[X86][AVX1] Improve 256-bit vector costs for integer unary intrinsics.

Account for subvector extraction/insertion, helps prevent the vectorizers from selecting 256-bit vectors that will have to be split anyhow on AVX1 targets.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@302378 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86TargetTransformInfo.cpp b/lib/Target/X86/X86TargetTransformInfo.cpp
index b742fb472372cd9f84c6159c6a98d7610caf94a0..f3b619a2956a02a5814532d10ab62ee19f70bc66 100644 (file)
--- a/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -1426,25 +1426,25 @@ int X86TTIImpl::getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy,
     { ISD::FSQRT,      MVT::v4f64,  28 }, // Haswell from http://www.agner.org/
   };
   static const CostTblEntry AVX1CostTbl[] = {
-    { ISD::BITREVERSE, MVT::v4i64,  10 },
-    { ISD::BITREVERSE, MVT::v8i32,  10 },
-    { ISD::BITREVERSE, MVT::v16i16, 10 },
-    { ISD::BITREVERSE, MVT::v32i8,  10 },
+    { ISD::BITREVERSE, MVT::v4i64,  12 }, // 2 x 128-bit Op + extract/insert
+    { ISD::BITREVERSE, MVT::v8i32,  12 }, // 2 x 128-bit Op + extract/insert
+    { ISD::BITREVERSE, MVT::v16i16, 12 }, // 2 x 128-bit Op + extract/insert
+    { ISD::BITREVERSE, MVT::v32i8,  12 }, // 2 x 128-bit Op + extract/insert
     { ISD::BSWAP,      MVT::v4i64,   4 },
     { ISD::BSWAP,      MVT::v8i32,   4 },
     { ISD::BSWAP,      MVT::v16i16,  4 },
-    { ISD::CTLZ,       MVT::v4i64,  46 },
-    { ISD::CTLZ,       MVT::v8i32,  36 },
-    { ISD::CTLZ,       MVT::v16i16, 28 },
-    { ISD::CTLZ,       MVT::v32i8,  18 },
-    { ISD::CTPOP,      MVT::v4i64,  14 },
-    { ISD::CTPOP,      MVT::v8i32,  22 },
-    { ISD::CTPOP,      MVT::v16i16, 18 },
-    { ISD::CTPOP,      MVT::v32i8,  12 },
-    { ISD::CTTZ,       MVT::v4i64,  20 },
-    { ISD::CTTZ,       MVT::v8i32,  28 },
-    { ISD::CTTZ,       MVT::v16i16, 24 },
-    { ISD::CTTZ,       MVT::v32i8,  18 },
+    { ISD::CTLZ,       MVT::v4i64,  48 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTLZ,       MVT::v8i32,  38 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTLZ,       MVT::v16i16, 30 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTLZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTPOP,      MVT::v4i64,  16 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTPOP,      MVT::v8i32,  24 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTPOP,      MVT::v16i16, 20 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTPOP,      MVT::v32i8,  14 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTTZ,       MVT::v4i64,  22 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTTZ,       MVT::v8i32,  30 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTTZ,       MVT::v16i16, 26 }, // 2 x 128-bit Op + extract/insert
+    { ISD::CTTZ,       MVT::v32i8,  20 }, // 2 x 128-bit Op + extract/insert
     { ISD::FSQRT,      MVT::f32,    14 }, // SNB from http://www.agner.org/
     { ISD::FSQRT,      MVT::v4f32,  14 }, // SNB from http://www.agner.org/
     { ISD::FSQRT,      MVT::v8f32,  28 }, // SNB from http://www.agner.org/

diff --git a/test/Analysis/CostModel/X86/bitreverse.ll b/test/Analysis/CostModel/X86/bitreverse.ll
index 2eb63babdc343263460fb48407ef05a54a74669f..8d5e1421eb829c5b10c95621067232893afd0663 100644 (file)
--- a/test/Analysis/CostModel/X86/bitreverse.ll
+++ b/test/Analysis/CostModel/X86/bitreverse.ll
@@ -79,7 +79,7 @@ define <4 x i64> @var_bitreverse_v4i64(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_bitreverse_v4i64':
 ; SSE2: Found an estimated cost of 58 for instruction:   %bitreverse
 ; SSE42: Found an estimated cost of 10 for instruction:   %bitreverse
-; AVX: Found an estimated cost of 10 for instruction:   %bitreverse
+; AVX: Found an estimated cost of 12 for instruction:   %bitreverse
 ; AVX2: Found an estimated cost of 5 for instruction:   %bitreverse
 ; XOP: Found an estimated cost of 4 for instruction:   %bitreverse
   %bitreverse = call <4 x i64> @llvm.bitreverse.v4i64(<4 x i64> %a)
@@ -101,7 +101,7 @@ define <8 x i32> @var_bitreverse_v8i32(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_bitreverse_v8i32':
