEXT:    vpmovsqb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x22,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qb.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2074,8 +2074,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qb.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qb_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qb_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x12,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x12,0x07]
 ; CHECK-NEXT:    vpmovusqb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x12,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qb.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2142,8 +2142,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qb.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qb_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qb_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x22,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x22,0x07]
 ; CHECK-NEXT:    vpmovsqb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x22,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qb.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2176,8 +2176,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qb.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qb_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qb_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x12,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x12,0x07]
 ; CHECK-NEXT:    vpmovusqb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x12,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qb.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2244,8 +2244,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qw.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qw_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qw_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x24,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x24,0x07]
 ; CHECK-NEXT:    vpmovsqw %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x24,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qw.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2278,8 +2278,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qw.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qw_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qw_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x14,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x14,0x07]
 ; CHECK-NEXT:    vpmovusqw %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x14,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qw.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2346,8 +2346,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qw.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qw_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qw_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x24,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x24,0x07]
 ; CHECK-NEXT:    vpmovsqw %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x24,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qw.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2380,8 +2380,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qw.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qw_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qw_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x14,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x14,0x07]
 ; CHECK-NEXT:    vpmovusqw %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x14,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qw.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2448,8 +2448,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qd.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qd_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qd_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqd %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x25,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqd %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x25,0x07]
 ; CHECK-NEXT:    vpmovsqd %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x25,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qd.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2482,8 +2482,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qd.mem.128(i8* %ptr, <2 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qd_mem_128(i8* %ptr, <2 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qd_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqd %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x15,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqd %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x15,0x07]
 ; CHECK-NEXT:    vpmovusqd %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x15,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qd.mem.128(i8* %ptr, <2 x i64> %x1, i8 -1)
@@ -2550,8 +2550,8 @@ declare void @llvm.x86.avx512.mask.pmovs.qd.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovs_qd_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_qd_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsqd %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x25,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsqd %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x25,0x07]
 ; CHECK-NEXT:    vpmovsqd %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x25,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.qd.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2584,8 +2584,8 @@ declare void @llvm.x86.avx512.mask.pmovus.qd.mem.256(i8* %ptr, <4 x i64>, i8)
 define void @test_int_x86_avx512_mask_pmovus_qd_mem_256(i8* %ptr, <4 x i64> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_qd_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusqd %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x15,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusqd %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x15,0x07]
 ; CHECK-NEXT:    vpmovusqd %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x15,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.qd.mem.256(i8* %ptr, <4 x i64> %x1, i8 -1)
@@ -2652,8 +2652,8 @@ declare void @llvm.x86.avx512.mask.pmovs.db.mem.128(i8* %ptr, <4 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovs_db_mem_128(i8* %ptr, <4 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_db_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x21,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsdb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x21,0x07]
 ; CHECK-NEXT:    vpmovsdb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x21,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.db.mem.128(i8* %ptr, <4 x i32> %x1, i8 -1)
@@ -2686,8 +2686,8 @@ declare void @llvm.x86.avx512.mask.pmovus.db.mem.128(i8* %ptr, <4 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovus_db_mem_128(i8* %ptr, <4 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_db_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x11,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusdb %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x11,0x07]
 ; CHECK-NEXT:    vpmovusdb %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x11,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.db.mem.128(i8* %ptr, <4 x i32> %x1, i8 -1)
@@ -2754,8 +2754,8 @@ declare void @llvm.x86.avx512.mask.pmovs.db.mem.256(i8* %ptr, <8 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovs_db_mem_256(i8* %ptr, <8 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_db_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x21,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsdb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x21,0x07]
 ; CHECK-NEXT:    vpmovsdb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x21,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.db.mem.256(i8* %ptr, <8 x i32> %x1, i8 -1)
@@ -2788,8 +2788,8 @@ declare void @llvm.x86.avx512.mask.pmovus.db.mem.256(i8* %ptr, <8 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovus_db_mem_256(i8* %ptr, <8 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_db_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x11,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusdb %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x11,0x07]
 ; CHECK-NEXT:    vpmovusdb %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x11,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.db.mem.256(i8* %ptr, <8 x i32> %x1, i8 -1)
@@ -2856,8 +2856,8 @@ declare void @llvm.x86.avx512.mask.pmovs.dw.mem.128(i8* %ptr, <4 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovs_dw_mem_128(i8* %ptr, <4 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_dw_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x23,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsdw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x23,0x07]
 ; CHECK-NEXT:    vpmovsdw %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x23,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.dw.mem.128(i8* %ptr, <4 x i32> %x1, i8 -1)
@@ -2890,8 +2890,8 @@ declare void @llvm.x86.avx512.mask.pmovus.dw.mem.128(i8* %ptr, <4 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovus_dw_mem_128(i8* %ptr, <4 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_dw_mem_128:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x13,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusdw %xmm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x08,0x13,0x07]
 ; CHECK-NEXT:    vpmovusdw %xmm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x09,0x13,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.dw.mem.128(i8* %ptr, <4 x i32> %x1, i8 -1)
@@ -2958,8 +2958,8 @@ declare void @llvm.x86.avx512.mask.pmovs.dw.mem.256(i8* %ptr, <8 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovs_dw_mem_256(i8* %ptr, <8 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovs_dw_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovsdw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x23,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovsdw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x23,0x07]
 ; CHECK-NEXT:    vpmovsdw %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x23,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovs.dw.mem.256(i8* %ptr, <8 x i32> %x1, i8 -1)
@@ -2992,8 +2992,8 @@ declare void @llvm.x86.avx512.mask.pmovus.dw.mem.256(i8* %ptr, <8 x i32>, i8)
 define void @test_int_x86_avx512_mask_pmovus_dw_mem_256(i8* %ptr, <8 x i32> %x1, i8 %x2) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_pmovus_dw_mem_256:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vpmovusdw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x13,0x07]
 ; CHECK-NEXT:    kmovw %esi, %k1 ## encoding: [0xc5,0xf8,0x92,0xce]
+; CHECK-NEXT:    vpmovusdw %ymm0, (%rdi) ## encoding: [0x62,0xf2,0x7e,0x28,0x13,0x07]
 ; CHECK-NEXT:    vpmovusdw %ymm0, (%rdi) {%k1} ## encoding: [0x62,0xf2,0x7e,0x29,0x13,0x07]
 ; CHECK-NEXT:    retq ## encoding: [0xc3]
     call void @llvm.x86.avx512.mask.pmovus.dw.mem.256(i8* %ptr, <8 x i32> %x1, i8 -1)