stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8239009: C2: Don't use PSHUF to load scalars from memory on x86

Browse Source 
Reviewed-by: kvn, dlong
This commit is contained in:
Vladimir Ivanov 2020-03-10 20:51:09 +03:00
parent 1dcd3d2c50
commit 02916dbb0b
1 changed files with 9 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/hotspot/cpu/x86/x86.ad
View File

@ -3227,21 +3227,12 @@ instruct ReplS_reg(vec dst, rRegI src) %{
%}
instruct ReplS_mem(vec dst, memory mem) %{
predicate(VM_Version::supports_avx()); // use VEX-encoded pshuflw to relax 16-byte alignment restriction on the source
predicate(VM_Version::supports_avx2());
match(Set dst (ReplicateS (LoadS mem)));
format %{ "replicateS $dst,$mem" %}
ins_encode %{
uint vlen = vector_length(this);
if (VM_Version::supports_avx2()) {
int vlen_enc = vector_length_encoding(this);
__ vpbroadcastw($dst$$XMMRegister, $mem$$Address, vlen_enc);
} else {
__ pshuflw($dst$$XMMRegister, $mem$$Address, 0x00);
if (vlen >= 8) {
assert(vlen == 8, "sanity");
__ punpcklqdq($dst$$XMMRegister, $dst$$XMMRegister);
}
}
int vlen_enc = vector_length_encoding(this);
__ vpbroadcastw($dst$$XMMRegister, $mem$$Address, vlen_enc);
%}
ins_pipe( pipe_slow );
%}
@ -3308,13 +3299,13 @@ instruct ReplI_reg(vec dst, rRegI src) %{
%}
instruct ReplI_mem(vec dst, memory mem) %{
predicate(VM_Version::supports_avx()); // use VEX-encoded pshufd to relax 16-byte alignment restriction on the source
match(Set dst (ReplicateI (LoadI mem)));
format %{ "replicateI $dst,$mem" %}
ins_encode %{
uint vlen = vector_length(this);
if (vlen <= 4) {
__ pshufd($dst$$XMMRegister, $mem$$Address, 0x00);
__ movdl($dst$$XMMRegister, $mem$$Address);
__ pshufd($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
} else {
assert(VM_Version::supports_avx2(), "sanity");
int vector_len = vector_length_encoding(this);
@ -3515,13 +3506,13 @@ instruct ReplF_reg(vec dst, vlRegF src) %{
%}
instruct ReplF_mem(vec dst, memory mem) %{
predicate(VM_Version::supports_avx()); // use VEX-encoded pshufd to relax 16-byte alignment restriction on the source
match(Set dst (ReplicateF (LoadF mem)));
format %{ "replicateF $dst,$mem" %}
ins_encode %{
uint vlen = vector_length(this);
if (vlen <= 4) {
__ pshufd($dst$$XMMRegister, $mem$$Address, 0x00);
__ movdl($dst$$XMMRegister, $mem$$Address);
__ pshufd($dst$$XMMRegister, $dst$$XMMRegister, 0x00);
} else {
assert(VM_Version::supports_avx(), "sanity");
int vector_len = vector_length_encoding(this);
@ -3569,13 +3560,13 @@ instruct ReplD_reg(vec dst, vlRegD src) %{
%}
instruct ReplD_mem(vec dst, memory mem) %{
predicate(VM_Version::supports_avx()); // use VEX-encoded pshufd to relax 16-byte alignment restriction on the source
match(Set dst (ReplicateD (LoadD mem)));
format %{ "replicateD $dst,$mem" %}
ins_encode %{
uint vlen = vector_length(this);
if (vlen == 2) {
__ pshufd($dst$$XMMRegister, $mem$$Address, 0x44);
__ movq($dst$$XMMRegister, $mem$$Address);
__ pshufd($dst$$XMMRegister, $dst$$XMMRegister, 0x44);
} else {
assert(VM_Version::supports_avx(), "sanity");
int vector_len = vector_length_encoding(this);