Skip to content

Commit 201877d

Browse files
RKSimonRKSimon
committed
[CostModel][X86] Improve accuracy of vXi8 multiply reduction costs
After rG47321c311bdbe0145b9bf45d822185c37b19fa50 we promote vXi8 reductions to vXi16 to create a much faster PMULLW mul reduction, followed by a (free) truncation. This avoids the high cost of repeated vXi8 multiplications (which extend+multiply+truncate to/from vXi16 types....). Fixes the missing vXi8 mul reduction vectorization in PR42674 (Comment #20) 'mul16' test case.
1 parent 6eb5b06 commit 201877d

File tree

2 files changed

+74
-44
lines changed

2 files changed

+74
-44
lines changed

‎llvm/lib/Target/X86/X86TargetTransformInfo.cpp

+10
Original file line numberDiff line numberDiff line change
@@ -3408,6 +3408,16 @@ int X86TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
34083408

34093409
auto *ValVTy = cast<FixedVectorType>(ValTy);
34103410

3411+
// Special case: vXi8 mul reductions are performed as vXi16.
3412+
if (ISD == ISD::MUL && MTy.getScalarType() == MVT::i8) {
3413+
auto *WideSclTy = IntegerType::get(ValVTy->getContext(), 16);
3414+
auto *WideVecTy = FixedVectorType::get(WideSclTy, ValVTy->getNumElements());
3415+
return getCastInstrCost(Instruction::ZExt, WideVecTy, ValTy,
3416+
TargetTransformInfo::CastContextHint::None,
3417+
CostKind) +
3418+
getArithmeticReductionCost(Opcode, WideVecTy, IsPairwise, CostKind);
3419+
}
3420+
34113421
unsigned ArithmeticCost = 0;
34123422
if (LT.first != 1 && MTy.isVector() &&
34133423
MTy.getVectorNumElements() < ValVTy->getNumElements()) {

‎llvm/test/Analysis/CostModel/X86/reduce-mul.ll

+64-44
Original file line numberDiff line numberDiff line change
@@ -187,64 +187,84 @@ define i32 @reduce_i16(i32 %arg) {
187187
}
188188

