Hi! My assumption that SIGN/ZERO_EXTEND will always have MODE_INT class mode was provably wrong, as can be seen on the attached testcase where mode is integer vector. Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, committed as obvious.
2011-06-29 Jakub Jelinek <ja...@redhat.com> PR debug/49567 * dwarf2out.c (mem_loc_descriptor) <case ZERO_EXTEND>: Give up for non-MODE_INT modes instead of asserting the mode has MODE_INT class. * gcc.target/i386/pr49567.c: New test. --- gcc/dwarf2out.c.jj 2011-06-23 10:13:58.000000000 +0200 +++ gcc/dwarf2out.c 2011-06-28 16:12:08.000000000 +0200 @@ -14762,7 +14762,8 @@ mem_loc_descriptor (rtx rtl, enum machin case SIGN_EXTEND: case ZERO_EXTEND: - gcc_assert (GET_MODE_CLASS (mode) == MODE_INT); + if (GET_MODE_CLASS (mode) != MODE_INT) + break; op0 = mem_loc_descriptor (XEXP (rtl, 0), GET_MODE (XEXP (rtl, 0)), mem_mode, VAR_INIT_STATUS_INITIALIZED); if (op0 == 0) --- gcc/testsuite/gcc.target/i386/pr49567.c.jj 2011-06-28 19:02:11.000000000 +0200 +++ gcc/testsuite/gcc.target/i386/pr49567.c 2011-06-28 19:01:41.000000000 +0200 @@ -0,0 +1,13 @@ +/* PR debug/49567 */ +/* { dg-do compile } */ +/* { dg-options "-g -O2 -msse4" } */ + +#include <x86intrin.h> + +__m128 +foo (__m128i x) +{ + __m128i y; + y = _mm_cvtepi16_epi32 (x); + return _mm_cvtepi32_ps (y); +} Jakub