[x86] make SLM extract vector element more expensive than default

I'm not sure what the effect of this change will be on all of the affected
tests or a larger benchmark, but it fixes the horizontal add/sub problems
noted here:
https://reviews.llvm.org/D59710?vs=227972&id=228095&whitespace=ignore-most#toc

The costs are based on reciprocal throughput numbers in Agner's tables for
PEXTR*; these appear to be very slow ops on Silvermont.

This is a small step towards the larger motivation discussed in PR43605:
https://bugs.llvm.org/show_bug.cgi?id=43605

Also, it seems likely that insert/extract is the source of perf regressions on
other CPUs (up to 30%) that were cited as part of the reason to revert D59710,
so maybe we'll extend the table-based approach to other subtargets.

Differential Revision: https://reviews.llvm.org/D70607

Author: rotateright

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

@@ -2377,6 +2377,13 @@ int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
 }
 
 int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
+  static const CostTblEntry SLMCostTbl[] = {
+     { ISD::EXTRACT_VECTOR_ELT,       MVT::i8,      4 },
+     { ISD::EXTRACT_VECTOR_ELT,       MVT::i16,     4 },
+     { ISD::EXTRACT_VECTOR_ELT,       MVT::i32,     4 },
+     { ISD::EXTRACT_VECTOR_ELT,       MVT::i64,     7 }
+   };
+
   assert(Val->isVectorTy() && "This must be a vector type");
 
   Type *ScalarType = Val->getScalarType();
@@ -2396,6 +2403,13 @@ int X86TTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
     // Floating point scalars are already located in index #0.
     if (ScalarType->isFloatingPointTy() && Index == 0)
       return 0;
+
+    int ISD = TLI->InstructionOpcodeToISD(Opcode);
+    assert(ISD && "Unexpected vector opcode");
+    MVT MScalarTy = LT.second.getScalarType();
+    if (ST->isSLM())
+      if (auto *Entry = CostTableLookup(SLMCostTbl, ISD, MScalarTy))
+        return LT.first * Entry->Cost;
   }
 
   // Add to the base cost if we know that the extracted element of a vector is

llvm/test/Analysis/CostModel/X86/fptosi.ll

@@ -6,7 +6,7 @@
 ; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,AVX512,AVX512F
 ; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mattr=+avx512f,+avx512dq | FileCheck %s --check-prefixes=CHECK,AVX512,AVX512DQ
 ;
-; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mcpu=slm | FileCheck %s --check-prefixes=CHECK,SSE,SSE42
+; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mcpu=slm | FileCheck %s --check-prefixes=SLM
 ; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mcpu=goldmont | FileCheck %s --check-prefixes=CHECK,SSE,SSE42
 ; RUN: opt < %s -mtriple=x86_64-apple-darwin -cost-model -analyze -mcpu=btver2 | FileCheck %s --check-prefixes=BTVER2
 
