Skip to content

Commit 98992d3

Browse files
dtcxzywdtcxzyw
committed
[InstCombine] Preserve signbit semantics of NaN with fold to fabs
1 parent 2ba455f commit 98992d3

File tree

2 files changed

+29
-12
lines changed

2 files changed

+29
-12
lines changed

‎llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2793,7 +2793,11 @@ static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,
27932793

27942794
// fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false
27952795
// fold (X > +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true
2796-
if (match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) {
2796+
// Note: We require "nnan" for this fold because fcmp ignores the signbit
2797+
// of NAN, but IEEE-754 specifies the signbit of NAN values with
2798+
// fneg/fabs operations.
2799+
if (cast<FPMathOperator>(CondVal)->hasNoNaNs() &&
2800+
match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) {
27972801
if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
27982802
Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
27992803
return IC.replaceInstUsesWith(SI, Fabs);

‎llvm/test/Transforms/InstCombine/fabs.ll‎

Lines changed: 24 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -256,6 +256,19 @@ define double @select_fcmp_ole_zero(double %x) {
256256
; CHECK-LABEL: @select_fcmp_ole_zero(
257257
; CHECK-NEXT: [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
258258
; CHECK-NEXT: ret double [[FABS]]
259+
;
260+
%lezero = fcmp nnan ole double %x, 0.0
261+
%negx = fsub double 0.0, %x
262+
%fabs = select i1 %lezero, double %negx, double %x
263+
ret double %fabs
264+
}
265+
266+
define double @select_fcmp_ole_zero_no_nnan(double %x) {
267+
; CHECK-LABEL: @select_fcmp_ole_zero_no_nnan(
268+
; CHECK-NEXT: [[LEZERO:%.*]] = fcmp ole double [[X:%.*]], 0.000000e+00
269+
; CHECK-NEXT: [[NEGX:%.*]] = fsub double 0.000000e+00, [[X]]
270+
; CHECK-NEXT: [[FABS:%.*]] = select i1 [[LEZERO]], double [[NEGX]], double [[X]]
271+
; CHECK-NEXT: ret double [[FABS]]
259272
;
260273
%lezero = fcmp ole double %x, 0.0
261274
%negx = fsub double 0.0, %x
@@ -268,7 +281,7 @@ define double @select_fcmp_nnan_ole_zero(double %x) {
268281
; CHECK-NEXT: [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
269282
; CHECK-NEXT: ret double [[FABS]]
270283
;
271-
%lezero = fcmp ole double %x, 0.0
284+
%lezero = fcmp nnan ole double %x, 0.0
272285
%negx = fsub nnan double 0.0, %x
273286
%fabs = select i1 %lezero, double %negx, double %x
274287
ret double %fabs
@@ -279,7 +292,7 @@ define double @select_nnan_fcmp_nnan_ole_zero(double %x) {
279292
; CHECK-NEXT: [[FABS:%.*]] = call nnan double @llvm.fabs.f64(double [[X:%.*]])
280293
; CHECK-NEXT: ret double [[FABS]]
281294
;
282-
%lezero = fcmp ole double %x, 0.0
295+
%lezero = fcmp nnan ole double %x, 0.0
283296
%negx = fsub nnan double 0.0, %x
284297
%fabs = select nnan i1 %lezero, double %negx, double %x
285298
ret double %fabs
@@ -292,7 +305,7 @@ define double @select_fcmp_nnan_ule_zero(double %x) {
292305
; CHECK-NEXT: [[FABS:%.*]] = call double @llvm.fabs.f64(double [[X:%.*]])
293306
; CHECK-NEXT: ret double [[FABS]]
294307
;
295-
%lezero = fcmp ule double %x, 0.0
308+
%lezero = fcmp nnan ule double %x, 0.0
296309
%negx = fsub nnan double 0.0, %x
297310
%fabs = select i1 %lezero, double %negx, double %x
298311
ret double %fabs
@@ -320,7 +333,7 @@ define <2 x float> @select_fcmp_nnan_ole_negzero(<2 x float> %x) {
320333
; CHECK-NEXT: [[FABS:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]])
