As noticed we are looking for INDIRECT_REF with allows_mem to avoid
a copy since then we're sure the operand is in memory (assuming
*& is folded). But INDIRECT_REFs are no longer a thing, the following
replaces the check with a check for a MEM_REF with a non-ADDR_EXPR
operand. This should fix the regression part without fully
exploring all possibilities around tcc_reference operands.
I've placed an assert that we do not see an INDIRECT_REF here.
While we gimplify asm operands we never do any checking on its
IL afterwards.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
OK if that succeeds?
Thanks,
Richard.
2022-03-09 Richard Biener <rguent...@suse.de>
* cfgexpand.c (expand_gimple_asm): Special-case MEM_REF
with non-decl operand, avoiding a copy.
---
gcc/cfgexpand.cc | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/gcc/cfgexpand.cc b/gcc/cfgexpand.cc
index 4f99f040450..d3cc77d2ca9 100644
--- a/gcc/cfgexpand.cc
+++ b/gcc/cfgexpand.cc
@@ -3290,7 +3290,10 @@ expand_asm_stmt (gasm *stmt)
generating_concat_p = 0;
- if ((TREE_CODE (val) == INDIRECT_REF && allows_mem)
+ gcc_assert (TREE_CODE (val) != INDIRECT_REF);
+ if (((TREE_CODE (val) == MEM_REF
+ && TREE_CODE (TREE_OPERAND (val, 0)) != ADDR_EXPR)
+ && allows_mem)
|| (DECL_P (val)
&& (allows_mem || REG_P (DECL_RTL (val)))
&& ! (REG_P (DECL_RTL (val))
--
2.34.1
