[CostModel][X86] Added mul costs for vXi8 vectors

More realistic v16i8/v32i8/v64i8 MUL costs - we have to extend to vXi16, use PMULLW and then truncate the result

llvm-svn: 286838

diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index 5b3091e..a2e4b3e 100644 (file)
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -218,15 +218,19 @@ int X86TTIImpl::getArithmeticInstrCost(
   }
 
   static const CostTblEntry AVX512BWCostTable[] = {
+    { ISD::MUL,   MVT::v64i8,     11 }, // extend/pmullw/trunc sequence.
+    { ISD::MUL,   MVT::v32i8,      4 }, // extend/pmullw/trunc sequence.
+    { ISD::MUL,   MVT::v16i8,      4 }, // extend/pmullw/trunc sequence.
+
     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
     { ISD::SDIV,  MVT::v64i8,  64*20 },
     { ISD::SDIV,  MVT::v32i16, 32*20 },
     { ISD::SDIV,  MVT::v16i32, 16*20 },
-    { ISD::SDIV,  MVT::v8i64,  8*20 },
+    { ISD::SDIV,  MVT::v8i64,   8*20 },
     { ISD::UDIV,  MVT::v64i8,  64*20 },
     { ISD::UDIV,  MVT::v32i16, 32*20 },
     { ISD::UDIV,  MVT::v16i32, 16*20 },
-    { ISD::UDIV,  MVT::v8i64,  8*20 },
+    { ISD::UDIV,  MVT::v8i64,   8*20 },
   };
 
   // Look for AVX512BW lowering tricks for custom cases.
@@ -240,9 +244,12 @@ int X86TTIImpl::getArithmeticInstrCost(
     { ISD::SHL,     MVT::v16i32,    1 },
     { ISD::SRL,     MVT::v16i32,    1 },
     { ISD::SRA,     MVT::v16i32,    1 },
-    { ISD::SHL,     MVT::v8i64,    1 },
-    { ISD::SRL,     MVT::v8i64,    1 },
-    { ISD::SRA,     MVT::v8i64,    1 },
+    { ISD::SHL,     MVT::v8i64,     1 },
+    { ISD::SRL,     MVT::v8i64,     1 },
+    { ISD::SRA,     MVT::v8i64,     1 },
+
+    { ISD::MUL,     MVT::v32i8,    13 }, // extend/pmullw/trunc sequence.
+    { ISD::MUL,     MVT::v16i8,     5 }, // extend/pmullw/trunc sequence.
   };
 
   if (ST->hasAVX512()) {
@@ -324,6 +331,10 @@ int X86TTIImpl::getArithmeticInstrCost(
     { ISD::SRA,  MVT::v16i16,     10 }, // extend/vpsravd/pack sequence.
     { ISD::SRA,  MVT::v2i64,       4 }, // srl/xor/sub sequence.
     { ISD::SRA,  MVT::v4i64,       4 }, // srl/xor/sub sequence.
+
+    { ISD::MUL,   MVT::v32i8,     17 }, // extend/pmullw/trunc sequence.
+    { ISD::MUL,   MVT::v16i8,      7 }, // extend/pmullw/trunc sequence.
+
     { ISD::FDIV,  MVT::f32,        7 }, // Haswell from http://www.agner.org/
     { ISD::FDIV,  MVT::v4f32,      7 }, // Haswell from http://www.agner.org/
     { ISD::FDIV,  MVT::v8f32,     14 }, // Haswell from http://www.agner.org/
@@ -340,12 +351,15 @@ int X86TTIImpl::getArithmeticInstrCost(
   }
 
