A DEBUG_INSN VAR_LOCATION can mention if_info->x without representing an
ordinary use. Exclude debug instructions from the overlap scan so that
debug-only payloads do not suppress min/max if-conversion.
Assisted-by: OpenAI Codex 5.6 Sol, co-reviewed by Claude and Yongqiang Tian.
PR rtl-optimization/127423
gcc/ChangeLog:
* ifcvt.cc (noce_get_alt_condition): Ignore debug instructions when
checking whether if_info->x is mentioned in the interval.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/pr127423.c: New test.
Signed-off-by: Yongqiang Tian <[email protected]>
---
Thanks for pointing this out. I checked the RTL immediately after `ce2`,
rather than relying only on what var-tracking eventually produces.
For the reported case, the transformed sequence first binds `xghost` to the
old value of `if_info->x`, then defines `if_info->x` with `smin`, and only then
binds `result` to the new value. The register is already defined at the first
debug bind, and no ordinary instruction changes it before that program point.
This ordering also follows from the interval involved here: the check scans
`(cond_earliest, jump]`, while the merged definition is emitted at the former
jump position. Ignoring the debug instruction in this membership check does
not move the bind across the new definition.
As a secondary check, var-tracking represents `xghost` using the entry value
when that value is recoverable. In an additional variant where an opaque
read-write asm operand prevents recovery of the old value, var-tracking makes
`xghost` unavailable instead of associating it with the new minimum. Based on
these checks, I do not think an explicit reset is needed in
`noce_get_alt_condition` for this transformation.
I also strengthened the regression test. `-fcompare-debug` alone can pass
vacuously if `ce2` stops performing the conversion in both compilations. The
test now compiles explicitly with `-g`, retains `-fcompare-debug`, and checks
that the resulting assembly contains exactly one `min`. The strengthened
test produces 14 expected passes with the patched compiler. The pristine
compiler reports seven `-fcompare-debug failure (length)` failures, and a
separate control with `ce2` disabled fails the assembly check.
Thank you very much for your help. Any further suggestions would be greatly
appreciated.
Changes in v2:
- Compile the regression test explicitly with `-g`.
- Retain `-fcompare-debug` to check debug/no-debug convergence.
- Require exactly one RISC-V `min` instruction so the test cannot pass without
exercising the intended if-conversion.
gcc/ifcvt.cc | 3 ++-
gcc/testsuite/gcc.target/riscv/pr127423.c | 18 ++++++++++++++++++
2 files changed, 20 insertions(+), 1 deletion(-)
create mode 100644 gcc/testsuite/gcc.target/riscv/pr127423.c
diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc
index 7d23ce115bd..618c2976c5d 100644
--- a/gcc/ifcvt.cc
+++ b/gcc/ifcvt.cc
@@ -3038,7 +3038,8 @@ noce_get_alt_condition (noce_if_info *if_info, rtx target,
/* X may not be mentioned in the range (cond_earliest, jump]. */
for (insn = if_info->jump; insn != *earliest; insn = PREV_INSN (insn))
- if (INSN_P (insn) && reg_overlap_mentioned_p (if_info->x, PATTERN (insn)))
+ if (NONDEBUG_INSN_P (insn)
+ && reg_overlap_mentioned_p (if_info->x, PATTERN (insn)))
return NULL;
/* A and B may not be modified in the range [cond_earliest, jump). */
diff --git a/gcc/testsuite/gcc.target/riscv/pr127423.c
b/gcc/testsuite/gcc.target/riscv/pr127423.c
new file mode 100644
index 00000000000..7c1e25bf640
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/pr127423.c
@@ -0,0 +1,18 @@
+/* PR rtl-optimization/127423 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -g -fno-tree-forwprop -fdisable-tree-phiopt1
-fdisable-tree-phiopt2 -fdisable-tree-phiopt3 -fdisable-tree-phiopt4
-march=rv64gc_zbb -mabi=lp64d -ffunction-sections -fcompare-debug" } */
+/* { dg-final { scan-assembler-times {\mmin\t} 1 } } */
+
+#define ATTR __attribute__ ((noinline, noclone))
+
+ATTR int
+debug_payload_min (int a, int b)
+{
+ _Bool pred = a < b;
+ asm volatile ("" : : "r" (pred));
+ int result = b;
+ int xghost = a;
+ if (pred)
+ result = a;
+ return result;
+}
--
2.34.1