 ; SSE2: Found an estimated cost of 54 for instruction:   %bitreverse
 ; SSE42: Found an estimated cost of 10 for instruction:   %bitreverse
-; AVX: Found an estimated cost of 10 for instruction:   %bitreverse
+; AVX: Found an estimated cost of 12 for instruction:   %bitreverse
 ; AVX2: Found an estimated cost of 5 for instruction:   %bitreverse
 ; XOP: Found an estimated cost of 4 for instruction:   %bitreverse
   %bitreverse = call <8 x i32> @llvm.bitreverse.v8i32(<8 x i32> %a)
@@ -123,7 +123,7 @@ define <16 x i16> @var_bitreverse_v16i16(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_bitreverse_v16i16':
 ; SSE2: Found an estimated cost of 54 for instruction:   %bitreverse
 ; SSE42: Found an estimated cost of 10 for instruction:   %bitreverse
-; AVX: Found an estimated cost of 10 for instruction:   %bitreverse
+; AVX: Found an estimated cost of 12 for instruction:   %bitreverse
 ; AVX2: Found an estimated cost of 5 for instruction:   %bitreverse
 ; XOP: Found an estimated cost of 4 for instruction:   %bitreverse
   %bitreverse = call <16 x i16> @llvm.bitreverse.v16i16(<16 x i16> %a)
@@ -145,7 +145,7 @@ define <32 x i8> @var_bitreverse_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_bitreverse_v32i8':
 ; SSE2: Found an estimated cost of 40 for instruction:   %bitreverse
 ; SSE42: Found an estimated cost of 10 for instruction:   %bitreverse
-; AVX: Found an estimated cost of 10 for instruction:   %bitreverse
+; AVX: Found an estimated cost of 12 for instruction:   %bitreverse
 ; AVX2: Found an estimated cost of 5 for instruction:   %bitreverse
 ; XOP: Found an estimated cost of 4 for instruction:   %bitreverse
   %bitreverse = call <32 x i8> @llvm.bitreverse.v32i8(<32 x i8> %a)

diff --git a/test/Analysis/CostModel/X86/ctbits-cost.ll b/test/Analysis/CostModel/X86/ctbits-cost.ll
index 8c7fa9d73151d90de25be690fcf213c9a316b3f3..aaf092c7b1d7571d0b8104e22dfb41286864400b 100644 (file)
--- a/test/Analysis/CostModel/X86/ctbits-cost.ll
+++ b/test/Analysis/CostModel/X86/ctbits-cost.ll
@@ -69,7 +69,7 @@ define <4 x i64> @var_ctpop_v4i64(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctpop_v4i64':
 ; SSE2: Found an estimated cost of 24 for instruction:   %ctpop
 ; SSE42: Found an estimated cost of 14 for instruction:   %ctpop
-; AVX1: Found an estimated cost of 14 for instruction:   %ctpop
+; AVX1: Found an estimated cost of 16 for instruction:   %ctpop
 ; AVX2: Found an estimated cost of 7 for instruction:   %ctpop
   %ctpop = call <4 x i64> @llvm.ctpop.v4i64(<4 x i64> %a)
   ret <4 x i64> %ctpop
@@ -88,7 +88,7 @@ define <8 x i32> @var_ctpop_v8i32(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctpop_v8i32':
 ; SSE2: Found an estimated cost of 30 for instruction:   %ctpop
 ; SSE42: Found an estimated cost of 22 for instruction:   %ctpop
-; AVX1: Found an estimated cost of 22 for instruction:   %ctpop
+; AVX1: Found an estimated cost of 24 for instruction:   %ctpop
 ; AVX2: Found an estimated cost of 11 for instruction:   %ctpop
   %ctpop = call <8 x i32> @llvm.ctpop.v8i32(<8 x i32> %a)
   ret <8 x i32> %ctpop
@@ -107,7 +107,7 @@ define <16 x i16> @var_ctpop_v16i16(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctpop_v16i16':
 ; SSE2: Found an estimated cost of 26 for instruction:   %ctpop
 ; SSE42: Found an estimated cost of 18 for instruction:   %ctpop
-; AVX1: Found an estimated cost of 18 for instruction:   %ctpop
+; AVX1: Found an estimated cost of 20 for instruction:   %ctpop
 ; AVX2: Found an estimated cost of 9 for instruction:   %ctpop
   %ctpop = call <16 x i16> @llvm.ctpop.v16i16(<16 x i16> %a)
   ret <16 x i16> %ctpop
@@ -126,7 +126,7 @@ define <32 x i8> @var_ctpop_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctpop_v32i8':
 ; SSE2: Found an estimated cost of 20 for instruction:   %ctpop
 ; SSE42: Found an estimated cost of 12 for instruction:   %ctpop
-; AVX1: Found an estimated cost of 12 for instruction:   %ctpop
+; AVX1: Found an estimated cost of 14 for instruction:   %ctpop
 ; AVX2: Found an estimated cost of 6 for instruction:   %ctpop