@@ -39,6 +39,13 @@ define i32 @fptosi_double_i64(i32 %arg) {
 ; AVX512DQ-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I64 = fptosi <8 x double> undef to <8 x i64>
 ; AVX512DQ-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_double_i64'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = fptosi double undef to i64
+; SLM-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: %V2I64 = fptosi <2 x double> undef to <2 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 37 for instruction: %V4I64 = fptosi <4 x double> undef to <4 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 75 for instruction: %V8I64 = fptosi <8 x double> undef to <8 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_double_i64'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = fptosi double undef to i64
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = fptosi <2 x double> undef to <2 x i64>
@@ -75,6 +82,13 @@ define i32 @fptosi_double_i32(i32 %arg) {
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = fptosi <8 x double> undef to <8 x i32>
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_double_i32'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = fptosi double undef to i32
+; SLM-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I32 = fptosi <2 x double> undef to <2 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %V4I32 = fptosi <4 x double> undef to <4 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %V8I32 = fptosi <8 x double> undef to <8 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_double_i32'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = fptosi double undef to i32
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V2I32 = fptosi <2 x double> undef to <2 x i32>
@@ -111,6 +125,13 @@ define i32 @fptosi_double_i16(i32 %arg) {
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I16 = fptosi <8 x double> undef to <8 x i16>
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_double_i16'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = fptosi double undef to i16
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I16 = fptosi <2 x double> undef to <2 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V4I16 = fptosi <4 x double> undef to <4 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %V8I16 = fptosi <8 x double> undef to <8 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_double_i16'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = fptosi double undef to i16
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V2I16 = fptosi <2 x double> undef to <2 x i16>
@@ -147,6 +168,13 @@ define i32 @fptosi_double_i8(i32 %arg) {
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I8 = fptosi <8 x double> undef to <8 x i8>
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_double_i8'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = fptosi double undef to i8
+; SLM-NEXT:  Cost Model: Found an estimated cost of 12 for instruction: %V2I8 = fptosi <2 x double> undef to <2 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %V4I8 = fptosi <4 x double> undef to <4 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 51 for instruction: %V8I8 = fptosi <8 x double> undef to <8 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_double_i8'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = fptosi double undef to i8
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I8 = fptosi <2 x double> undef to <2 x i8>
@@ -194,6 +222,14 @@ define i32 @fptosi_float_i64(i32 %arg) {
 ; AVX512DQ-NEXT:  Cost Model: Found an estimated cost of 3 for instruction: %V16I64 = fptosi <16 x float> undef to <16 x i64>
 ; AVX512DQ-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_float_i64'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = fptosi float undef to i64
+; SLM-NEXT:  Cost Model: Found an estimated cost of 18 for instruction: %V2I64 = fptosi <2 x float> undef to <2 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 37 for instruction: %V4I64 = fptosi <4 x float> undef to <4 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 75 for instruction: %V8I64 = fptosi <8 x float> undef to <8 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 151 for instruction: %V16I64 = fptosi <16 x float> undef to <16 x i64>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_float_i64'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I64 = fptosi float undef to i64
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 6 for instruction: %V2I64 = fptosi <2 x float> undef to <2 x i64>
@@ -218,6 +254,13 @@ define i32 @fptosi_float_i32(i32 %arg) {
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = fptosi <16 x float> undef to <16 x i32>
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_float_i32'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = fptosi float undef to i32
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = fptosi <4 x float> undef to <4 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V8I32 = fptosi <8 x float> undef to <8 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I32 = fptosi <16 x float> undef to <16 x i32>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_float_i32'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I32 = fptosi float undef to i32
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I32 = fptosi <4 x float> undef to <4 x i32>
@@ -254,6 +297,13 @@ define i32 @fptosi_float_i16(i32 %arg) {
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I16 = fptosi <16 x float> undef to <16 x i16>
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_float_i16'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = fptosi float undef to i16
+; SLM-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I16 = fptosi <4 x float> undef to <4 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 5 for instruction: %V8I16 = fptosi <8 x float> undef to <8 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %V16I16 = fptosi <16 x float> undef to <16 x i16>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_float_i16'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I16 = fptosi float undef to i16
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %V4I16 = fptosi <4 x float> undef to <4 x i16>
@@ -290,6 +340,13 @@ define i32 @fptosi_float_i8(i32 %arg) {
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V16I8 = fptosi <16 x float> undef to <16 x i8>
 ; AVX512-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
+; SLM-LABEL: 'fptosi_float_i8'
+; SLM-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = fptosi float undef to i8
+; SLM-NEXT:  Cost Model: Found an estimated cost of 24 for instruction: %V4I8 = fptosi <4 x float> undef to <4 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 49 for instruction: %V8I8 = fptosi <8 x float> undef to <8 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 99 for instruction: %V16I8 = fptosi <16 x float> undef to <16 x i8>
+; SLM-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
+;
 ; BTVER2-LABEL: 'fptosi_float_i8'
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %I8 = fptosi float undef to i8
 ; BTVER2-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %V4I8 = fptosi <4 x float> undef to <4 x i8>