[PATCH] D142996: DAG: Fix broken lowering of is.fplcass fcZero with DAZ
Matt Arsenault via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 31 10:30:44 PST 2023
arsenm created this revision.
arsenm added reviewers: scanon, jcranmer-intel, sepavloff, kpn, andrew.w.kaylor.
Herald added subscribers: pengfei, hiraditya.
Herald added a project: All.
arsenm requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.
is.fpclass x, fcZero is not equivalent to fcmp with 0 if
denormals are treated as 0. It would be equivalent to fcZero|fcSubnormal
which can be done separately; this is the minimal correctness fix.
The same optimization was not ported to the GlobalISel version.
https://reviews.llvm.org/D142996
Files:
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
llvm/test/CodeGen/X86/is_fpclass.ll
Index: llvm/test/CodeGen/X86/is_fpclass.ll
===================================================================
--- llvm/test/CodeGen/X86/is_fpclass.ll
+++ llvm/test/CodeGen/X86/is_fpclass.ll
@@ -393,24 +393,15 @@
define i1 @iszero_f_daz(float %x) #0 {
; CHECK-32-LABEL: iszero_f_daz:
; CHECK-32: # %bb.0: # %entry
-; CHECK-32-NEXT: flds {{[0-9]+}}(%esp)
-; CHECK-32-NEXT: fldz
-; CHECK-32-NEXT: fucompp
-; CHECK-32-NEXT: fnstsw %ax
-; CHECK-32-NEXT: # kill: def $ah killed $ah killed $ax
-; CHECK-32-NEXT: sahf
-; CHECK-32-NEXT: setnp %cl
+; CHECK-32-NEXT: testl $2147483647, {{[0-9]+}}(%esp) # imm = 0x7FFFFFFF
; CHECK-32-NEXT: sete %al
-; CHECK-32-NEXT: andb %cl, %al
; CHECK-32-NEXT: retl
;
; CHECK-64-LABEL: iszero_f_daz:
; CHECK-64: # %bb.0: # %entry
-; CHECK-64-NEXT: xorps %xmm1, %xmm1
-; CHECK-64-NEXT: cmpeqss %xmm0, %xmm1
-; CHECK-64-NEXT: movd %xmm1, %eax
-; CHECK-64-NEXT: andl $1, %eax
-; CHECK-64-NEXT: # kill: def $al killed $al killed $eax
+; CHECK-64-NEXT: movd %xmm0, %eax
+; CHECK-64-NEXT: testl $2147483647, %eax # imm = 0x7FFFFFFF
+; CHECK-64-NEXT: sete %al
; CHECK-64-NEXT: retq
entry:
%0 = tail call i1 @llvm.is.fpclass.f32(float %x, i32 96) ; 0x60 = "zero"
Index: llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8045,10 +8045,20 @@
// exceptions are ignored.
if (Flags.hasNoFPExcept() &&
isOperationLegalOrCustom(ISD::SETCC, OperandVT.getScalarType())) {
- if (Test == fcZero)
- return DAG.getSetCC(DL, ResultVT, Op,
- DAG.getConstantFP(0.0, DL, OperandVT),
- IsInverted ? ISD::SETUNE : ISD::SETOEQ);
+ if (Test == fcZero) {
+ DenormalMode Mode = DAG.getMachineFunction().getDenormalMode(Semantics);
+ if (Mode.Input == DenormalMode::IEEE) {
+ // If denormals could be implicitly treated as 0, this is not equivalent
+ // to a compare with 0 since it will also be true for denormals.
+ //
+ // TODO: With DAZ, check == fcZero | fcSubnormal
+ return DAG.getSetCC(DL, ResultVT, Op,
+ DAG.getConstantFP(0.0, DL, OperandVT),
+ IsInverted ? ISD::SETUNE : ISD::SETOEQ);
+ }
+
+ }
+
if (Test == fcNan)
return DAG.getSetCC(DL, ResultVT, Op, Op,
IsInverted ? ISD::SETO : ISD::SETUO);
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D142996.493668.patch
Type: text/x-patch
Size: 2584 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230131/6ec657b3/attachment.bin>
More information about the llvm-commits
mailing list