x86: Preserve ymm/zmm registers under option #22692

BradleyWood · 2025-09-29T14:02:42Z

This commit restores #14632 and applies a number of changes. Preserving extended registers (ymm/zmm/opmask) registers is only done if the "-XPreserveExtendedRegs" vm option is provided. Two new VM flags are introduced. One flag marks AVX-256 register preservation, the other for AVX-512. Vector lengths 256/512 are preserved in an out-of-line sequence assigned by a function pointer if extended vector preservation is enabled. This is done to prevent speculative execution from causing CPU frequency regressions even when AVX-2 / AVX-512 code is never executed.

gita-omr · 2025-10-04T01:47:50Z

runtime/codert_vm/xvector.m4

+
+ifdef({ASM_J9VM_ENV_DATA64},{
+	pop r8
+	forloop({REG_CTR}, 0, 31, {SAVE_ZMM_REG(REG_CTR, J9TR_cframe_jitFPRs+(REG_CTR*64))})


I think original PR (14632) was only saving registers 0 to 5. We need to save only volatile registers. @0xdaryl @gacholio do you know what the system ABI is for zmm and ymm registers?

On Linux, the System ABI specifies that all XMM, YMM, and ZMM registers are volatile.

On Windows, the x64 ABI specifies that XMM6-XMM15 are preserved. For YMM/ZMM6-15, the upper bits are volatile and the lower bits preserved. XMM/YMM/ZMM 0-5 and 16-31 are all volatile.

gita-omr · 2025-10-04T01:49:39Z

runtime/codert_vm/xvector.m4

+
+ifdef({ASM_J9VM_ENV_DATA64},{
+	pop r8
+	forloop({REG_CTR}, 0, 15, {vmovdqu ymmword ptr J9TR_cframe_jitFPRs+(REG_CTR*32)[_rsp],ymm{}REG_CTR})


Similar to zmm, which exactly registers do we need to save/restore?

keithc-ca · 2025-10-06T14:43:06Z

runtime/codert_vm/xvector.m4

@@ -0,0 +1,126 @@
+dnl Copyright IBM Corp. and others 2023


The commit was authored in 2021, so this should say 2021.

I am not completely sure because xvector.m4 did not exist at the time.

It never got committed before, so I suppose the date should be this year.

Devin worked on this and committed it no later than in 2021. I believe that should be the first copyright year.

runtime/codert_vm/cnathelp.cpp

runtime/codert_vm/xvector.m4

runtime/oti/xhelpers.m4

runtime/vm/jvminit.c

keithc-ca · 2025-10-06T15:56:19Z

runtime/vm/jvminit.c


+#if JAVA_SPEC_VERSION >= 17
+			/* Extended vector register preservation (ymm/zmm/opmask registers) requires a VM option */
+			if ((argIndex = FIND_AND_CONSUME_VMARG(EXACT_MATCH, VMOPT_PRESERVE_VECTORS, NULL)) < 0) {


I think you meant >=, not > here; using -XPreserveExtendedRegs should preserve those registers, but as written will do so only if that option is not specified.

We should have a pair of options:

#define VMOPT_PRESERVE_VECTORS "-XX:+PreserveExtendedRegs" #define VMOPT_NO_PRESERVE_VECTORS "-XX:-PreserveExtendedRegs"

so that the command line can override environment variables.

This should then be:

argIndex = FIND_AND_CONSUME_VMARG(EXACT_MATCH, VMOPT_PRESERVE_VECTORS, NULL); argIndex2 = FIND_AND_CONSUME_VMARG(EXACT_MATCH, VMOPT_NO_PRESERVE_VECTORS, NULL); if (argIndex2 > argIndex) { vm->extendedRuntimeFlags3 &= ~J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_256; vm->extendedRuntimeFlags3 &= ~J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_512; }

I think you meant >=, not > here; using -XPreserveExtendedRegs should preserve those registers, but as written will do so only if that option is not specified.

So extended preservation is supposed to be disabled unless explicitly enabled by an option. This code resets those flags if no option is specified.

We should still have a pair of options (one might appear in an environment variable), the other is needed to override that (on the command line).