   %ctpop = call <32 x i8> @llvm.ctpop.v32i8(<32 x i8> %a)
   ret <32 x i8> %ctpop
@@ -229,7 +229,7 @@ define <4 x i64> @var_ctlz_v4i64(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v4i64':
 ; SSE2: Found an estimated cost of 50 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 46 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 46 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 48 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 23 for instruction:   %ctlz
   %ctlz = call <4 x i64> @llvm.ctlz.v4i64(<4 x i64> %a, i1 0)
   ret <4 x i64> %ctlz
@@ -239,7 +239,7 @@ define <4 x i64> @var_ctlz_v4i64u(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v4i64u':
 ; SSE2: Found an estimated cost of 50 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 46 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 46 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 48 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 23 for instruction:   %ctlz
   %ctlz = call <4 x i64> @llvm.ctlz.v4i64(<4 x i64> %a, i1 1)
   ret <4 x i64> %ctlz
@@ -267,7 +267,7 @@ define <8 x i32> @var_ctlz_v8i32(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v8i32':
 ; SSE2: Found an estimated cost of 52 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 36 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 36 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 38 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 18 for instruction:   %ctlz
   %ctlz = call <8 x i32> @llvm.ctlz.v8i32(<8 x i32> %a, i1 0)
   ret <8 x i32> %ctlz
@@ -277,7 +277,7 @@ define <8 x i32> @var_ctlz_v8i32u(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v8i32u':
 ; SSE2: Found an estimated cost of 52 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 36 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 36 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 38 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 18 for instruction:   %ctlz
   %ctlz = call <8 x i32> @llvm.ctlz.v8i32(<8 x i32> %a, i1 1)
   ret <8 x i32> %ctlz
@@ -305,7 +305,7 @@ define <16 x i16> @var_ctlz_v16i16(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v16i16':
 ; SSE2: Found an estimated cost of 40 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 28 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 28 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 30 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 14 for instruction:   %ctlz
   %ctlz = call <16 x i16> @llvm.ctlz.v16i16(<16 x i16> %a, i1 0)
   ret <16 x i16> %ctlz
@@ -315,7 +315,7 @@ define <16 x i16> @var_ctlz_v16i16u(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v16i16u':
 ; SSE2: Found an estimated cost of 40 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 28 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 28 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 30 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 14 for instruction:   %ctlz
   %ctlz = call <16 x i16> @llvm.ctlz.v16i16(<16 x i16> %a, i1 1)
   ret <16 x i16> %ctlz
@@ -343,7 +343,7 @@ define <32 x i8> @var_ctlz_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v32i8':
 ; SSE2: Found an estimated cost of 34 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 18 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 18 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 20 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 9 for instruction:   %ctlz
   %ctlz = call <32 x i8> @llvm.ctlz.v32i8(<32 x i8> %a, i1 0)
   ret <32 x i8> %ctlz
@@ -353,7 +353,7 @@ define <32 x i8> @var_ctlz_v32i8u(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_ctlz_v32i8u':
 ; SSE2: Found an estimated cost of 34 for instruction:   %ctlz
 ; SSE42: Found an estimated cost of 18 for instruction:   %ctlz
-; AVX1: Found an estimated cost of 18 for instruction:   %ctlz
+; AVX1: Found an estimated cost of 20 for instruction:   %ctlz
 ; AVX2: Found an estimated cost of 9 for instruction:   %ctlz
   %ctlz = call <32 x i8> @llvm.ctlz.v32i8(<32 x i8> %a, i1 1)
   ret <32 x i8> %ctlz
