Fix SSE Instructions for XED v2022.04.17 (#597)

* added logging

* add logging

* add op corresponding to read

* same fix for movhps

* add fillers to MOVLPD

* add filler to MOVLPS

* add filler to CVTSD2SS

* filler for CVTSS2SD

* remove logging

* bump cxx-common

* drop llvm-12 support

* drop in other runs

* add dummy read to MOVHLPS

* change ISEL to match size from xed update

* fix MOVLHPS

* fix VMOVDDUP_XMMdq_XMMdq size

Author: 2over12

.github/workflows/ci.yml

@@ -23,7 +23,7 @@ jobs:
       matrix:
         image:
           - { name: 'ubuntu', tag: '20.04' }
-        llvm: ['12', '13']
+        llvm: ['13']
 
     runs-on: ubuntu-20.04
     container:
@@ -94,7 +94,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ['macos-11']
-        llvm: ['12', '13']
+        llvm: ['13']
 
     runs-on: ${{ matrix.os }}
 
@@ -215,7 +215,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        llvm: ["12", "13"]
+        llvm: ["13"]
         ubuntu: ["20.04"]
     steps:
     - uses: actions/checkout@v2

lib/Arch/X86/Semantics/CONVERT.cpp

@@ -307,7 +307,7 @@ IF_AVX(IF_64BIT(
 namespace {
 
 template <typename S1>
-DEF_SEM(CVTSD2SS, V128W dst, S1 src) {
+DEF_SEM(CVTSD2SS, V128W dst, V128 _nop_read, S1 src) {
   FWriteV32(dst, FInsertV32(FReadV32(dst), 0,
                             Float32(FExtractV64(FReadV64(src), 0))));
   return memory;
@@ -349,7 +349,7 @@ DEF_SEM(CVTSI2SD, V128W dst, V128 src1, S2 src2) {
 }
 
 template <typename S2>
-DEF_SEM(CVTSS2SD, VV128W dst_src1, S2 src2) {
+DEF_SEM(CVTSS2SD, VV128W dst_src1, V128 _nop_read, S2 src2) {
   auto src1_vec = FReadV64(dst_src1);
   auto src2_vec = FReadV32(src2);
   auto conv_val = Float64(FExtractV32(src2_vec, 0));

lib/Arch/X86/Semantics/DATAXFER.cpp

@@ -84,16 +84,16 @@ DEF_SEM(MOVDQx, D dst, S src) {
   return memory;
 }
 
-template <typename D, typename S>
-DEF_SEM(MOVLPS, D dst, S src) {
+template <typename D, typename S, typename... Fs>
+DEF_SEM(MOVLPS, D dst, Fs... _nop_fillers, S src) {
   auto src_vec = FReadV32(src);
   auto low1 = FExtractV32(src_vec, 0);
   auto low2 = FExtractV32(src_vec, 1);
   FWriteV32(dst, FInsertV32(FInsertV32(FReadV32(dst), 0, low1), 1, low2));
   return memory;
 }
 
-DEF_SEM(MOVLHPS, V128W dst, V128 src) {
+DEF_SEM(MOVLHPS, V128W dst, V128 _noop_read, V128 src) {
   auto res = FReadV32(dst);
   auto src1 = FReadV32(src);
   res = FInsertV32(res, 2, FExtractV32(src1, 0));
@@ -102,7 +102,7 @@ DEF_SEM(MOVLHPS, V128W dst, V128 src) {
   return memory;
 }
 
-DEF_SEM(MOVHLPS, V128W dst, V128 src) {
+DEF_SEM(MOVHLPS, V128W dst, V128 _nop_read, V128 src) {
   auto res = FReadV32(dst);
   auto src1 = FReadV32(src);
   res = FInsertV32(res, 0, FExtractV32(src1, 2));
@@ -111,8 +111,8 @@ DEF_SEM(MOVHLPS, V128W dst, V128 src) {
   return memory;
 }
 
-template <typename D, typename S>
-DEF_SEM(MOVLPD, D dst, S src) {
+template <typename D, typename S, typename... Fs>
+DEF_SEM(MOVLPD, D dst, Fs... fargs, S src) {
   FWriteV64(dst, FInsertV64(FReadV64(dst), 0, FExtractV64(FReadV64(src), 0)));
   return memory;
 }
@@ -466,8 +466,8 @@ DEF_ISEL(VMOVDQA_MEMqq_YMMqq) = MOVDQx<MV256W, VV256>;
 DEF_ISEL(VMOVDQA_YMMqq_YMMqq_7F) = MOVDQx<VV256W, VV256>;
 #endif  // HAS_FEATURE_AVX
 