This commit restores eclipse-openj9#14632 and applies a number of changes. Preserving extended registers (ymm/zmm/opmask) registers is only done if the "-XPreserveExtendedRegs" vm option is provided. Two new VM flags are introduced. One flag marks AVX-256 register preservation, the other for AVX-512. Vector lengths 256/512 are preserved in an out-of-line sequence assigned by a function pointer if extended vector preservation is enabled. This is done to prevent speculative execution from causing CPU frequency regressions even when AVX-2 / AVX-512 code is never executed. Co-authored-by: Devin Nakamura <[email protected]> Co-authored-by: Bradley Wood <[email protected]>

gacholio

Looks reasonable to me (once all of Keith's comments are addressed). Question: why the fence in the vector code?

keithc-ca · 2025-10-20T14:59:05Z

runtime/codert_vm/cnathelp.cpp

+	if (J9_ARE_ANY_BITS_SET(vm->extendedRuntimeFlags3, J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_512))
+	{


Formatting: braces:

if (J9_ARE_ANY_BITS_SET(vm->extendedRuntimeFlags3, J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_512)) { jitConfig->saveVectorRegisters = (void *)jitSaveVectorRegistersAVX512; jitConfig->restoreVectorRegisters = (void *)jitRestoreVectorRegistersAVX512; } else if (J9_ARE_ANY_BITS_SET(vm->extendedRuntimeFlags3, J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_256)) { jitConfig->saveVectorRegisters = (void *)jitSaveVectorRegistersAVX; jitConfig->restoreVectorRegisters = (void *)jitRestoreVectorRegistersAVX; }

keithc-ca · 2025-10-20T15:01:01Z

runtime/codert_vm/xvector.m4

@@ -0,0 +1,126 @@
+dnl Copyright IBM Corp. and others 2023


Devin worked on this and committed it no later than in 2021. I believe that should be the first copyright year.

keithc-ca · 2025-10-20T15:02:18Z

runtime/oti/jvminit.h

 #define VMOPT_XCONCURRENTBACKGROUND "-Xconcurrentbackground"
 #define VMOPT_XGCTHREADS "-Xgcthreads"
 #define VMOPT_XGCMAXTHREADS "-Xgcmaxthreads"
+#define VMOPT_PRESERVE_VECTORS "-XPreserveExtendedRegs"


I repeat my suggestion that we should have both positive and negative versions of this option:

#define VMOPT_PRESERVE_VECTORS "-XX:+PreserveExtendedRegs" #define VMOPT_NO_PRESERVE_VECTORS "-XX:-PreserveExtendedRegs"

Agreed - the convention should be followed.

keithc-ca · 2025-10-20T15:05:02Z

runtime/oti/xhelpers.m4

 	LABEL(skip_vzu{}VZU_COUNT):
 })

+dnl for(<symbol>=<start>; <symbol> <= <end>; ++<symbol>) { <expr> }


Space both sides of =, please.

keithc-ca · 2025-10-20T15:06:47Z

runtime/oti/xhelpers.m4

+define({forloop},
+	{define({$1}, {$2})$4
+	ifelse({$2}, {$3}, {},{$0({$1}, incr({$2}), {$3}, {$4})})})


My earlier comments have not been addressed; see #22692.

keithc-ca · 2025-10-20T15:09:04Z

runtime/vm/jvminit.c


+#if JAVA_SPEC_VERSION >= 17
+			/* Extended vector register preservation (ymm/zmm/opmask registers) requires a VM option */
+			if ((argIndex = FIND_AND_CONSUME_VMARG(EXACT_MATCH, VMOPT_PRESERVE_VECTORS, NULL)) < 0) {


We should still have a pair of options (one might appear in an environment variable), the other is needed to override that (on the command line).

keithc-ca · 2025-10-20T15:09:24Z

runtime/vm/jvminit.c

+				vm->extendedRuntimeFlags3 &= ~J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_256;
+				vm->extendedRuntimeFlags3 &= ~J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_512;
+			}
+#endif


Missing comment:

#endif /* JAVA_SPEC_VERSION >= 17 */

gita-omr requested review from 0xdaryl, gacholio, gita-omr, keithc-ca and tajila October 4, 2025 00:44

gita-omr reviewed Oct 4, 2025

View reviewed changes

keithc-ca requested changes Oct 6, 2025

View reviewed changes

BradleyWood force-pushed the vapi_preserve_ool_ptr_f branch 2 times, most recently from d1b4280 to 3ced5a4 Compare October 17, 2025 15:06

BradleyWood force-pushed the vapi_preserve_ool_ptr_f branch from 3ced5a4 to 2724324 Compare October 17, 2025 15:09

gacholio approved these changes Oct 20, 2025

View reviewed changes

keithc-ca requested changes Oct 20, 2025

View reviewed changes

		if (J9_ARE_ANY_BITS_SET(vm->extendedRuntimeFlags3, J9_EXTENDED_RUNTIME3_USE_VECTOR_LENGTH_512))
		{

x86: Preserve ymm/zmm registers under option #22692

Are you sure you want to change the base?

x86: Preserve ymm/zmm registers under option #22692

Uh oh!

Conversation

BradleyWood commented Sep 29, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gacholio left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

keithc-ca Oct 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

keithc-ca Oct 20, 2025 •

edited

Loading