release/20.x: [ValueTracking] Bail out on x86_fp80 when computing fpclass with knownbits (#130477)

[llvmbot]

Backport 029e102

Requested by: @dtcxzyw

[llvmbot]

@arsenm What do you think about merging this PR to the release branch?

[llvmbot]

@llvm/pr-subscribers-llvm-analysis

@llvm/pr-subscribers-llvm-transforms

Author: None (llvmbot)

Changes

Backport 029e102

Requested by: @dtcxzyw

---

Full diff: https://github.com/llvm/llvm-project/pull/130489.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/ValueTracking.cpp (+2-1)
• (modified) llvm/test/Transforms/InstSimplify/fcmp.ll (+17)

diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 8a674914641a8..4e79f41df2eb9 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -6135,13 +6135,14 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
     else if (Bits.isNegative())
       Known.signBitMustBeOne();
 
-    if (Ty->isIEEE()) {
+    if (Ty->isIEEELikeFPTy()) {
       // IEEE floats are NaN when all bits of the exponent plus at least one of
       // the fraction bits are 1. This means:
       //   - If we assume unknown bits are 0 and the value is NaN, it will
       //     always be NaN
       //   - If we assume unknown bits are 1 and the value is not NaN, it can
       //     never be NaN
+      // Note: They do not hold for x86_fp80 format.
       if (APFloat(Ty->getFltSemantics(), Bits.One).isNaN())
         Known.KnownFPClasses = fcNan;
       else if (!APFloat(Ty->getFltSemantics(), ~Bits.Zero).isNaN())
diff --git a/llvm/test/Transforms/InstSimplify/fcmp.ll b/llvm/test/Transforms/InstSimplify/fcmp.ll
index 64132f5fb7db7..0c2be5210a741 100644
--- a/llvm/test/Transforms/InstSimplify/fcmp.ll
+++ b/llvm/test/Transforms/InstSimplify/fcmp.ll
@@ -16,3 +16,20 @@ define i1 @poison2(float %x) {
   %v = fcmp ueq float %x, poison
   ret i1 %v
 }
+
+define i1 @pr130408(x86_fp80 %x) {
+; CHECK-LABEL: @pr130408(
+; CHECK-NEXT:    [[BITS:%.*]] = bitcast x86_fp80 [[X:%.*]] to i80
+; CHECK-NEXT:    [[MASKED:%.*]] = and i80 [[BITS]], -604444463063240877801473
+; CHECK-NEXT:    [[OR:%.*]] = or i80 [[MASKED]], 302194561415509874573312
+; CHECK-NEXT:    [[FP:%.*]] = bitcast i80 [[OR]] to x86_fp80
+; CHECK-NEXT:    [[RES:%.*]] = fcmp uno x86_fp80 [[FP]], 0xK00000000000000000000
+; CHECK-NEXT:    ret i1 [[RES]]
+;
+  %bits = bitcast x86_fp80 %x to i80
+  %masked = and i80 %bits, -604444463063240877801473
+  %or = or i80 %masked, 302194561415509874573312
+  %fp = bitcast i80 %or to x86_fp80
+  %res = fcmp uno x86_fp80 %fp, 0xK00000000000000000000
+  ret i1 %res
+}

[github-actions]

@dtcxzyw (or anyone else). If you would like to add a note about this fix in the release notes (completely optional). Please reply to this comment with a one or two sentence description of the fix. When you are done, please add the release:note label to this PR.