189189
define i32 @reduce_i8(i32 %arg) {
190-
; SSE-LABEL: 'reduce_i8'
191-
; SSE-NEXT: Cost Model: Found an estimated cost of 14 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
192-
; SSE-NEXT: Cost Model: Found an estimated cost of 27 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
193-
; SSE-NEXT: Cost Model: Found an estimated cost of 40 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
194-
; SSE-NEXT: Cost Model: Found an estimated cost of 53 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
195-
; SSE-NEXT: Cost Model: Found an estimated cost of 65 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
196-
; SSE-NEXT: Cost Model: Found an estimated cost of 89 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
197-
; SSE-NEXT: Cost Model: Found an estimated cost of 137 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
198-
; SSE-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
190+
; SSE2-LABEL: 'reduce_i8'
191+
; SSE2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
192+
; SSE2-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
193+
; SSE2-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
194+
; SSE2-NEXT: Cost Model: Found an estimated cost of 11 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
195+
; SSE2-NEXT: Cost Model: Found an estimated cost of 16 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
196+
; SSE2-NEXT: Cost Model: Found an estimated cost of 26 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
197+
; SSE2-NEXT: Cost Model: Found an estimated cost of 46 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
198+
; SSE2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
199+
;
200+
; SSSE3-LABEL: 'reduce_i8'
201+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
202+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
203+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
204+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 11 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
205+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 16 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
206+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 26 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
207+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 46 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
208+
; SSSE3-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
209+
;
210+
; SSE42-LABEL: 'reduce_i8'
211+
; SSE42-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
212+
; SSE42-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
213+
; SSE42-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
214+
; SSE42-NEXT: Cost Model: Found an estimated cost of 10 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
215+
; SSE42-NEXT: Cost Model: Found an estimated cost of 14 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
216+
; SSE42-NEXT: Cost Model: Found an estimated cost of 22 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
217+
; SSE42-NEXT: Cost Model: Found an estimated cost of 38 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
218+
; SSE42-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
199219
;
200220
; AVX1-LABEL: 'reduce_i8'
201-
; AVX1-NEXT: Cost Model: Found an estimated cost of 14 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
202-
; AVX1-NEXT: Cost Model: Found an estimated cost of 27 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
203-
; AVX1-NEXT: Cost Model: Found an estimated cost of 40 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
204-
; AVX1-NEXT: Cost Model: Found an estimated cost of 53 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
205-
; AVX1-NEXT: Cost Model: Found an estimated cost of 66 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
206-
; AVX1-NEXT: Cost Model: Found an estimated cost of 92 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
207-
; AVX1-NEXT: Cost Model: Found an estimated cost of 144 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
221+
; AVX1-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
222+
; AVX1-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
223+
; AVX1-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
224+
; AVX1-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
225+
; AVX1-NEXT: Cost Model: Found an estimated cost of 22 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
226+
; AVX1-NEXT: Cost Model: Found an estimated cost of 39 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
227+
; AVX1-NEXT: Cost Model: Found an estimated cost of 73 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
208228
; AVX1-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
209229
;
210230
; AVX2-LABEL: 'reduce_i8'
211-
; AVX2-NEXT: Cost Model: Found an estimated cost of 9 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
212-
; AVX2-NEXT: Cost Model: Found an estimated cost of 17 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
213-
; AVX2-NEXT: Cost Model: Found an estimated cost of 25 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
214-
; AVX2-NEXT: Cost Model: Found an estimated cost of 33 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
215-
; AVX2-NEXT: Cost Model: Found an estimated cost of 41 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
216-
; AVX2-NEXT: Cost Model: Found an estimated cost of 58 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
217-
; AVX2-NEXT: Cost Model: Found an estimated cost of 92 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
231+
; AVX2-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
232+
; AVX2-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
233+
; AVX2-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
234+
; AVX2-NEXT: Cost Model: Found an estimated cost of 10 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
235+
; AVX2-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
236+
; AVX2-NEXT: Cost Model: Found an estimated cost of 18 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
237+
; AVX2-NEXT: Cost Model: Found an estimated cost of 28 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
218238
; AVX2-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
219239
;
220240
; AVX512F-LABEL: 'reduce_i8'
221-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 7 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
222-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
223-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 19 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
224-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 25 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
225-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 31 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
226-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 45 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
227-
; AVX512F-NEXT: Cost Model: Found an estimated cost of 71 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
241+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
242+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
243+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
244+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 10 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
245+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 14 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
246+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 20 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
247+
; AVX512F-NEXT: Cost Model: Found an estimated cost of 31 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
228248
; AVX512F-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
229249
;
230250
; AVX512BW-LABEL: 'reduce_i8'
231-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
232-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 11 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
233-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 16 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
234-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 21 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
235-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 26 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
236-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 31 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
237-
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 42 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
251+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
252+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
253+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
254+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 10 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
255+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 12 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
256+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 15 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
257+
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 20 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
238258
; AVX512BW-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
239259
;
240260
; AVX512DQ-LABEL: 'reduce_i8'
241-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 7 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
242-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 13 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
243-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 19 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
244-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 25 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
245-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 31 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
246-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 45 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
247-
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 71 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
261+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 4 for instruction: %V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)
262+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %V4 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> undef)
263+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %V8 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> undef)
264+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 10 for instruction: %V16 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> undef)
265+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 14 for instruction: %V32 = call i8 @llvm.vector.reduce.mul.v32i8(<32 x i8> undef)
266+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 20 for instruction: %V64 = call i8 @llvm.vector.reduce.mul.v64i8(<64 x i8> undef)
267+
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 31 for instruction: %V128 = call i8 @llvm.vector.reduce.mul.v128i8(<128 x i8> undef)
248268
; AVX512DQ-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret i32 undef
249269
;
250270
%V2 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> undef)

0 commit comments

Comments
 (0)