vect: Don't allow vect_emulated_vector_p type in vectorizable_call [PR106322]
authorKewen Lin <linkw@linux.ibm.com>
Tue, 16 Aug 2022 05:18:51 +0000 (00:18 -0500)
committerKewen Lin <linkw@linux.ibm.com>
Tue, 16 Aug 2022 05:19:38 +0000 (00:19 -0500)
commit5239e2bd48fb1e6a1d1b06a1bac49bee0a742e98
tree16683c59b23c4afdd6ccc3b93f421b69779e830a
parenta2ff389099c325bc1244b1d72744a18c5fa6fda3
vect: Don't allow vect_emulated_vector_p type in vectorizable_call [PR106322]

As PR106322 shows, in some cases for some vector type whose
TYPE_MODE is a scalar integral mode instead of a vector mode,
it's possible to obtain wrong target support information when
querying with the scalar integral mode.  For example, for the
test case in PR106322, on ppc64 32bit vectorizer gets vector
type "vector(2) short unsigned int" for scalar type "short
unsigned int", its mode is SImode instead of V2HImode.  The
target support querying checks umul_highpart optab with SImode
and considers it's supported, then vectorizer further generates
.MULH IFN call for that vector type.  Unfortunately it's wrong
to use SImode support for that vector type multiply highpart
here.

This patch is to teach vectorizable_call analysis not to allow
vect_emulated_vector_p type for both vectype_in and vectype_out
as Richi suggested.

PR tree-optimization/106322

gcc/ChangeLog:

* tree-vect-stmts.cc (vectorizable_call): Don't allow
vect_emulated_vector_p type for both vectype_in and vectype_out.

gcc/testsuite/ChangeLog:

* gcc.target/i386/pr106322.c: New test.
* gcc.target/powerpc/pr106322.c: New test.
gcc/testsuite/gcc.target/i386/pr106322.c [new file with mode: 0644]
gcc/testsuite/gcc.target/powerpc/pr106322.c [new file with mode: 0644]
gcc/tree-vect-stmts.cc