-DEF_ISEL(MOVLPS_MEMq_XMMps) = MOVLPS<MV64W, V128>;
-DEF_ISEL(MOVLPS_XMMq_MEMq) = MOVLPS<V128W, MV64>;
+DEF_ISEL(MOVLPS_MEMq_XMMq) = MOVLPS<MV64W, V128>;
+DEF_ISEL(MOVLPS_XMMq_MEMq) = MOVLPS<V128W, MV64, V128>;
 IF_AVX(DEF_ISEL(VMOVLPS_MEMq_XMMq) = MOVLPS<MV64W, VV128>;)
 IF_AVX(DEF_ISEL(VMOVLPS_XMMdq_XMMdq_MEMq) = VMOVLPS;)
 
@@ -487,7 +487,7 @@ IF_AVX(DEF_ISEL(VMOVLHPS_XMMdq_XMMdq_XMMdq) = VMOVLHPS;)
 #  endif  // HAS_FEATURE_AVX512
 #endif  // HAS_FEATURE_AVX
 
-DEF_ISEL(MOVLPD_XMMsd_MEMq) = MOVLPD<V128W, MV64>;
+DEF_ISEL(MOVLPD_XMMsd_MEMq) = MOVLPD<V128W, MV64, V128>;
 DEF_ISEL(MOVLPD_MEMq_XMMsd) = MOVLPD<MV64W, V128>;
 IF_AVX(DEF_ISEL(VMOVLPD_MEMq_XMMq) = MOVLPD<MV64W, VV128>;)
 IF_AVX(DEF_ISEL(VMOVLPD_XMMdq_XMMdq_MEMq) = VMOVLPD;)
@@ -585,7 +585,8 @@ DEF_ISEL(MOVNTSS_MEMd_XMMd) = MOVSS_MEM<MV32W, V128>;
 
 namespace {
 
-DEF_SEM(MOVHPD, V128W dst, MV64 src) {
+// NOTE(Ian): Latest xed adds a read operand referring to the read of lower bits.
+DEF_SEM(MOVHPD, V128W dst, V128 _nop_read, MV64 src) {
   FWriteV64(dst, FInsertV64(FReadV64(dst), 1, FExtractV64(FReadV64(src), 0)));
   return memory;
 }
@@ -616,7 +617,8 @@ IF_AVX(DEF_ISEL(VMOVHPD_MEMq_XMMdq) = MOVHPD_STORE;)
 
 namespace {
 
-DEF_SEM(MOVHPS, V128W dst, MV64 src) {
+// NOTE(Ian): Xed adds a read op for the bits read but not written to.
+DEF_SEM(MOVHPS, V128W dst, V128 _nop_read, MV64 src) {
   auto dst_vec = FReadV32(dst);
   auto src_vec = FReadV32(src);
   auto low_entry = FExtractV32(src_vec, 0);

lib/Arch/X86/Semantics/SSE.cpp

@@ -1616,7 +1616,7 @@ DEF_SEM(MOVDDUP, D dst, S1 src) {
 DEF_ISEL(MOVDDUP_XMMdq_MEMq) = MOVDDUP<V128W, MV64>;
 DEF_ISEL(MOVDDUP_XMMdq_XMMq) = MOVDDUP<V128W, V128>;
 IF_AVX(DEF_ISEL(VMOVDDUP_XMMdq_MEMq) = MOVDDUP<VV128W, MV64>;)
-IF_AVX(DEF_ISEL(VMOVDDUP_XMMdq_XMMdq) = MOVDDUP<VV128W, V128>;)
+IF_AVX(DEF_ISEL(VMOVDDUP_XMMdq_XMMq) = MOVDDUP<VV128W, V128>;)
 /*
 2320 VMOVDDUP VMOVDDUP_YMMqq_MEMqq DATAXFER AVX AVX ATTRIBUTES:
 2321 VMOVDDUP VMOVDDUP_YMMqq_YMMqq DATAXFER AVX AVX ATTRIBUTES:

scripts/build.sh

@@ -27,7 +27,7 @@ LLVM_VERSION=llvm-13
 OS_VERSION=
 ARCH_VERSION=
 BUILD_FLAGS=
-CXX_COMMON_VERSION="v0.1.8"
+CXX_COMMON_VERSION="v0.2.7"
 
 # There are pre-build versions of various libraries for specific
 # Ubuntu releases.