@@ -456,7 +456,7 @@ define <4 x i64> @var_cttz_v4i64(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v4i64':
 ; SSE2: Found an estimated cost of 28 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 20 for instruction:   %cttz
-; AVX1: Found an estimated cost of 20 for instruction:   %cttz
+; AVX1: Found an estimated cost of 22 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 10 for instruction:   %cttz
   %cttz = call <4 x i64> @llvm.cttz.v4i64(<4 x i64> %a, i1 0)
   ret <4 x i64> %cttz
@@ -466,7 +466,7 @@ define <4 x i64> @var_cttz_v4i64u(<4 x i64> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v4i64u':
 ; SSE2: Found an estimated cost of 28 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 20 for instruction:   %cttz
-; AVX1: Found an estimated cost of 20 for instruction:   %cttz
+; AVX1: Found an estimated cost of 22 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 10 for instruction:   %cttz
   %cttz = call <4 x i64> @llvm.cttz.v4i64(<4 x i64> %a, i1 1)
   ret <4 x i64> %cttz
@@ -494,7 +494,7 @@ define <8 x i32> @var_cttz_v8i32(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v8i32':
 ; SSE2: Found an estimated cost of 36 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 28 for instruction:   %cttz
-; AVX1: Found an estimated cost of 28 for instruction:   %cttz
+; AVX1: Found an estimated cost of 30 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 14 for instruction:   %cttz
   %cttz = call <8 x i32> @llvm.cttz.v8i32(<8 x i32> %a, i1 0)
   ret <8 x i32> %cttz
@@ -504,7 +504,7 @@ define <8 x i32> @var_cttz_v8i32u(<8 x i32> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v8i32u':
 ; SSE2: Found an estimated cost of 36 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 28 for instruction:   %cttz
-; AVX1: Found an estimated cost of 28 for instruction:   %cttz
+; AVX1: Found an estimated cost of 30 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 14 for instruction:   %cttz
   %cttz = call <8 x i32> @llvm.cttz.v8i32(<8 x i32> %a, i1 1)
   ret <8 x i32> %cttz
@@ -532,7 +532,7 @@ define <16 x i16> @var_cttz_v16i16(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v16i16':
 ; SSE2: Found an estimated cost of 32 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 24 for instruction:   %cttz
-; AVX1: Found an estimated cost of 24 for instruction:   %cttz
+; AVX1: Found an estimated cost of 26 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 12 for instruction:   %cttz
   %cttz = call <16 x i16> @llvm.cttz.v16i16(<16 x i16> %a, i1 0)
   ret <16 x i16> %cttz
@@ -542,7 +542,7 @@ define <16 x i16> @var_cttz_v16i16u(<16 x i16> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v16i16u':
 ; SSE2: Found an estimated cost of 32 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 24 for instruction:   %cttz
-; AVX1: Found an estimated cost of 24 for instruction:   %cttz
+; AVX1: Found an estimated cost of 26 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 12 for instruction:   %cttz
   %cttz = call <16 x i16> @llvm.cttz.v16i16(<16 x i16> %a, i1 1)
   ret <16 x i16> %cttz
@@ -570,7 +570,7 @@ define <32 x i8> @var_cttz_v32i8(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v32i8':
 ; SSE2: Found an estimated cost of 26 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 18 for instruction:   %cttz
-; AVX1: Found an estimated cost of 18 for instruction:   %cttz
+; AVX1: Found an estimated cost of 20 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 9 for instruction:   %cttz
   %cttz = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %a, i1 0)
   ret <32 x i8> %cttz
@@ -580,7 +580,7 @@ define <32 x i8> @var_cttz_v32i8u(<32 x i8> %a) {
 ; CHECK: 'Cost Model Analysis' for function 'var_cttz_v32i8u':
 ; SSE2: Found an estimated cost of 26 for instruction:   %cttz
 ; SSE42: Found an estimated cost of 18 for instruction:   %cttz
-; AVX1: Found an estimated cost of 18 for instruction:   %cttz
+; AVX1: Found an estimated cost of 20 for instruction:   %cttz
 ; AVX2: Found an estimated cost of 9 for instruction:   %cttz
   %cttz = call <32 x i8> @llvm.cttz.v32i8(<32 x i8> %a, i1 1)
   ret <32 x i8> %cttz