[PATCH] PR middle-end/80270: ICE in extract_bit_field_1

[PATCH] PR middle-end/80270: ICE in extract_bit_field_1

[Roger Sayle]

[-- Attachment #1: Type: text/plain, Size: 1067 bytes --]


This patch fixes PR middle-end/80270, an ICE-on-valid regression, where
performing a bitfield extraction on a variable explicitly stored in a
hard register by the user causes a segmentation fault during RTL
expansion.  Nearly identical source code without the "asm" qualifier
compiles fine.  The point of divergence is in simplify_gen_subreg
which tries to avoid creating non-trivial SUBREGs of hard registers,
to avoid problems during register allocation.  This suggests the
simple solution proposed here, to copy hard registers to a new pseudo
in extract_integral_bit_field, just before calling simplify_gen_subreg.

This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures.  Ok for mainline?


2022-02-27  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
	PR middle-end/80270
	* expmed.cc (extract_integral_bit_field): If OP0 is a hard
	register, copy it to a pseudo before calling simplify_gen_subreg.

gcc/testsuite/ChangeLog
	* gcc.target/i386/pr80270.c: New test case.

Thanks in advance,
Roger
--


[-- Attachment #2: patchbfb.txt --]
[-- Type: text/plain, Size: 1293 bytes --]

diff --git a/gcc/expmed.cc b/gcc/expmed.cc
index 80a16ce..b51450d 100644
--- a/gcc/expmed.cc
+++ b/gcc/expmed.cc
@@ -1975,6 +1975,14 @@ extract_integral_bit_field (rtx op0, opt_scalar_int_mode op0_mode,
 					    unsignedp, reverse);
 	  return convert_extracted_bit_field (target, mode, tmode, unsignedp);
 	}
+      /* If OP0 is a hard register, copy it to a pseudo before calling
+	 simplify_gen_subreg.  */
+      if (REG_P (op0) && HARD_REGISTER_P (op0))
+	{
+	  rtx tmp = gen_reg_rtx (GET_MODE (op0));
+	  emit_move_insn (tmp, op0);
+	  op0 = tmp;
+	}
       op0 = simplify_gen_subreg (word_mode, op0, op0_mode.require (),
 				 bitnum / BITS_PER_WORD * UNITS_PER_WORD);
       op0_mode = word_mode;
diff --git a/gcc/testsuite/gcc.target/i386/pr80270.c b/gcc/testsuite/gcc.target/i386/pr80270.c
new file mode 100644
index 0000000..89e9c33
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr80270.c
@@ -0,0 +1,21 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -msse" } */
+
+typedef int v8 __attribute__((vector_size(8)));
+struct S1 {
+  v8 s1f;
+};
+struct S2 {
+  struct S1 s2f1;
+  v8 s2f2;
+};
+
+extern void foo(int);
+
+void bar() {
+  int tmp, i = 3;
+  register struct S2 b asm("xmm0");
+  tmp = b.s2f1.s1f[i];
+  foo(tmp);
+}
+

Re: [PATCH] PR middle-end/80270: ICE in extract_bit_field_1

[Eric Botcazou]

> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check with no new failures.  Ok for mainline?
> 
> 
> 2022-02-27  Roger Sayle  <roger@nextmovesoftware.com>
> 
> gcc/ChangeLog
> 	PR middle-end/80270
> 	* expmed.cc (extract_integral_bit_field): If OP0 is a hard
> 	register, copy it to a pseudo before calling simplify_gen_subreg.

Looks good to me, but why not using copy_to_reg here?

-- 
Eric Botcazou