Skip to content

Commit e6cf292

Browse files
RKSimonRKSimon
committed
[InstCombine][X86] Add some basic knownbits tests for PMULH/PMULHU
Recognise known zero upperbits
1 parent 2c0add9 commit e6cf292

File tree

2 files changed

+98
-0
lines changed

2 files changed

+98
-0
lines changed

‎llvm/test/Transforms/InstCombine/X86/x86-pmulh.ll

Lines changed: 49 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -233,3 +233,52 @@ define <32 x i16> @elts_pmulh_512(<32 x i16> %a0, <32 x i16> %a1) {
233233
%4 = shufflevector <32 x i16> %3, <32 x i16> poison, <32 x i32> zeroinitializer
234234
ret <32 x i16> %4
235235
}
236+
237+
;
238+
; Known Bits
239+
;
240+
241+
define <8 x i16> @known_pmulh_128(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> %a2) {
242+
; CHECK-LABEL: @known_pmulh_128(
243+
; CHECK-NEXT: [[X0:%.*]] = lshr <8 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
244+
; CHECK-NEXT: [[X1:%.*]] = and <8 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
245+
; CHECK-NEXT: [[M:%.*]] = tail call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> [[X0]], <8 x i16> [[X1]])
246+
; CHECK-NEXT: [[R:%.*]] = add <8 x i16> [[M]], [[A2:%.*]]
247+
; CHECK-NEXT: ret <8 x i16> [[R]]
248+
;
249+
%x0 = lshr <8 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
250+
%x1 = and <8 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
251+
%m = tail call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> %x0, <8 x i16> %x1)
252+
%r = add <8 x i16> %m, %a2
253+
ret <8 x i16> %r
254+
}
255+
256+
define <16 x i16> @known_pmulh_256(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> %a2) {
257+
; CHECK-LABEL: @known_pmulh_256(
258+
; CHECK-NEXT: [[X0:%.*]] = lshr <16 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
259+
; CHECK-NEXT: [[X1:%.*]] = and <16 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
260+
; CHECK-NEXT: [[M:%.*]] = tail call <16 x i16> @llvm.x86.avx2.pmulh.w(<16 x i16> [[X0]], <16 x i16> [[X1]])
261+
; CHECK-NEXT: [[R:%.*]] = add <16 x i16> [[M]], [[A2:%.*]]
262+
; CHECK-NEXT: ret <16 x i16> [[R]]
263+
;
264+
%x0 = lshr <16 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
265+
%x1 = and <16 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
266+
%m = tail call <16 x i16> @llvm.x86.avx2.pmulh.w(<16 x i16> %x0, <16 x i16> %x1)
267+
%r = add <16 x i16> %m, %a2
268+
ret <16 x i16> %r
269+
}
270+
271+
define <32 x i16> @known_pmulh_512(<32 x i16> %a0, <32 x i16> %a1, <32 x i16> %a2) {
272+
; CHECK-LABEL: @known_pmulh_512(
273+
; CHECK-NEXT: [[X0:%.*]] = lshr <32 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
274+
; CHECK-NEXT: [[X1:%.*]] = and <32 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
275+
; CHECK-NEXT: [[M:%.*]] = tail call <32 x i16> @llvm.x86.avx512.pmulh.w.512(<32 x i16> [[X0]], <32 x i16> [[X1]])
276+
; CHECK-NEXT: [[R:%.*]] = add <32 x i16> [[M]], [[A2:%.*]]
277+
; CHECK-NEXT: ret <32 x i16> [[R]]
278+
;
279+
%x0 = lshr <32 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
280+
%x1 = and <32 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
281+
%m = tail call <32 x i16> @llvm.x86.avx512.pmulh.w.512(<32 x i16> %x0, <32 x i16> %x1)
282+
%r = add <32 x i16> %m, %a2
283+
ret <32 x i16> %r
284+
}

