[X86] Add IMUL scheduling info on sandybridge, fix it on >=haswell.
Summary:
Only IMUL16rri uses an extra P0156. IMUL32* and IMUL16rr only use
P1.
This was computed using https://github.com/google/EXEgesis/blob/master/exegesis/tools/compute_itineraries.cc
This can easily be validated by running perf on the following code:
```
int main(int argc, char**argv) {
int a = argc;
int b = argc;
int c = argc;
int d = argc;
for (int i = 0; i < LOOP_ITERATIONS; ++i) {
asm volatile(
R"(
.rept 10000
imull $0x2, %%edx, %%eax
imull $0x2, %%ecx, %%ebx
imull $0x2, %%eax, %%edx
imull $0x2, %%ebx, %%ecx
.endr
)"
: "+a"(a), "+b"(b), "+c"(c), "+d"(d)
:
:);
}
return a+b+c+d;
}
```
-> test.cc
perf stat -x, -e cycles --pfm-events=uops_executed_port:port_0:u,uops_executed_port:port_1:u,uops_executed_port:port_2:u,uops_executed_port:port_3:u,uops_executed_port:port_4:u,uops_executed_port:port_5:u,uops_executed_port:port_6:u,uops_executed_port:port_7:u test
Reviewers: craig.topper, RKSimon, gadi.haber
Subscribers: llvm-commits, gchatelet, chandlerc
Differential Revision: https://reviews.llvm.org/D43460
llvm-svn: 326877
diff --git a/llvm/lib/Target/X86/X86SchedBroadwell.td b/llvm/lib/Target/X86/X86SchedBroadwell.td
index 79deb4b..6441fd5 100755
--- a/llvm/lib/Target/X86/X86SchedBroadwell.td
+++ b/llvm/lib/Target/X86/X86SchedBroadwell.td
@@ -1217,7 +1217,7 @@
def: InstRW<[BWWriteResGroup27], (instregex "CVTDQ2PSrr")>;
def: InstRW<[BWWriteResGroup27], (instregex "CVTPS2DQrr")>;
def: InstRW<[BWWriteResGroup27], (instregex "CVTTPS2DQrr")>;
-def: InstRW<[BWWriteResGroup27], (instrs IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
+def: InstRW<[BWWriteResGroup27], (instrs IMUL16rr, IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
def: InstRW<[BWWriteResGroup27], (instrs IMUL8r)>;
def: InstRW<[BWWriteResGroup27], (instregex "LZCNT(16|32|64)rr")>;
def: InstRW<[BWWriteResGroup27], (instregex "MAX(C?)PDrr")>;
@@ -1298,7 +1298,7 @@
let NumMicroOps = 2;
let ResourceCycles = [1,1];
}
-def: InstRW<[BWWriteResGroup27_16], (instrs IMUL16rr, IMUL16rri, IMUL16rri8)>;
+def: InstRW<[BWWriteResGroup27_16], (instrs IMUL16rri, IMUL16rri8)>;
def BWWriteResGroup28 : SchedWriteRes<[BWPort5]> {
let Latency = 3;
diff --git a/llvm/lib/Target/X86/X86SchedHaswell.td b/llvm/lib/Target/X86/X86SchedHaswell.td
index ce30ee8..fbaf745 100644
--- a/llvm/lib/Target/X86/X86SchedHaswell.td
+++ b/llvm/lib/Target/X86/X86SchedHaswell.td
@@ -2498,7 +2498,7 @@
def: InstRW<[HWWriteResGroup50], (instregex "CVTDQ2PSrr")>;
def: InstRW<[HWWriteResGroup50], (instregex "CVTPS2DQrr")>;
def: InstRW<[HWWriteResGroup50], (instregex "CVTTPS2DQrr")>;
-def: InstRW<[HWWriteResGroup50], (instrs IMUL64rr, IMUL64rri32, IMUL64rri8)>;
