https://gcc.gnu.org/g:2b3402f56531e384e90f5a4228bb3bc0ed8e1250

commit r17-1462-g2b3402f56531e384e90f5a4228bb3bc0ed8e1250
Author: Jeff Law <[email protected]>
Date:   Wed Jun 10 07:35:14 2026 -0600

    [RISC-V] Don't modify existing RTL in combined sCC splitters, generate new 
nodes instead
    
    I was testing spec2017 late last week on the K3 design and stumbled across a
    couple of functional fails.
    
    In this particular case we mis-compiled omnet, thankfully in such a way 
that it
    didn't really run at all, so time to failure was exceedingly short.
    
    A couple months ago I extended our ability to recognize more czero 
sequences,
    particularly in cases where we're combining the output of 2 or more sCC 
style
    insns.  I took a bit of a short-cut and adjusted the code on a node I knew
    would not be shared in the IL when doing a split.  I should have known 
better.
    PUT_MODE and PUT_CODE are generally bad to use and I've called out others 
for
    similar changes.
    
    Consider if we're doing a 3->2 split, but when one of the two split insns do
    not match.  In that scenario the PUT_CODE trick is going to cause problems
    because we change the IL, but the transformation as a whole isn't applied.  
In
    this case we're inverting the mode (EQ->NE and NE->EQ), the net is we flip 
the
    tense of a branch/sCC.
    
    I never narrowed down a testcase.  This was found by inspecting codegen and
    dump differences after bisecting to a change and bisecting down to a single 
.o
    file.
    
    This fixes the omnet failure.  There's a second cluster of failures 
(502.gcc)
    which are unrelated and I'm still debugging those failures.
    
    Bootstrapped and regression tested on riscv64.  Of course the original patch
    survived that test as well, so perhaps take it with a mountain of salt 
rather
    than the usual grain...  Waiting on pre-commit before moving forward.
    
            * config/riscv/zicond.md (combined sCC splitters): Avoid using
            PUT_MODE to change existing RTL.  Instead just generate a new node.

Diff:
---
 gcc/config/riscv/zicond.md | 30 ++++++++++++++++++++++++------
 1 file changed, 24 insertions(+), 6 deletions(-)

diff --git a/gcc/config/riscv/zicond.md b/gcc/config/riscv/zicond.md
index 522b0e75cd86..a890322a5189 100644
--- a/gcc/config/riscv/zicond.md
+++ b/gcc/config/riscv/zicond.md
@@ -143,7 +143,10 @@
                                        [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 5)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 (define_split
   [(set (match_operand:X 0 "register_operand")
@@ -158,7 +161,10 @@
                                        [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 5)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 
 ;; Similarly but GE/GEU which requires (const_int 1) as an operand.
@@ -175,7 +181,10 @@
                                       [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 4)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 (define_split
   [(set (match_operand:X 0 "register_operand")
@@ -190,7 +199,10 @@
                                       [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 4)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 ;; Similarly but LU/LTU which allows an arith_operand
 (define_split
@@ -206,7 +218,10 @@
                                       [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 5)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 ;; Finally LE/LEU which requires sle_operand.
 (define_split
@@ -222,7 +237,10 @@
                                       [(match_dup 2) (const_int 0)])
                                      (const_int 0)
                                      (match_dup 5)))]
-  { PUT_CODE (operands[1], GET_CODE (operands[1]) == EQ ? NE : EQ); })
+  { operands[1] = gen_rtx_fmt_ee (GET_CODE (operands[1]) == EQ ? NE : EQ,
+                                 GET_MODE (operands[1]),
+                                 operands[2],
+                                 CONST0_RTX (GET_MODE (operands[1]))); })
 
 ;; We can splat the sign bit across a GPR with an arithmetic right shift
 ;; which gives us a 0, -1 result.  We then turn on bit #0 unconditionally

Reply via email to