| Message ID | 563C86EA.3070000@t-online.de |
|---|---|
| State | New |
| Headers | show |
On 11/06/2015 03:54 AM, Bernd Schmidt wrote: > I have a patch which makes use of the renamer more often, and this > exposed a bug with earlyclobber operands. The code that does the > terminate_write step has the following comment: > > /* Step 5: Close open chains that overlap writes. Similar to > step 2, we hide in-out operands, since we do not want to > close these chains. We also hide earlyclobber operands, > since we've opened chains for them in step 1, and earlier > chains they would overlap with must have been closed at > the previous insn at the latest, as such operands cannot > possibly overlap with any input operands. */ > > That's all right as far as it goes, but the problem is that this means > there isn't a terminate_write step for earlyclobbers. > > The following seems like the simplest possible fix. It was bootstrapped > and tested with -frename-registers enabled at -O1 on x86_64-linux. Ok? > > (Incidentally there are some avx tests that fail if they are renamed, > apparently because the scan-assembler doesn't allow register numbers > like %zmm10. avx512bw-vptestmb-1.c is one of those). > > > Bernd > > rr-ec-term.diff > > > * regrename.c (record_out_operands): Terminate earlyclobbered > operands here. Do you want to terminate those chains after step 6 is complete so that the earlyclobber output conflicts with all the other outputs? Isn't that how we typically do things for earlyclobbers in life analysis? earlyclobbers birth before the insn so they conflict with the inputs and die after the insn so they conflict with the outputs? Or am I fundamentally wrong in even thinking about this in a manner similar to lifetime analysis? Jeff
On 11/06/2015 08:30 PM, Jeff Law wrote: >> * regrename.c (record_out_operands): Terminate earlyclobbered >> operands here. > Do you want to terminate those chains after step 6 is complete so that > the earlyclobber output conflicts with all the other outputs? I'm not sure what you mean by your question. This patch is about terminating live registers that are overwritten by earlyclobber output operands. We open a new chain for every output, ec or not, and since they are live at the same time, they are all considered conflicting. Bernd
On 11/06/2015 04:04 PM, Bernd Schmidt wrote: > On 11/06/2015 08:30 PM, Jeff Law wrote: >>> * regrename.c (record_out_operands): Terminate earlyclobbered >>> operands here. >> Do you want to terminate those chains after step 6 is complete so that >> the earlyclobber output conflicts with all the other outputs? > > I'm not sure what you mean by your question. This patch is about > terminating live registers that are overwritten by earlyclobber output > operands. We open a new chain for every output, ec or not, and since > they are live at the same time, they are all considered conflicting. I must have completely mis-understood the code. OK for the trunk. jeff
* regrename.c (record_out_operands): Terminate earlyclobbered operands here. Index: gcc/regrename.c =================================================================== --- gcc/regrename.c (revision 229049) +++ gcc/regrename.c (working copy) @@ -1513,6 +1525,8 @@ record_out_operands (rtx_insn *insn, boo cur_operand = insn_info->op_info + i; prev_open = open_chains; + if (earlyclobber) + scan_rtx (insn, loc, cl, terminate_write, OP_OUT); scan_rtx (insn, loc, cl, mark_write, OP_OUT); /* ??? Many targets have output constraints on the SET_DEST
I have a patch which makes use of the renamer more often, and this exposed a bug with earlyclobber operands. The code that does the terminate_write step has the following comment: /* Step 5: Close open chains that overlap writes. Similar to step 2, we hide in-out operands, since we do not want to close these chains. We also hide earlyclobber operands, since we've opened chains for them in step 1, and earlier chains they would overlap with must have been closed at the previous insn at the latest, as such operands cannot possibly overlap with any input operands. */ That's all right as far as it goes, but the problem is that this means there isn't a terminate_write step for earlyclobbers. The following seems like the simplest possible fix. It was bootstrapped and tested with -frename-registers enabled at -O1 on x86_64-linux. Ok? (Incidentally there are some avx tests that fail if they are renamed, apparently because the scan-assembler doesn't allow register numbers like %zmm10. avx512bw-vptestmb-1.c is one of those). Bernd