From: Jeff Law <jlaw@ventanamicro.com>
Date: Thu, 2 Oct 2025 12:25:47 +0000 (-0600)
Subject: [RISC-V][PR target/122051] Fix pmode_reg_or_uimm5_operand for thead vector
X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=c34ccc8ba23aa7cbf5e2d88186abf0c384a675a3;p=thirdparty%2Fgcc.git

[RISC-V][PR target/122051] Fix pmode_reg_or_uimm5_operand for thead vector

For this bug we're failing during vsetvl insertion, but the real problem
happens earlier.

Basically the slide instructions are using pmode_reg_or_uimm5_operand which has
an implementation that was appropriate when we integrated RVV, but which is
bogus once thead vector was added.  It was just a thin wrapper around
vector_length_operand.

vector_length_operand rejects most constants when thead vector is enabled.  LRA
saw the rK constraint, so it figured the (const_int 1) was a sensible
substitution for the relevant operand.  It was only during vsetvl insertion
that we made another change to the insn and tried to recognize it and boom
things blew up.

This patch makes pmode_reg_or_uimm5_operand independent of
vector_length_operand and everything is happy again.

Tested on riscv32-elf (verifying the selector properly skips the test) and
riscv64-elf where the ICE could be seen.  Bootstrap on the Pioneer and BPI just
started a short while ago, so no data for another 7/24 hours respectively, but
not expecting issues.

	PR target/122051
gcc/
	* config/riscv/predicates.md (pmode_reg_or_uimm5_operand): Implement
	directly rather than using vector_length_operand.

gcc/testsuite/
	* gcc.target/riscv/pr122051.c: New test.
---

diff --git a/gcc/config/riscv/predicates.md b/gcc/config/riscv/predicates.md
index 777e71b14e3..056f9e21909 100644
--- a/gcc/config/riscv/predicates.md
+++ b/gcc/config/riscv/predicates.md
@@ -607,13 +607,12 @@
 (define_predicate "ge_operator"
   (match_code "ge,geu"))
 
-;; pmode_reg_or_uimm5_operand can be used by vsll.vx/vsrl.vx/vsra.vx instructions.
-;; Since it has the same predicate with vector_length_operand which allows register
-;; or immediate (0 ~ 31), we define this predicate same as vector_length_operand here.
-;; We don't use vector_length_operand directly to predicate vsll.vx/vsrl.vx/vsra.vx
-;; since it may be confusing.
+;; pmode_reg_or_uimm5_operand can be used by vsll.vx/vsrl.vx/vsra.vx instructions
+;; It is *not* equivalent to vector_length_operand due to the vector_length_operand
+;; needing to conditionalize some behavior on XTHEADVECTOR.
 (define_special_predicate "pmode_reg_or_uimm5_operand"
-  (match_operand 0 "vector_length_operand"))
+  (ior (match_operand 0 "pmode_register_operand")
+       (match_operand 0 "const_csr_operand")))
 
 (define_special_predicate "pmode_reg_or_0_operand"
   (ior (match_operand 0 "const_0_operand")
diff --git a/gcc/testsuite/gcc.target/riscv/pr122051.c b/gcc/testsuite/gcc.target/riscv/pr122051.c
new file mode 100644
index 00000000000..c2f4b877d5e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/pr122051.c
@@ -0,0 +1,24 @@
+/* { dg-do compile { target int128 } } */
+/* { dg-additional-options "-mrvv-vector-bits=zvl -mcpu=xt-c920 -w" } */
+
+typedef __attribute__((__vector_size__(4))) char B;
+typedef __attribute__((__vector_size__(16))) long V;
+typedef __attribute__((__vector_size__(32))) double W;
+typedef __attribute__((__vector_size__(32))) char U;
+unsigned u;
+B o;
+char *p;
+int q;
+V v;
+W w;
+
+void
+foo(__int128, __int128, __int128, __int128, B a, B b, B c, B d, B e, B f, B g, B h) {
+  do {
+    w -= q;
+    v ^= u;
+  } while (__builtin_memcmp(p, 1 + p, 7));
+  o = ((U)w)[0] + c + d + e + f + g + h + a + b;
+}
+
+