Package: src:qd Version: 2.3.22+dfsg.1-3 Control: tags -1 patch ftbfs --
Hi, after contacting the ppc64 toolchain team about bug #935289, especially Pat Haugen, he said : --- Basically the testcase is sensitive to precision differences when FMA instructions are used (floating point multiply-add). The compiler will generate these insns when it sees expressions of the form (a * b) + c. If you want to continue to compile the code at -O2 you'll need to specify the compiler option "-ffp-contract=off" in order to prevent the FMA transformation(s). this is just something specific to the values being used in the testcase. The compiler will continue to generate 'fmadd' type instructions at -O2 unless -ffp-contract=off is specified. My digging into the testcase showed that different results were produced based on whether a fmadd insn was generated for a specific expression. If it was not the result was 0, if it was the result was some very small value. That value is then used further and differences compounded, as you could guess. --- so, so we may improve the patch by changing -fno-tree-pre into -ffp-contract=off Thanks, F.
pgpKWm33_tlEt.pgp
Description: PGP signature