   static const CostTblEntry AVXCustomCostTable[] = {
+    { ISD::MUL,   MVT::v32i8,  26 }, // extend/pmullw/trunc sequence.
+
     { ISD::FDIV,  MVT::f32,    14 }, // SNB from http://www.agner.org/
     { ISD::FDIV,  MVT::v4f32,  14 }, // SNB from http://www.agner.org/
     { ISD::FDIV,  MVT::v8f32,  28 }, // SNB from http://www.agner.org/
     { ISD::FDIV,  MVT::f64,    22 }, // SNB from http://www.agner.org/
     { ISD::FDIV,  MVT::v2f64,  22 }, // SNB from http://www.agner.org/
     { ISD::FDIV,  MVT::v4f64,  44 }, // SNB from http://www.agner.org/
+
     // Vectorizing division is a bad idea. See the SSE2 table for more comments.
     { ISD::SDIV,  MVT::v32i8,  32*20 },
     { ISD::SDIV,  MVT::v16i16, 16*20 },
@@ -494,6 +508,8 @@ int X86TTIImpl::getArithmeticInstrCost(
     { ISD::SRA,  MVT::v2i64,    12 }, // srl/xor/sub sequence.
     { ISD::SRA,  MVT::v4i64,  2*12 }, // srl/xor/sub sequence.
 
+    { ISD::MUL,  MVT::v16i8,    12 }, // extend/pmullw/trunc sequence.
+
     { ISD::FDIV, MVT::f32,      23 }, // Pentium IV from http://www.agner.org/
     { ISD::FDIV, MVT::v4f32,    39 }, // Pentium IV from http://www.agner.org/
     { ISD::FDIV, MVT::f64,      38 }, // Pentium IV from http://www.agner.org/

diff --git a/llvm/test/Analysis/CostModel/X86/arith.ll b/llvm/test/Analysis/CostModel/X86/arith.ll
index aa204db..55766e2 100644 (file)
--- a/llvm/test/Analysis/CostModel/X86/arith.ll
+++ b/llvm/test/Analysis/CostModel/X86/arith.ll
@@ -490,24 +490,26 @@ define i32 @mul(i32 %arg) {
   ; AVX512BW: cost of 1 {{.*}} %I = mul
   %I = mul <32 x i16> undef, undef
 
-  ; SSSE3: cost of 2 {{.*}} %J = mul
-  ; SSE42: cost of 2 {{.*}} %J = mul
-  ; AVX: cost of 2 {{.*}} %J = mul
-  ; AVX2: cost of 2 {{.*}} %J = mul
-  ; AVX512: cost of 2 {{.*}} %J = mul
+  ; SSSE3: cost of 12 {{.*}} %J = mul
+  ; SSE42: cost of 12 {{.*}} %J = mul
+  ; AVX: cost of 12 {{.*}} %J = mul
+  ; AVX2: cost of 7 {{.*}} %J = mul
+  ; AVX512F: cost of 5 {{.*}} %J = mul
+  ; AVX512BW: cost of 4 {{.*}} %J = mul
   %J = mul <16 x i8> undef, undef
-  ; SSSE3: cost of 4 {{.*}} %K = mul
-  ; SSE42: cost of 4 {{.*}} %K = mul
-  ; AVX: cost of 2 {{.*}} %K = mul
-  ; AVX2: cost of 2 {{.*}} %K = mul
-  ; AVX512: cost of 2 {{.*}} %K = mul
+  ; SSSE3: cost of 24 {{.*}} %K = mul
+  ; SSE42: cost of 24 {{.*}} %K = mul
+  ; AVX: cost of 26 {{.*}} %K = mul
+  ; AVX2: cost of 17 {{.*}} %K = mul
+  ; AVX512F: cost of 13 {{.*}} %K = mul
+  ; AVX512BW: cost of 4 {{.*}} %K = mul
   %K = mul <32 x i8> undef, undef
-  ; SSSE3: cost of 8 {{.*}} %L = mul
-  ; SSE42: cost of 8 {{.*}} %L = mul
-  ; AVX: cost of 4 {{.*}} %L = mul
-  ; AVX2: cost of 4 {{.*}} %L = mul
-  ; AVX512F: cost of 4 {{.*}} %L = mul
-  ; AVX512BW: cost of 2 {{.*}} %L = mul
+  ; SSSE3: cost of 48 {{.*}} %L = mul
+  ; SSE42: cost of 48 {{.*}} %L = mul
+  ; AVX: cost of 52 {{.*}} %L = mul
+  ; AVX2: cost of 34 {{.*}} %L = mul
+  ; AVX512F: cost of 26 {{.*}} %L = mul
+  ; AVX512BW: cost of 11 {{.*}} %L = mul
   %L = mul <64 x i8> undef, undef
 
   ret i32 undef