================
@@ -307,8 +320,8 @@ def generate_saturated_conversion(src, dst, size):
# Conversion from float to int
print(
""" {DST}{N} y = convert_{DST}{N}(x);
- y = select(y, ({DST}{N}){DST_MIN}, {BP}(x < ({SRC}{N}){DST_MIN}){BS});
- y = select(y, ({DST}{N}){DST_MAX}, {BP}(x > ({SRC}{N}){DST_MAX}){BS});
+ y = select(y, ({DST}{N}){DST_MIN}, {BP}(x <= ({SRC}{N}){DST_MIN}){BS});
+ y = select(y, ({DST}{N}){DST_MAX}, {BP}(x >= ({SRC}{N}){DST_MAX}){BS});
----------------
rjodinchr wrote:
Yes this change was intended.
It is not a standalone bugfix as for all C platforms, this should be the same.
But with Vulkan SPIR-V it leads to issues with zeros, infs & nans, where
comparisons are unspecified (if I remember correctly).
I understand that changing the default path might get a bit more dangerous
there, so I have added a if statement for it as well.
https://github.com/llvm/llvm-project/pull/66902
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
