Make uitofp and sitofp defined on overflow.

[efriedma]

Summary

IEEE 754 defines the expected result on overflow. As far as I know, hardware implementations (of f16), and compiler-rt (__floatuntisf) correctly return +-Inf on overflow. And I can't think of any useful transform that would take advantage of overflow being undefined here.

[spatel]

This is reversing part of D5603 / rL219542 . I see there was some controversy in the post-commit thread, although it was for the FP -> int direction. Let me add those commenters as reviewers in any case.

For reference, IEEE754, section 7.4 says:
"The overflow exception shall be signaled if and only if the destination format’s largest finite number is exceeded in magnitude by what would have been the rounded floating-point result (see 4) were the exponent range unbounded. The default result shall be determined by the rounding-direction attribute and the sign of the intermediate result as follows:
a) roundTiesToEven and roundTiesToAway carry all overflows to ∞ with the sign of the intermediate result."

And the C standard, section 6.3.1.4 says:
"When a value of integer type is converted to a real floating type, if the value being converted can be represented exactly in the new type, it is unchanged. If the value being converted is in the range of values that can be represented but cannot be represented exactly, the result is either the nearest higher or nearest lower representable value, chosen in an implementation-defined manner. If the value being converted is outside the range of values that can be represented, the behavior is undefined."

That language (like the LangRef regardless of what we decide here) probably needs refinement because there is no "outside of the range" of infinity?

Note: I moved/corrected the tests at rL334107 (shouldn't have been under instcombine in the first place). Please rebase with full context for the record.

[spatel]

Adding some more possible reviewers (people that commented on the recent junk-in-the-ftrunc changes): D46135.

[efriedma]

The C standard language is weird because it was written before IEEE fp was used universally; some formats didn't have infinity.

[efriedma]

Rebased, and uploading with full context.

[spatel]

I agree that this is a good change, so LGTM. But we should wait a bit for more comments in case anyone else sees potential downside.

docs/LangRef.rst

3289

Nit - I'd find it easier to read if we rewrote as:

``uitofp (CONSTANT to TYPE)``
    Convert an unsigned integer constant to floating-point. 
    CONSTANT must be a scalar or vector integer type. 
    TYPE must be a scalar or vector floating-point type. 
    If CONSTANT is a vector, TYPE must be a vector with the same number of elements.

[nikic]

Rust currently generates additional code for u128 to f32 casts to work around this problem (see https://github.com/rust-lang/rust/blob/c131bdcaff68d35f96e954baac4340206779335f/src/librustc_codegen_llvm/mir/rvalue.rs#L795). Making this defined would allow directly using uitofp, so this would be a welcome change.

docs/LangRef.rst

     Convert an unsigned integer constant to the corresponding 
     floating-point constant. TYPE must be a scalar or vector floating-point
     type.  CST must be of scalar or vector integer type. Both CST and TYPE must
-    be scalars, or vectors of the same number of elements. If the value
-    won't fit in the floating-point type, the results are undefined.
+    be scalars, or vectors of the same number of elements.
 ``sitofp (CST to TYPE)``
     Convert a signed integer constant to the corresponding floating-point
     constant. TYPE must be a scalar or vector floating-point type.
     CST must be of scalar or vector integer type. Both CST and TYPE must
-    be scalars, or vectors of the same number of elements. If the value
-    won't fit in the floating-point type, the results are undefined.
+    be scalars, or vectors of the same number of elements.
 ``ptrtoint (CST to TYPE)``
     Perform the :ref:`ptrtoint operation <i_ptrtoint>` on constants.
 ``inttoptr (CST to TYPE)``

 
 The '``uitofp``' instruction interprets its operand as an unsigned
 integer quantity and converts it to the corresponding floating-point
-value. If the value cannot fit in the floating-point value, the results
-are undefined.
+value. If the value cannot be exactly represented, it is rounded using
+the default rounding mode.
+
 
 Example:
 """"""""

 """"""""""
 
 The '``sitofp``' instruction interprets its operand as a signed integer
-quantity and converts it to the corresponding floating-point value. If
-the value cannot fit in the floating-point value, the results are
-undefined.
+quantity and converts it to the corresponding floating-point value. If the
+value cannot be exactly represented, it is rounded using the default rounding
+mode.
 
 Example:
 """"""""

lib/IR/ConstantFold.cpp

       const APInt &api = CI->getValue();
       APFloat apf(DestTy->getFltSemantics(),
                   APInt::getNullValue(DestTy->getPrimitiveSizeInBits()));
-      if (APFloat::opOverflow &
-          apf.convertFromAPInt(api, opc==Instruction::SIToFP,
-                              APFloat::rmNearestTiesToEven)) {
-        // Undefined behavior invoked - the destination type can't represent
-        // the input constant.
-        return UndefValue::get(DestTy);
-      }
+      apf.convertFromAPInt(api, opc==Instruction::SIToFP,
+                           APFloat::rmNearestTiesToEven);
       return ConstantFP::get(V->getContext(), apf);
     }
     return nullptr;

test/Transforms/InstCombine/cast.ll

 ; number; 129-bits for the value + 1 bit for the sign.
 define float @overflow_uitofp() {
 ; CHECK-LABEL: @overflow_uitofp(
-; CHECK-NEXT:    ret float undef
+; CHECK-NEXT:    ret float 0x7FF0000000000000
 ;
   %i = uitofp i130 400000000000000000000000000000000000000 to float
   ret float %i

 
 define float @overflow_sitofp() {
 ; CHECK-LABEL: @overflow_sitofp(
-; CHECK-NEXT:    ret float undef
+; CHECK-NEXT:    ret float 0x7FF0000000000000
 ;
   %i = sitofp i130 400000000000000000000000000000000000000 to float
   ret float %i