+def: InstRW<[HWWriteResGroup50], (instrs IMUL16rr, IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
def: InstRW<[HWWriteResGroup50], (instrs IMUL8r)>;
def: InstRW<[HWWriteResGroup50], (instregex "LZCNT(16|32|64)rr")>;
def: InstRW<[HWWriteResGroup50], (instregex "MAX(C?)PDrr")>;
@@ -2574,17 +2574,12 @@
def: InstRW<[HWWriteResGroup50], (instregex "VUCOMISDrr")>;
def: InstRW<[HWWriteResGroup50], (instregex "VUCOMISSrr")>;
-def HWWriteResGroup50_16 : SchedWriteRes<[HWPort1, HWPort0156]> {
+def HWWriteResGroup50_16i : SchedWriteRes<[HWPort1, HWPort0156]> {
let Latency = 3;
- let NumMicroOps = 4;
+ let NumMicroOps = 2;
+ let ResourceCycles = [1,1];
}
-def: InstRW<[HWWriteResGroup50_16], (instrs IMUL16rr, IMUL16rri, IMUL16rri8)>;
-
-def HWWriteResGroup50_32 : SchedWriteRes<[HWPort1, HWPort0156]> {
- let Latency = 3;
- let NumMicroOps = 3;
-}
-def: InstRW<[HWWriteResGroup50_32], (instrs IMUL32rr, IMUL32rri, IMUL32rri8)>;
+def: InstRW<[HWWriteResGroup50_16i], (instrs IMUL16rri, IMUL16rri8)>;
def HWWriteResGroup51 : SchedWriteRes<[HWPort5]> {
let Latency = 3;
diff --git a/llvm/lib/Target/X86/X86SchedSandyBridge.td b/llvm/lib/Target/X86/X86SchedSandyBridge.td
index bd1a271..bec8f62 100644
--- a/llvm/lib/Target/X86/X86SchedSandyBridge.td
+++ b/llvm/lib/Target/X86/X86SchedSandyBridge.td
@@ -925,6 +925,7 @@
def: InstRW<[SBWriteResGroup21], (instregex "CVTDQ2PSrr")>;
def: InstRW<[SBWriteResGroup21], (instregex "CVTPS2DQrr")>;
def: InstRW<[SBWriteResGroup21], (instregex "CVTTPS2DQrr")>;
+def: InstRW<[SBWriteResGroup21], (instrs IMUL16rr, IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
def: InstRW<[SBWriteResGroup21], (instregex "MAX(C?)PDrr")>;
def: InstRW<[SBWriteResGroup21], (instregex "MAX(C?)PSrr")>;
def: InstRW<[SBWriteResGroup21], (instregex "MAX(C?)SDrr")>;
@@ -1000,6 +1001,13 @@
def: InstRW<[SBWriteResGroup21], (instregex "VSUBSDrr")>;
def: InstRW<[SBWriteResGroup21], (instregex "VSUBSSrr")>;
+def SBWriteResGroup21_16i : SchedWriteRes<[SBPort1, SBPort015]> {
+ let Latency = 3;
+ let NumMicroOps = 2;
+ let ResourceCycles = [1,1];
+}
+def: InstRW<[SBWriteResGroup21_16i], (instrs IMUL16rri, IMUL16rri8)>;
+
def SBWriteResGroup22 : SchedWriteRes<[SBPort0,SBPort5]> {
let Latency = 3;
let NumMicroOps = 2;
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeClient.td b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
index 1ff268a..abad9ce 100644
--- a/llvm/lib/Target/X86/X86SchedSkylakeClient.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
@@ -1214,7 +1214,7 @@
}
def: InstRW<[SKLWriteResGroup29], (instregex "BSF(16|32|64)rr")>;
def: InstRW<[SKLWriteResGroup29], (instregex "BSR(16|32|64)rr")>;
