[PATCH] D115804: [CodeGen] use saturating FP casts when compiling with "no-strict-float-cast-overflow"

[Sanjay Patel via Phabricator via cfe-commits]

spatel created this revision.
spatel added reviewers: MatzeB, neildhar, nikic, aqjune, dmgreen.
Herald added a subscriber: mcrosier.
spatel requested review of this revision.

We got an unintended consequence of the optimizer getting smarter when compiling in a non-standard mode, and there's no good way to inhibit those optimizations at a later stage. The test is based on an example linked from D92270 <https://reviews.llvm.org/D92270>.

We allow the "no-strict-float-cast-overflow" exception to normal C cast rules to preserve legacy code that does not expect overflowing casts from FP to int to produce UB. See D46236 <https://reviews.llvm.org/D46236> for details.


https://reviews.llvm.org/D115804

Files:
  clang/lib/CodeGen/CGExprScalar.cpp
  clang/test/CodeGen/no-junk-ftrunc.c


Index: clang/test/CodeGen/no-junk-ftrunc.c
===================================================================
--- clang/test/CodeGen/no-junk-ftrunc.c
+++ clang/test/CodeGen/no-junk-ftrunc.c
@@ -1,14 +1,23 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
 // RUN: %clang_cc1 -S -fno-strict-float-cast-overflow %s -emit-llvm -o - | FileCheck %s --check-prefix=NOSTRICT
+
+// When compiling with non-standard semantics, use intrinsics to inhibit the optimizer.
+
 // NOSTRICT-LABEL: main
+// NOSTRICT: call i32 @llvm.fptosi.sat.i32.f64
+// NOSTRICT: call i32 @llvm.fptoui.sat.i32.f64
 // NOSTRICT: attributes #0 = {{.*}}"strict-float-cast-overflow"="false"{{.*}}
 
 // The workaround attribute is not applied by default.
 
 // RUN: %clang_cc1 -S %s -emit-llvm -o - | FileCheck %s --check-prefix=STRICT
 // STRICT-LABEL: main
+// STRICT: = fptosi
+// STRICT: = fptoui
 // STRICT-NOT: strict-float-cast-overflow
 
+
 int main() {
-  return 0;
+  double d = 1e20;
+  return (int)d != 1e20 && (unsigned)d != 1e20;
 }
-
Index: clang/lib/CodeGen/CGExprScalar.cpp
===================================================================
--- clang/lib/CodeGen/CGExprScalar.cpp
+++ clang/lib/CodeGen/CGExprScalar.cpp
@@ -1240,7 +1240,18 @@
 
   if (isa<llvm::IntegerType>(DstElementTy)) {
     assert(SrcElementTy->isFloatingPointTy() && "Unknown real conversion");
-    if (DstElementType->isSignedIntegerOrEnumerationType())
+    bool IsSigned = DstElementType->isSignedIntegerOrEnumerationType();
+
+    // If we can't recognize overflow as undefined behavior, assume that
+    // overflow saturates. This protects against normal optimizations if we are
+    // compiling with non-standard FP semantics.
+    if (!CGF.CGM.getCodeGenOpts().StrictFloatCastOverflow) {
+      llvm::Intrinsic::ID IID =
+          IsSigned ? llvm::Intrinsic::fptosi_sat : llvm::Intrinsic::fptoui_sat;
+      return Builder.CreateCall(CGF.CGM.getIntrinsic(IID, {DstTy, SrcTy}), Src);
+    }
+
+    if (IsSigned)
       return Builder.CreateFPToSI(Src, DstTy, "conv");
     return Builder.CreateFPToUI(Src, DstTy, "conv");
   }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D115804.394571.patch
Type: text/x-patch
Size: 2125 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20211215/2fa4bdf4/attachment-0001.bin>