[llvm] 1ba9ec0 - [AArch64] Update FP16 vector cmp costs
David Green via llvm-commits
llvm-commits at lists.llvm.org
Sun May 14 15:28:17 PDT 2023
Author: David Green
Date: 2023-05-14T23:28:11+01:00
New Revision: 1ba9ec0d10c8c95216d8d7fc47f028344c1640bb
URL: https://github.com/llvm/llvm-project/commit/1ba9ec0d10c8c95216d8d7fc47f028344c1640bb
DIFF: https://github.com/llvm/llvm-project/commit/1ba9ec0d10c8c95216d8d7fc47f028344c1640bb.diff
LOG: [AArch64] Update FP16 vector cmp costs
Without FP16, a fp16 v4f16 comparison will be converted to a v4f32 and back.
v8f16 get scalarized currently. Update the costs of v4f16 to match.
Added:
Modified:
llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
llvm/test/Analysis/CostModel/AArch64/fptoi_sat.ll
llvm/test/Analysis/CostModel/AArch64/reduce-minmax.ll
llvm/test/Analysis/CostModel/AArch64/vector-select.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 91224cedea8b0..063fdd2dfa6a1 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -2585,21 +2585,11 @@ InstructionCost AArch64TTIImpl::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
}
}
- // Cost vXf16 FCmp without FP16 support as if it is scalarized.
- // FIXME: This isn't correct, but matches previous behavior.
- if (isa<FixedVectorType>(ValTy) && ISD == ISD::SETCC &&
- ValTy->getScalarType()->isHalfTy() && !ST->hasFullFP16()) {
- unsigned Num = cast<FixedVectorType>(ValTy)->getNumElements();
- if (CondTy)
- CondTy = CondTy->getScalarType();
- InstructionCost Cost = getCmpSelInstrCost(
- Opcode, ValTy->getScalarType(), CondTy, VecPred, CostKind, I);
-
- // Return the cost of multiple scalar invocation plus the cost of
- // inserting and extracting the values.
- return getScalarizationOverhead(cast<FixedVectorType>(ValTy), /*Insert*/ true,
- /*Extract*/ false, CostKind) +
- Num * Cost;
+ if (isa<FixedVectorType>(ValTy) && ISD == ISD::SETCC) {
+ auto LT = getTypeLegalizationCost(ValTy);
+ // Cost v4f16 FCmp without FP16 support via converting to v4f32 and back.
+ if (LT.second == MVT::v4f16 && !ST->hasFullFP16())
+ return LT.first * 4; // fcvtl + fcvtl + fcmp + xtn
}
// The base case handles scalable vectors fine for now, since it treats the
diff --git a/llvm/test/Analysis/CostModel/AArch64/fptoi_sat.ll b/llvm/test/Analysis/CostModel/AArch64/fptoi_sat.ll
index eca69189245d7..342fb6da1c285 100644
--- a/llvm/test/Analysis/CostModel/AArch64/fptoi_sat.ll
+++ b/llvm/test/Analysis/CostModel/AArch64/fptoi_sat.ll
@@ -233,25 +233,25 @@ define void @fp16() {
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %f16u32 = call i32 @llvm.fptoui.sat.i32.f16(half undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 5 for instruction: %f16s64 = call i64 @llvm.fptosi.sat.i64.f16(half undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 3 for instruction: %f16u64 = call i64 @llvm.fptoui.sat.i64.f16(half undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 33 for instruction: %v2f16s1 = call <2 x i1> @llvm.fptosi.sat.v2i1.v2f16(<2 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 32 for instruction: %v2f16s1 = call <2 x i1> @llvm.fptosi.sat.v2i1.v2f16(<2 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 23 for instruction: %v2f16u1 = call <2 x i1> @llvm.fptoui.sat.v2i1.v2f16(<2 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 33 for instruction: %v2f16s8 = call <2 x i8> @llvm.fptosi.sat.v2i8.v2f16(<2 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 32 for instruction: %v2f16s8 = call <2 x i8> @llvm.fptosi.sat.v2i8.v2f16(<2 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 23 for instruction: %v2f16u8 = call <2 x i8> @llvm.fptoui.sat.v2i8.v2f16(<2 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 33 for instruction: %v2f16s16 = call <2 x i16> @llvm.fptosi.sat.v2i16.v2f16(<2 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 32 for instruction: %v2f16s16 = call <2 x i16> @llvm.fptosi.sat.v2i16.v2f16(<2 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 23 for instruction: %v2f16u16 = call <2 x i16> @llvm.fptoui.sat.v2i16.v2f16(<2 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 33 for instruction: %v2f16s32 = call <2 x i32> @llvm.fptosi.sat.v2i32.v2f16(<2 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 32 for instruction: %v2f16s32 = call <2 x