Ping [IA-64] Work around thinko in 'x' constraint implementation
I'd like to ping this patch as it fixed an ICE visible on both ia64 linux and ia64 openvms. Tristan. On Mar 6, 2012, at 11:07 PM, Eric Botcazou wrote: We have a regression on one of the testcases of our internal testsuite on IA-64 with a 4.7-based compiler, which is of the form: test_vec_madd.adb: In function 'Test_Vec_Madd': test_vec_madd.adb:160:5: error: could not split insn (insn 887 4859 889 16 (set (reg:TI 158 f30 [orig:417 m ] [417]) (mem/c:TI (reg/f:DI 14 r14 [1025]) [0 S16 A128])) /gnu/lib/gcc/ia64-hp- openvms/4_7_0/adainclude/g-altcon.adb:277 125 {movti_internal} (nil)) +===GNAT BUG DETECTED==+ | Pro 7.1.0w (20120221-head) (ia64-hp-openvms) GCC error: | | in final_scan_insn, at final.c:2716 | | Error detected around test_vec_madd.adb:160:5| The compiler aborts during the final pass because it couldn't split the insn. The pattern for movti_internal is: (define_insn_and_split movti_internal [(set (match_operand:TI 0 destination_operand =r, *fm,*x,*f, Q) (match_operand:TI 1 general_operand r*fim,r, Q, *fOQ,*f))] ia64_move_ok (operands[0], operands[1]) @ # # ldfp8 %X0 = %1%P1 # # reload_completed !ia64_load_pair_ok(operands[0], operands[1]) [(const_int 0)] The problem is that the operands satisfy ia64_load_pair_ok so the splitter cannot be invoked on them. The root cause is a discrepancy between this predicate and how the 'x' constraint is interpreted. The predicate uses FP_REGNO_P to check the destination and this returns true for %f30 (but would return false for the immediately following register %f31). But recog interprets the 'x' constraint as meaning that every hard register in the destination must be in the FP_REGS class; now the mode is TImode so both %f30 and %f31 are taken into account and %f31 isn't in the FP_REGS class, so the operand is rejected. AFAICS the problem dates back to the introduction of the code (r102463), so I'm not sure that we want to rewrite it at this point. That's why the attached patch is a simple workaround that just avoid ICEing. Bootstrapped/regtested on IA-64/Linux, OK for the mainline? Do we also want it for 4.7.1 (I assume that some RA change makes the issue visible in 4.7.x)? 2012-03-06 Eric Botcazou ebotca...@adacore.com * config/ia64/ia64.c (ia64_load_pair_ok): Return 0 if the second member of the destination isn't also a FP_REGS register. -- Eric Botcazou p1.diff
Re: Ping [IA-64] Work around thinko in 'x' constraint implementation
Ok. r~ On 04/04/2012 03:34 AM, Tristan Gingold wrote: I'd like to ping this patch as it fixed an ICE visible on both ia64 linux and ia64 openvms. Tristan. On Mar 6, 2012, at 11:07 PM, Eric Botcazou wrote: We have a regression on one of the testcases of our internal testsuite on IA-64 with a 4.7-based compiler, which is of the form: test_vec_madd.adb: In function 'Test_Vec_Madd': test_vec_madd.adb:160:5: error: could not split insn (insn 887 4859 889 16 (set (reg:TI 158 f30 [orig:417 m ] [417]) (mem/c:TI (reg/f:DI 14 r14 [1025]) [0 S16 A128])) /gnu/lib/gcc/ia64-hp- openvms/4_7_0/adainclude/g-altcon.adb:277 125 {movti_internal} (nil)) +===GNAT BUG DETECTED==+ | Pro 7.1.0w (20120221-head) (ia64-hp-openvms) GCC error: | | in final_scan_insn, at final.c:2716 | | Error detected around test_vec_madd.adb:160:5| The compiler aborts during the final pass because it couldn't split the insn. The pattern for movti_internal is: (define_insn_and_split movti_internal [(set (match_operand:TI 0 destination_operand =r, *fm,*x,*f, Q) (match_operand:TI 1 general_operand r*fim,r, Q, *fOQ,*f))] ia64_move_ok (operands[0], operands[1]) @ # # ldfp8 %X0 = %1%P1 # # reload_completed !ia64_load_pair_ok(operands[0], operands[1]) [(const_int 0)] The problem is that the operands satisfy ia64_load_pair_ok so the splitter cannot be invoked on them. The root cause is a discrepancy between this predicate and how the 'x' constraint is interpreted. The predicate uses FP_REGNO_P to check the destination and this returns true for %f30 (but would return false for the immediately following register %f31). But recog interprets the 'x' constraint as meaning that every hard register in the destination must be in the FP_REGS class; now the mode is TImode so both %f30 and %f31 are taken into account and %f31 isn't in the FP_REGS class, so the operand is rejected. AFAICS the problem dates back to the introduction of the code (r102463), so I'm not sure that we want to rewrite it at this point. That's why the attached patch is a simple workaround that just avoid ICEing. Bootstrapped/regtested on IA-64/Linux, OK for the mainline? Do we also want it for 4.7.1 (I assume that some RA change makes the issue visible in 4.7.x)? 2012-03-06 Eric Botcazou ebotca...@adacore.com * config/ia64/ia64.c (ia64_load_pair_ok): Return 0 if the second member of the destination isn't also a FP_REGS register. -- Eric Botcazou p1.diff