On Tue, Mar 26, 2019 at 8:05 PM Uros Bizjak <[email protected]> wrote:
>
> Attached patch fixes a corner case in STV pass where the shift operand
> register equals shift count register. The specialization for shift
> insns marked register as processed, but didn't process shift input
> operand, leaving an unprocessed DImode register.
There is another instance of the same problem in make_vector_copies.
2019-03-28 Uroš Bizjak <[email protected]>
PR target/89848
* config/i386/i386.c (dimode_scalar_chain::make_vector_copies):
Also process XEXP (src, 0) of a shift insn.
testsuite/ChangeLog:
2019-03-28 Uroš Bizjak <[email protected]>
PR target/89848
* gcc.target/i386/pr89848.c: New test.
Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.
Committed to mainline SVN, will be backported to gcc-8 branch.
Uros.
Index: testsuite/gcc.target/i386/pr89848.c
===================================================================
--- testsuite/gcc.target/i386/pr89848.c (nonexistent)
+++ testsuite/gcc.target/i386/pr89848.c (revision 270003)
@@ -0,0 +1,11 @@
+/* PR target/89848 */
+/* { dg-do compile { target ia32 } } */
+/* { dg-options "-O2 -msse2 -mtune=pentium3m" } */
+
+long long
+foo (long long x)
+{
+ x >>= 3;
+ x <<= x;
+ return x;
+}
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 270002)
+++ config/i386/i386.c (revision 270003)
@@ -1901,7 +1901,10 @@
|| GET_CODE (src) == LSHIFTRT)
&& !CONST_INT_P (XEXP (src, 1))
&& reg_or_subregno (XEXP (src, 1)) == regno)
- XEXP (src, 1) = vreg;
+ {
+ XEXP (src, 0) = replace_with_subreg (XEXP (src, 0), reg, reg);
+ XEXP (src, 1) = vreg;
+ }
}
else
replace_with_subreg_in_insn (insn, reg, vreg);