[llvm] d181fd9 - [CostModel][X86] Drop some hard coded fp<->int scalarization costs

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2021-07-02T14:29:32+01:00
New Revision: d181fd918d18cbd99768f025e14a69d35d275f14

URL: https://github.com/llvm/llvm-project/commit/d181fd918d18cbd99768f025e14a69d35d275f14
DIFF: https://github.com/llvm/llvm-project/commit/d181fd918d18cbd99768f025e14a69d35d275f14.diff

LOG: [CostModel][X86] Drop some hard coded fp<->int scalarization costs

Scalarization costs handling is a lot better now, and the hard coded costs were higher than the worse case numbers from the script in D103695

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86TargetTransformInfo.cpp
    llvm/test/Analysis/CostModel/X86/sitofp.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index d55cd8a8c7a8..9eb5abe4dd9b 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -1977,13 +1977,6 @@ InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
     { ISD::UINT_TO_FP,  MVT::v8f64, MVT::v8i32, 10 },
     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i64, 5 },
     { ISD::UINT_TO_FP,  MVT::v4f64, MVT::v4i64, 6 },
-    // The generic code to compute the scalar overhead is currently broken.
-    // Workaround this limitation by estimating the scalarization overhead
-    // here. We have roughly 10 instructions per scalar element.
-    // Multiply that by the vector width.
-    // FIXME: remove that when PR19268 is fixed.
-    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
-    { ISD::SINT_TO_FP,  MVT::v4f64, MVT::v4i64, 13 },
 
     { ISD::FP_TO_SINT,  MVT::v8i8,  MVT::v8f32, 4 },
     { ISD::FP_TO_SINT,  MVT::v4i8,  MVT::v4f64, 3 },
@@ -2003,12 +1996,6 @@ InstructionCost X86TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
     { ISD::FP_TO_UINT,  MVT::v8i16, MVT::v8f32, 3 },
     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f32, 9 },
     { ISD::FP_TO_UINT,  MVT::v8i32, MVT::v8f64, 19 },
-    // This node is expanded into scalarized operations but BasicTTI is overly
-    // optimistic estimating its cost.  It computes 3 per element (one
-    // vector-extract, one scalar conversion and one vector-insert).  The
-    // problem is that the inserts form a read-modify-write chain so latency
-    // should be factored in too.  Inflating the cost per element by 1.
-    { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f64, 4*4 },
 
     { ISD::FP_EXTEND,   MVT::v4f64,  MVT::v4f32,  1 },
     { ISD::FP_ROUND,    MVT::v4f32,  MVT::v4f64,  1 },

diff  --git a/llvm/test/Analysis/CostModel/X86/sitofp.ll b/llvm/test/Analysis/CostModel/X86/sitofp.ll
index b3c400c93b9f..b327454c1d09 100644
--- a/llvm/test/Analysis/CostModel/X86/sitofp.ll
+++ b/llvm/test/Analysis/CostModel/X86/sitofp.ll
@@ -122,14 +122,14 @@ define i32 @sitofp_i64_double() {
 ; AVX-LABEL: 'sitofp_i64_double'
 ; AVX-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %cvt_i64_f64 = sitofp i64 undef to double
 ; AVX-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %cvt_v2i64_v2f64 = sitofp <2 x i64> undef to <2 x double>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 13 for instruction: %cvt_v4i64_v4f64 = sitofp <4 x i64> undef to <4 x double>
-; AVX-NEXT:  Cost Model: Found an estimated cost of 26 for instruction: %cvt_v8i64_v8f64 = sitofp <8 x i64> undef to <8 x double>
+; AVX-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %cvt_v4i64_v4f64 = sitofp <4 x i64> undef to <4 x double>
+; AVX-NEXT:  Cost Model: Found an estimated cost of 22 for instruction: %cvt_v8i64_v8f64 = sitofp <8 x i64> undef to <8 x double>
 ; AVX-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;
 ; AVX512F-LABEL: 'sitofp_i64_double'
 ; AVX512F-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: %cvt_i64_f64 = sitofp i64 undef to double
 ; AVX512F-NEXT:  Cost Model: Found an estimated cost of 8 for instruction: %cvt_v2i64_v2f64 = sitofp <2 x i64> undef to <2 x double>
-; AVX512F-NEXT:  Cost Model: Found an estimated cost of 13 for instruction: %cvt_v4i64_v4f64 = sitofp <4 x i64> undef to <4 x double>
+; AVX512F-NEXT:  Cost Model: Found an estimated cost of 11 for instruction: %cvt_v4i64_v4f64 = sitofp <4 x i64> undef to <4 x double>
 ; AVX512F-NEXT:  Cost Model: Found an estimated cost of 25 for instruction: %cvt_v8i64_v8f64 = sitofp <8 x i64> undef to <8 x double>
 ; AVX512F-NEXT:  Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
 ;