DAG: Handle AssertNoFPClass in computeKnownBits

[arsenm]

It's possible to determine the sign bit if the value is known
one of the positive/negative classes and not-nan.

[arsenm]

• DAG: Handle AssertNoFPClass in computeKnownBits #167289 👈 (View in Graphite)
• AMDGPU: Add baseline test for known bits of AssertNoFPClass #167288
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-selectiondag

Author: Matt Arsenault (arsenm)

Changes

It's possible to determine the sign bit if the value is known
one of the positive/negative classes and not-nan.

---

Full diff: https://github.com/llvm/llvm-project/pull/167289.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+21)
• (modified) llvm/test/CodeGen/AMDGPU/compute-known-bits-nofpclass.ll (-2)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 80bbfea7fb83c..27dcd8a546d91 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -4121,6 +4121,27 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
     Known.One.clearLowBits(LogOfAlign);
     break;
   }
+  case ISD::AssertNoFPClass: {
+    Known = computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
+
+    FPClassTest NoFPClass =
+        static_cast<FPClassTest>(Op.getConstantOperandVal(1));
+    const FPClassTest NegativeTestMask = fcNan | fcNegative;
+    if ((NoFPClass & NegativeTestMask) == NegativeTestMask) {
+      // Cannot be negative.
+      Known.Zero.setSignBit();
+      Known.One.clearSignBit();
+    }
+
+    const FPClassTest PositiveTestMask = fcNan | fcPositive;
+    if ((NoFPClass & PositiveTestMask) == PositiveTestMask) {
+      // Cannot be positive.
+      Known.Zero.clearSignBit();
+      Known.One.setSignBit();
+    }
+
+    break;
+  }
   case ISD::FGETSIGN:
     // All bits are zero except the low bit.
     Known.Zero.setBitsFrom(1);
diff --git a/llvm/test/CodeGen/AMDGPU/compute-known-bits-nofpclass.ll b/llvm/test/CodeGen/AMDGPU/compute-known-bits-nofpclass.ll
index d440d58246333..244c3f7c2a96a 100644
--- a/llvm/test/CodeGen/AMDGPU/compute-known-bits-nofpclass.ll
+++ b/llvm/test/CodeGen/AMDGPU/compute-known-bits-nofpclass.ll
@@ -5,7 +5,6 @@ define i32 @known_positive(float nofpclass(nan ninf nzero nsub nnorm) %signbit.z
 ; CHECK-LABEL: known_positive:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
   %cast = bitcast float %signbit.zero to i32
   %and = and i32 %cast, 2147483647
@@ -27,7 +26,6 @@ define i32 @known_negative(float nofpclass(nan pinf pzero psub pnorm) %signbit.o
 ; CHECK-LABEL: known_negative:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    v_or_b32_e32 v0, 0x80000000, v0
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
   %cast = bitcast float %signbit.one to i32
   %or = or i32 %cast, -2147483648

[dtcxzyw]

Does SDAG's version of computeKnownBits also work for floating-point values?

[RKSimon]

Mainly just through bitcasts to wrapped integer logic etc. But X86 does successfully use it for some of its FP node types