321334
; CHECK-NEXT: ret <2 x float> [[FABS]]
322335
;
323-
%lezero = fcmp ole <2 x float> %x, <float -0.0, float -0.0>
336+
%lezero = fcmp nnan ole <2 x float> %x, <float -0.0, float -0.0>
324337
%negx = fsub nnan <2 x float> <float 0.0, float poison>, %x
325338
%fabs = select <2 x i1> %lezero, <2 x float> %negx, <2 x float> %x
326339
ret <2 x float> %fabs
@@ -331,7 +344,7 @@ define <2 x float> @select_nnan_fcmp_nnan_ole_negzero(<2 x float> %x) {
331344
; CHECK-NEXT: [[FABS:%.*]] = call nnan <2 x float> @llvm.fabs.v2f32(<2 x float> [[X:%.*]])
332345
; CHECK-NEXT: ret <2 x float> [[FABS]]
333346
;
334-
%lezero = fcmp ole <2 x float> %x, <float -0.0, float -0.0>
347+
%lezero = fcmp nnan ole <2 x float> %x, <float -0.0, float -0.0>
335348
%negx = fsub nnan <2 x float> <float 0.0, float poison>, %x
336349
%fabs = select nnan <2 x i1> %lezero, <2 x float> %negx, <2 x float> %x
337350
ret <2 x float> %fabs
@@ -344,7 +357,7 @@ define fp128 @select_fcmp_ogt_zero(fp128 %x) {
344357
; CHECK-NEXT: [[FABS:%.*]] = call fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
345358
; CHECK-NEXT: ret fp128 [[FABS]]
346359
;
347-
%gtzero = fcmp ogt fp128 %x, zeroinitializer
360+
%gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
348361
%negx = fsub fp128 zeroinitializer, %x
349362
%fabs = select i1 %gtzero, fp128 %x, fp128 %negx
350363
ret fp128 %fabs
@@ -382,7 +395,7 @@ define fp128 @select_fcmp_nnan_ogt_zero(fp128 %x) {
382395
; CHECK-NEXT: [[FABS:%.*]] = call fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
383396
; CHECK-NEXT: ret fp128 [[FABS]]
384397
;
385-
%gtzero = fcmp ogt fp128 %x, zeroinitializer
398+
%gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
386399
%negx = fsub nnan fp128 zeroinitializer, %x
387400
%fabs = select i1 %gtzero, fp128 %x, fp128 %negx
388401
ret fp128 %fabs
@@ -393,7 +406,7 @@ define fp128 @select_nnan_fcmp_nnan_ogt_zero(fp128 %x) {
393406
; CHECK-NEXT: [[FABS:%.*]] = call nnan fp128 @llvm.fabs.f128(fp128 [[X:%.*]])
394407
; CHECK-NEXT: ret fp128 [[FABS]]
395408
;
396-
%gtzero = fcmp ogt fp128 %x, zeroinitializer
409+
%gtzero = fcmp nnan ogt fp128 %x, zeroinitializer
397410
%negx = fsub nnan fp128 zeroinitializer, %x
398411
%fabs = select nnan i1 %gtzero, fp128 %x, fp128 %negx
399412
ret fp128 %fabs
@@ -406,7 +419,7 @@ define half @select_fcmp_nnan_ogt_negzero(half %x) {
406419
; CHECK-NEXT: [[FABS:%.*]] = call half @llvm.fabs.f16(half [[X:%.*]])
407420
; CHECK-NEXT: ret half [[FABS]]
408421
;
409-
%gtzero = fcmp ogt half %x, -0.0
422+
%gtzero = fcmp nnan ogt half %x, -0.0
410423
%negx = fsub nnan half 0.0, %x
411424
%fabs = select i1 %gtzero, half %x, half %negx
412425
ret half %fabs
@@ -417,7 +430,7 @@ define half @select_nnan_fcmp_nnan_ogt_negzero(half %x) {
417430
; CHECK-NEXT: [[FABS:%.*]] = call nnan half @llvm.fabs.f16(half [[X:%.*]])
418431
; CHECK-NEXT: ret half [[FABS]]
419432
;
420-
%gtzero = fcmp ogt half %x, -0.0
433+
%gtzero = fcmp nnan ogt half %x, -0.0
421434
%negx = fsub nnan half 0.0, %x
422435
%fabs = select nnan i1 %gtzero, half %x, half %negx
423436
ret half %fabs
@@ -430,7 +443,7 @@ define half @select_fcmp_nnan_ugt_negzero(half %x) {
430443
; CHECK-NEXT: [[FABS:%.*]] = call half @llvm.fabs.f16(half [[X:%.*]])
431444
; CHECK-NEXT: ret half [[FABS]]
432445
;
433-
%gtzero = fcmp ugt half %x, -0.0
446+
%gtzero = fcmp nnan ugt half %x, -0.0
434447
%negx = fsub nnan half 0.0, %x
435448
%fabs = select i1 %gtzero, half %x, half %negx
436449
ret half %fabs

0 commit comments

Comments
 (0)