-def: InstRW<[SKLWriteResGroup29], (instrs IMUL64rr, IMUL64rri32, IMUL64rri8)>;
+def: InstRW<[SKLWriteResGroup29], (instrs IMUL16rr, IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
def: InstRW<[SKLWriteResGroup29], (instrs IMUL8r)>;
def: InstRW<[SKLWriteResGroup29], (instregex "LZCNT(16|32|64)rr")>;
def: InstRW<[SKLWriteResGroup29], (instrs MUL8r)>;
@@ -1225,18 +1225,12 @@
def: InstRW<[SKLWriteResGroup29], (instregex "SHRD(16|32|64)rri8")>;
def: InstRW<[SKLWriteResGroup29], (instregex "TZCNT(16|32|64)rr")>;
-def SKLWriteResGroup29_16 : SchedWriteRes<[SKLPort1, SKLPort0156]> {
+def SKLWriteResGroup29_16i : SchedWriteRes<[SKLPort1, SKLPort0156]> {
let Latency = 3;
let NumMicroOps = 2;
let ResourceCycles = [1,1];
}
-def: InstRW<[SKLWriteResGroup29_16], (instrs IMUL16rr, IMUL16rri, IMUL16rri8)>;
-
-def SKLWriteResGroup29_32 : SchedWriteRes<[SKLPort1]> {
- let Latency = 3;
- let NumMicroOps = 1;
-}
-def: InstRW<[SKLWriteResGroup29_32], (instrs IMUL32rr, IMUL32rri, IMUL32rri8)>;
+def: InstRW<[SKLWriteResGroup29_16i], (instrs IMUL16rri, IMUL16rri8)>;
def SKLWriteResGroup30 : SchedWriteRes<[SKLPort5]> {
let Latency = 3;
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeServer.td b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
index 44ba5b4..bc5ecc6 100755
--- a/llvm/lib/Target/X86/X86SchedSkylakeServer.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
@@ -1758,7 +1758,7 @@
}
def: InstRW<[SKXWriteResGroup31], (instregex "BSF(16|32|64)rr")>;
def: InstRW<[SKXWriteResGroup31], (instregex "BSR(16|32|64)rr")>;
-def: InstRW<[SKXWriteResGroup31], (instrs IMUL64rr, IMUL64rri32, IMUL64rri8)>;
+def: InstRW<[SKXWriteResGroup31], (instrs IMUL16rr, IMUL32rr, IMUL32rri, IMUL32rri8, IMUL64rr, IMUL64rri32, IMUL64rri8)>;
def: InstRW<[SKXWriteResGroup31], (instrs IMUL8r)>;
def: InstRW<[SKXWriteResGroup31], (instregex "LZCNT(16|32|64)rr")>;
def: InstRW<[SKXWriteResGroup31], (instrs MUL8r)>;
@@ -1769,18 +1769,13 @@
def: InstRW<[SKXWriteResGroup31], (instregex "SHRD(16|32|64)rri8")>;
def: InstRW<[SKXWriteResGroup31], (instregex "TZCNT(16|32|64)rr")>;
-def SKXWriteResGroup31_16 : SchedWriteRes<[SKXPort1, SKXPort0156]> {
+def SKXWriteResGroup31_16i : SchedWriteRes<[SKXPort1, SKXPort0156]> {
let Latency = 3;
let NumMicroOps = 2;
let ResourceCycles = [1,1];
}
-def: InstRW<[SKXWriteResGroup31_16], (instrs IMUL16rr, IMUL16rri, IMUL16rri8)>;
+def: InstRW<[SKXWriteResGroup31_16i], (instrs IMUL16rri, IMUL16rri8)>;
-def SKXWriteResGroup31_32 : SchedWriteRes<[SKXPort1]> {
- let Latency = 3;
- let NumMicroOps = 1;
-}
-def: InstRW<[SKXWriteResGroup31_32], (instrs IMUL32rr, IMUL32rri, IMUL32rri8)>;
def SKXWriteResGroup32 : SchedWriteRes<[SKXPort5]> {
let Latency = 3;