Unified build on sh-unknown-elf with enabling c++ ICEs during compiling libstdc++-v3/src/ios_init.cc with -m2a. It starts to fail my recent change r158208 | kkojima | 2010-04-12 07:59:36 +0900 (Mon, 12 Apr 2010) | 7 lines * config/sh/sh-protos.h (sh_legitimize_reload_address): Declare. * config/sh/sh.c: Include reload.h. (sh_legitimize_reload_address): New. * config/sh/sh.h (LEGITIMIZE_RELOAD_ADDRESS): Use sh_legitimize_reload_address.
Created attachment 20374 [details] a reduced test case xxx.ii: In constructor 'std::ios_base::Init::Init()': xxx.ii:161:3: error: insn does not satisfy its constraints: (insn 284 282 285 6 xxx.ii:113 (set (reg:SI 1 r1) (reg/f:SI 319)) 176 {movsi_ie} (nil)) xxx.ii:161:3: internal compiler error: in reload_cse_simplify_operands, at postreload.c:396 which shows that the pseudo register remains after reload. The .ira rtl dump looks that sh_legitimize_reload_address makes a too complex reload for this case. I've confirmed the patch below fix the above failure. I'll apply it if it survives the usual tests and doesn't make the code quality worse with CSiBE tests. --- ORIG/trunk/gcc/config/sh/sh.c 2010-04-12 09:52:36.000000000 +0900 +++ trunk/gcc/config/sh/sh.c 2010-04-13 08:55:31.000000000 +0900 @@ -9649,8 +9649,7 @@ sh_legitimize_reload_address (rtx *p, en || XEXP (*p, 0) == hard_frame_pointer_rtx)) { rtx index_rtx = XEXP (*p, 1); - HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; - rtx sum; + HOST_WIDE_INT offset = INTVAL (index_rtx); if (TARGET_SH2A && mode == DFmode && (offset & 0x7)) { @@ -9665,24 +9664,6 @@ sh_legitimize_reload_address (rtx *p, en BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); goto win; } - /* Instead of offset_base 128..131 use 124..127, so that - simple add suffices. */ - if (offset > 127) - offset_base = ((offset + 4) & ~60) - 4; - else - offset_base = offset & ~60; - /* Sometimes the normal form does not suit DImode. We could avoid - that by using smaller ranges, but that would give less optimized - code when SImode is prevalent. */ - if (offset_base != 0 - && GET_MODE_SIZE (mode) + offset - offset_base <= 64) - { - sum = gen_rtx_PLUS (Pmode, XEXP (*p, 0), GEN_INT (offset_base)); - *p = gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); - push_reload (sum, NULL_RTX, &XEXP (*p, 0), NULL, - BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); - goto win; - } } /* We must re-recognize what we created before. */ else if (GET_CODE (*p) == PLUS
I've noticed that I applied a wrong patch at revision 158208 and it causes the problematic reload in the failed case. I'll fix it.
Fixed.