i32> @llvm.fptosi.sat.v2i32.v2f16(<2 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 23 for instruction: %v2f16u32 = call <2 x i32> @llvm.fptoui.sat.v2i32.v2f16(<2 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 40 for instruction: %v2f16s64 = call <2 x i64> @llvm.fptosi.sat.v2i64.v2f16(<2 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 39 for instruction: %v2f16s64 = call <2 x i64> @llvm.fptosi.sat.v2i64.v2f16(<2 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 30 for instruction: %v2f16u64 = call <2 x i64> @llvm.fptoui.sat.v2i64.v2f16(<2 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 89 for instruction: %v4f16s1 = call <4 x i1> @llvm.fptosi.sat.v4i1.v4f16(<4 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 80 for instruction: %v4f16s1 = call <4 x i1> @llvm.fptosi.sat.v4i1.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 63 for instruction: %v4f16u1 = call <4 x i1> @llvm.fptoui.sat.v4i1.v4f16(<4 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 89 for instruction: %v4f16s8 = call <4 x i8> @llvm.fptosi.sat.v4i8.v4f16(<4 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 80 for instruction: %v4f16s8 = call <4 x i8> @llvm.fptosi.sat.v4i8.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 63 for instruction: %v4f16u8 = call <4 x i8> @llvm.fptoui.sat.v4i8.v4f16(<4 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 89 for instruction: %v4f16s16 = call <4 x i16> @llvm.fptosi.sat.v4i16.v4f16(<4 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 80 for instruction: %v4f16s16 = call <4 x i16> @llvm.fptosi.sat.v4i16.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 63 for instruction: %v4f16u16 = call <4 x i16> @llvm.fptoui.sat.v4i16.v4f16(<4 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 110 for instruction: %v4f16s32 = call <4 x i32> @llvm.fptosi.sat.v4i32.v4f16(<4 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 101 for instruction: %v4f16s32 = call <4 x i32> @llvm.fptosi.sat.v4i32.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 84 for instruction: %v4f16u32 = call <4 x i32> @llvm.fptoui.sat.v4i32.v4f16(<4 x half> undef)
-; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 172 for instruction: %v4f16s64 = call <4 x i64> @llvm.fptosi.sat.v4i64.v4f16(<4 x half> undef)
+; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 163 for instruction: %v4f16s64 = call <4 x i64> @llvm.fptosi.sat.v4i64.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 79 for instruction: %v4f16u64 = call <4 x i64> @llvm.fptoui.sat.v4i64.v4f16(<4 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 250 for instruction: %v8f16s1 = call <8 x i1> @llvm.fptosi.sat.v8i1.v8f16(<8 x half> undef)
; CHECK-NOFP16-NEXT: Cost Model: Found an estimated cost of 192 for instruction: %v8f16u1 = call <8 x i1> @llvm.fptoui.sat.v8i1.v8f16(<8 x half> undef)
diff --git a/llvm/test/Analysis/CostModel/AArch64/reduce-minmax.ll b/llvm/test/Analysis/CostModel/AArch64/reduce-minmax.ll
index 4b58a8a6be482..cf813bae839c6 100644
--- a/llvm/test/Analysis/CostModel/AArch64/reduce-minmax.ll
+++ b/llvm/test/Analysis/CostModel/AArch64/reduce-minmax.ll
@@ -162,8 +162,8 @@ define void @reduce_smax() {
define void @reduce_fmin16() {
; CHECK-NOF16-LABEL: 'reduce_fmin16'
-; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V2f16 = call half @llvm.vector.reduce.fmin.v2f16(<2 x half> undef)
-; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 58 for instruction: %V4f16 = call half @llvm.vector.reduce.fmin.v4f16(<4 x half> undef)
+; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 12 for instruction: %V2f16 = call half @llvm.vector.reduce.fmin.v2f16(<2 x half> undef)
+; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 40 for instruction: %V4f16 = call half @llvm.vector.reduce.fmin.v4f16(<4 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 198 for instruction: %V8f16 = call half @llvm.vector.reduce.fmin.v8f16(<8 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 298 for instruction: %V16f16 = call half @llvm.vector.reduce.fmin.v16f16(<16 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret void
@@ -184,8 +184,8 @@ define void @reduce_fmin16() {
define void @reduce_fmax16() {
; CHECK-NOF16-LABEL: 'reduce_fmax16'
-; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V2f16 = call half @llvm.vector.reduce.fmax.v2f16(<2 x half> undef)
-; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 58 for instruction: %V4f16 = call half @llvm.vector.reduce.fmax.v4f16(<4 x half> undef)
+; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 12 for instruction: %V2f16 = call half @llvm.vector.reduce.fmax.v2f16(<2 x half> undef)
+; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 40 for instruction: %V4f16 = call half @llvm.vector.reduce.fmax.v4f16(<4 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 198 for instruction: %V8f16 = call half @llvm.vector.reduce.fmax.v8f16(<8 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 298 for instruction: %V16f16 = call half @llvm.vector.reduce.fmax.v16f16(<16 x half> undef)
; CHECK-NOF16-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret void
diff --git a/llvm/test/Analysis/CostModel/AArch64/vector-select.ll b/llvm/test/Analysis/CostModel/AArch64/vector-select.ll
index a576c7e3fd293..7d36ec2281978 100644
--- a/llvm/test/Analysis/CostModel/AArch64/vector-select.ll
+++ b/llvm/test/Analysis/CostModel/AArch64/vector-select.ll
@@ -157,7 +157,7 @@ define <2 x i64> @v2i64_select_no_cmp(<2 x i64> %a, <2 x i64> %b, <2 x i1> %cond
define <4 x half> @v4f16_select_ogt(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_ogt
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp ogt <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp ogt <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp ogt <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -241,7 +241,7 @@ define <2 x double> @v2f64_select_ogt(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_oge(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_oge
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp oge <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp oge <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp oge <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -325,7 +325,7 @@ define <2 x double> @v2f64_select_oge(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_olt(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_olt
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp olt <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp olt <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp olt <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -409,7 +409,7 @@ define <2 x double> @v2f64_select_olt(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_ole(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_ole
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp ole <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp ole <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp ole <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -493,7 +493,7 @@ define <2 x double> @v2f64_select_ole(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_oeq(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_oeq
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp oeq <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp oeq <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp oeq <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -577,7 +577,7 @@ define <2 x double> @v2f64_select_oeq(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_one(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_one
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp one <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp one <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp one <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
@@ -671,7 +671,7 @@ define <2 x double> @v2f64_select_one(<2 x double> %a, <2 x double> %b, <2 x dou
define <4 x half> @v4f16_select_une(<4 x half> %a, <4 x half> %b, <4 x half> %c) {
; COST-LABEL: v4f16_select_une
-; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %cmp.1 = fcmp une <4 x half> %a, %b
+; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %cmp.1 = fcmp une <4 x half> %a, %b
; COST-NOFP16-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %cmp.1 = fcmp une <4 x half> %a, %b
; COST-FULLFP16-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %s.1 = select <4 x i1> %cmp.1, <4 x half> %a, <4 x half> %c
More information about the llvm-commits
mailing list