‎llvm/test/Transforms/InstCombine/X86/x86-pmulhu.ll

Lines changed: 49 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -227,3 +227,52 @@ define <32 x i16> @elts_pmulhu_512(<32 x i16> %a0, <32 x i16> %a1) {
227227
%4 = shufflevector <32 x i16> %3, <32 x i16> poison, <32 x i32> zeroinitializer
228228
ret <32 x i16> %4
229229
}
230+
231+
;
232+
; Known Bits
233+
;
234+
235+
define <8 x i16> @known_pmulhu_128(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> %a2) {
236+
; CHECK-LABEL: @known_pmulhu_128(
237+
; CHECK-NEXT: [[X0:%.*]] = lshr <8 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
238+
; CHECK-NEXT: [[X1:%.*]] = and <8 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
239+
; CHECK-NEXT: [[M:%.*]] = tail call <8 x i16> @llvm.x86.sse2.pmulhu.w(<8 x i16> [[X0]], <8 x i16> [[X1]])
240+
; CHECK-NEXT: [[R:%.*]] = add <8 x i16> [[M]], [[A2:%.*]]
241+
; CHECK-NEXT: ret <8 x i16> [[R]]
242+
;
243+
%x0 = lshr <8 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
244+
%x1 = and <8 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
245+
%m = tail call <8 x i16> @llvm.x86.sse2.pmulhu.w(<8 x i16> %x0, <8 x i16> %x1)
246+
%r = add <8 x i16> %m, %a2
247+
ret <8 x i16> %r
248+
}
249+
250+
define <16 x i16> @known_pmulhu_256(<16 x i16> %a0, <16 x i16> %a1, <16 x i16> %a2) {
251+
; CHECK-LABEL: @known_pmulhu_256(
252+
; CHECK-NEXT: [[X0:%.*]] = lshr <16 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
253+
; CHECK-NEXT: [[X1:%.*]] = and <16 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
254+
; CHECK-NEXT: [[M:%.*]] = tail call <16 x i16> @llvm.x86.avx2.pmulhu.w(<16 x i16> [[X0]], <16 x i16> [[X1]])
255+
; CHECK-NEXT: [[R:%.*]] = add <16 x i16> [[M]], [[A2:%.*]]
256+
; CHECK-NEXT: ret <16 x i16> [[R]]
257+
;
258+
%x0 = lshr <16 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
259+
%x1 = and <16 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
260+
%m = tail call <16 x i16> @llvm.x86.avx2.pmulhu.w(<16 x i16> %x0, <16 x i16> %x1)
261+
%r = add <16 x i16> %m, %a2
262+
ret <16 x i16> %r
263+
}
264+
265+
define <32 x i16> @known_pmulhu_512(<32 x i16> %a0, <32 x i16> %a1, <32 x i16> %a2) {
266+
; CHECK-LABEL: @known_pmulhu_512(
267+
; CHECK-NEXT: [[X0:%.*]] = lshr <32 x i16> [[A0:%.*]], <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
268+
; CHECK-NEXT: [[X1:%.*]] = and <32 x i16> [[A1:%.*]], <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
269+
; CHECK-NEXT: [[M:%.*]] = tail call <32 x i16> @llvm.x86.avx512.pmulhu.w.512(<32 x i16> [[X0]], <32 x i16> [[X1]])
270+
; CHECK-NEXT: [[R:%.*]] = add <32 x i16> [[M]], [[A2:%.*]]
271+
; CHECK-NEXT: ret <32 x i16> [[R]]
272+
;
273+
%x0 = lshr <32 x i16> %a0, <i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>
274+
%x1 = and <32 x i16> %a1, <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>
275+
%m = tail call <32 x i16> @llvm.x86.avx512.pmulhu.w.512(<32 x i16> %x0, <32 x i16> %x1)
276+
%r = add <32 x i16> %m, %a2
277+
ret <32 x i16> %r
278+
}

0 commit comments

Comments
 (0)