Hi all,
This patch fixes an issue with the final_prescan workaround for the
Cortex-A53 erratum 835769
where calling recog_memoized could modify the recog data for the
multiply-accumulate instruction
when looking at a preceding asm block. This can lead to wrong code
generation.
The solution is to call extract_constrain_insn_cached to restore the
recog data before exiting
aarch64_madd_needs_nop.
Bootstrapped and tested on aarch64-none-linux-gnu.
A compile testcase is added demonstrating the issue.
Ok for trunk?
Thanks,
Kyrill
2014-10-28 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* config/aarch64/aarch64.c (aarch64_madd_needs_nop): Restore
recog state after aarch64_prev_real_insn call.
2014-10-28 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* gcc.target/aarch64/madd_after_asm_1.c: New test.commit e11cd63678d7c334d2f9f124e197695ea470025c
Author: Kyrylo Tkachov <kyrylo.tkac...@arm.com>
Date: Mon Oct 27 13:29:25 2014 +0000
[AArch64] Restore recog state after finding pre-madd instruction
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index cbbc482..002cb44 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -7769,6 +7769,10 @@ aarch64_madd_needs_nop (rtx_insn* insn)
return false;
prev = aarch64_prev_real_insn (insn);
+ /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN.
+ Restore recog state to INSN to avoid state corruption. */
+ extract_constrain_insn_cached (insn);
+
if (!prev || !has_memory_op (prev))
return false;
diff --git a/gcc/testsuite/gcc.target/aarch64/madd_after_asm_1.c b/gcc/testsuite/gcc.target/aarch64/madd_after_asm_1.c
new file mode 100644
index 0000000..523941d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/madd_after_asm_1.c
@@ -0,0 +1,14 @@
+/* { dg-do assemble } */
+/* { dg-options "-O2 -mfix-cortex-a53-835769" } */
+
+int
+test (int a, double b, int c, int d, int e)
+{
+ double result;
+ __asm__ __volatile ("// %0, %1"
+ : "=w" (result)
+ : "0" (b)
+ : /* No clobbers */
+ );
+ return c * d + e;
+}
