[X86][AVX] Fixed v16i16/v32i8 ADD/SUB costs on AVX1 subtargets

Add explicit v16i16/v32i8 ADD/SUB costs, matching the costs of v4i64/v8i32 - they were missing for some reason.

This has side effects on the LV max bandwidth tests (AVX1 now prefers 128-bit vectors vs AVX2 which still prefers 256-bit)

llvm-svn: 286832

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

@@ -526,6 +526,10 @@ int X86TTIImpl::getArithmeticInstrCost(
     // Two ops + 1 extract + 1 insert = 4.
     { ISD::MUL,     MVT::v16i16,   4 },
     { ISD::MUL,     MVT::v8i32,    4 },
+    { ISD::SUB,     MVT::v32i8,    4 },
+    { ISD::ADD,     MVT::v32i8,    4 },
+    { ISD::SUB,     MVT::v16i16,   4 },
+    { ISD::ADD,     MVT::v16i16,   4 },
     { ISD::SUB,     MVT::v8i32,    4 },
     { ISD::ADD,     MVT::v8i32,    4 },
     { ISD::SUB,     MVT::v4i64,    4 },

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

@@ -57,13 +57,13 @@ define i32 @add(i32 %arg) {
   %G = add <8 x i16> undef, undef
   ; SSSE3: cost of 2 {{.*}} %H = add
   ; SSE42: cost of 2 {{.*}} %H = add
-  ; AVX: cost of 2 {{.*}} %H = add
+  ; AVX: cost of 4 {{.*}} %H = add
   ; AVX2: cost of 1 {{.*}} %H = add
   ; AVX512: cost of 1 {{.*}} %H = add
   %H = add <16 x i16> undef, undef
   ; SSSE3: cost of 4 {{.*}} %I = add
   ; SSE42: cost of 4 {{.*}} %I = add
-  ; AVX: cost of 4 {{.*}} %I = add
+  ; AVX: cost of 8 {{.*}} %I = add
   ; AVX2: cost of 2 {{.*}} %I = add
   ; AVX512F: cost of 2 {{.*}} %I = add
   ; AVX512BW: cost of 1 {{.*}} %I = add
@@ -77,13 +77,13 @@ define i32 @add(i32 %arg) {
   %J = add <16 x i8> undef, undef
   ; SSSE3: cost of 2 {{.*}} %K = add
   ; SSE42: cost of 2 {{.*}} %K = add
-  ; AVX: cost of 2 {{.*}} %K = add
+  ; AVX: cost of 4 {{.*}} %K = add
   ; AVX2: cost of 1 {{.*}} %K = add
   ; AVX512: cost of 1 {{.*}} %K = add
   %K = add <32 x i8> undef, undef
   ; SSSE3: cost of 4 {{.*}} %L = add
   ; SSE42: cost of 4 {{.*}} %L = add
-  ; AVX: cost of 4 {{.*}} %L = add
+  ; AVX: cost of 8 {{.*}} %L = add
   ; AVX2: cost of 2 {{.*}} %L = add
   ; AVX512F: cost of 2 {{.*}} %L = add
   ; AVX512BW: cost of 1 {{.*}} %L = add
@@ -140,13 +140,13 @@ define i32 @sub(i32 %arg) {
   %G = sub <8 x i16> undef, undef
   ; SSSE3: cost of 2 {{.*}} %H = sub
   ; SSE42: cost of 2 {{.*}} %H = sub
-  ; AVX: cost of 2 {{.*}} %H = sub
+  ; AVX: cost of 4 {{.*}} %H = sub
   ; AVX2: cost of 1 {{.*}} %H = sub
   ; AVX512: cost of 1 {{.*}} %H = sub
   %H = sub <16 x i16> undef, undef
   ; SSSE3: cost of 4 {{.*}} %I = sub
   ; SSE42: cost of 4 {{.*}} %I = sub
-  ; AVX: cost of 4 {{.*}} %I = sub
+  ; AVX: cost of 8 {{.*}} %I = sub
   ; AVX2: cost of 2 {{.*}} %I = sub
   ; AVX512F: cost of 2 {{.*}} %I = sub
   ; AVX512BW: cost of 1 {{.*}} %I = sub
@@ -160,13 +160,13 @@ define i32 @sub(i32 %arg) {
   %J = sub <16 x i8> undef, undef
   ; SSSE3: cost of 2 {{.*}} %K = sub
   ; SSE42: cost of 2 {{.*}} %K = sub
-  ; AVX: cost of 2 {{.*}} %K = sub
+  ; AVX: cost of 4 {{.*}} %K = sub
   ; AVX2: cost of 1 {{.*}} %K = sub
   ; AVX512: cost of 1 {{.*}} %K = sub
   %K = sub <32 x i8> undef, undef
   ; SSSE3: cost of 4 {{.*}} %L = sub
   ; SSE42: cost of 4 {{.*}} %L = sub
-  ; AVX: cost of 4 {{.*}} %L = sub
+  ; AVX: cost of 8 {{.*}} %L = sub
   ; AVX2: cost of 2 {{.*}} %L = sub
   ; AVX512F: cost of 2 {{.*}} %L = sub
   ; AVX512BW: cost of 1 {{.*}} %L = sub

llvm/test/Transforms/LoopVectorize/X86/vector_max_bandwidth.ll

-; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -mcpu=corei7-avx -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck %s
+; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -mcpu=corei7-avx -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-AVX1
+; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -mcpu=core-avx2 -debug-only=loop-vectorize -S < %s 2>&1 | FileCheck %s --check-prefix=CHECK-AVX2
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -16,7 +17,8 @@ target triple = "x86_64-unknown-linux-gnu"
 ; -vectorizer-maximize-bandwidth is indicated.
 ;
 ; CHECK-label: foo
-; CHECK: LV: Selecting VF: 32.
+; CHECK-AVX1: LV: Selecting VF: 16.
+; CHECK-AVX2: LV: Selecting VF: 32.
 define void @foo() {
 entry:
   br label %for.body