[X86] Enable -mprefer-vector-width=256 by default for Skylake-avx512 and later Intel CPUs.

AVX512 instructions can cause a frequency drop on these CPUs. This
can negate the performance gains from using wider vectors. Enabling
prefer-vector-width=256 will prevent generation of zmm registers
unless explicit 512 bit operations are used in the original source
code.

I believe gcc and icc both do something similar to this by default.

Differential Revision: https://reviews.llvm.org/D67259

llvm-svn: 371694

3 files changed, 14 insertions, 0 deletions

diff --git a/llvm/docs/ReleaseNotes.rst b/llvm/docs/ReleaseNotes.rst
index eb340235f5f..1d501b26063 100644
--- a/llvm/docs/ReleaseNotes.rst
+++ b/llvm/docs/ReleaseNotes.rst
@@ -96,6 +96,10 @@ Changes to the X86 Target
   be passed in ZMM registers for calls and returns. Previously they were passed
   in two YMM registers. Old behavior can be enabled by passing
   -x86-enable-old-knl-abi
+* -mprefer-vector-width=256 is now the default behavior skylake-avx512 and later
+  Intel CPUs. This tries to limit the use of 512-bit registers which can cause a
+  decrease in CPU frequency on these CPUs. This can be re-enabled by passing
+  -mprefer-vector-width=512 to clang or passing -mattr=-prefer-256-bit to llc.
 
 Changes to the AMDGPU Target
 -----------------------------
diff --git a/llvm/lib/Target/X86/X86.td b/llvm/lib/Target/X86/X86.td
index 96d24228253..53a05f6a9c4 100644
--- a/llvm/lib/Target/X86/X86.td
+++ b/llvm/lib/Target/X86/X86.td
@@ -601,6 +601,7 @@ def ProcessorFeatures {
 
   // Skylake-AVX512
   list<SubtargetFeature> SKXAdditionalFeatures = [FeatureAVX512,
+                                                  FeaturePrefer256Bit,
                                                   FeatureCDI,
                                                   FeatureDQI,
                                                   FeatureBWI,
@@ -634,6 +635,7 @@ def ProcessorFeatures {
 
   // Cannonlake
   list<SubtargetFeature> CNLAdditionalFeatures = [FeatureAVX512,
+                                                  FeaturePrefer256Bit,
                                                   FeatureCDI,
                                                   FeatureDQI,
                                                   FeatureBWI,
diff --git a/llvm/test/CodeGen/X86/min-legal-vector-width.ll b/llvm/test/CodeGen/X86/min-legal-vector-width.ll
index baee49ca295..b69525deb41 100644
--- a/llvm/test/CodeGen/X86/min-legal-vector-width.ll
+++ b/llvm/test/CodeGen/X86/min-legal-vector-width.ll
@@ -1,6 +1,14 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=fast-variable-shuffle,avx512vl,avx512bw,avx512dq,prefer-256-bit | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=fast-variable-shuffle,avx512vl,avx512bw,avx512dq,prefer-256-bit,avx512vbmi | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
+; Make sure CPUs default to prefer-256-bit. avx512vnni isn't interesting as it just adds an isel peephole for vpmaddwd+vpaddd
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake-avx512 | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=cascadelake | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=cooperlake | FileCheck %s --check-prefixes=CHECK,CHECK-AVX512
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=cannonlake | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=icelake-client | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=icelake-server | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-avx512vnni -mcpu=tigerlake | FileCheck %s --check-prefixes=CHECK,CHECK-VBMI
 
 ; This file primarily contains tests for specific places in X86ISelLowering.cpp that needed be made aware of the legalizer not allowing 512-bit vectors due to prefer-256-bit even though AVX512 is enabled.