|
 |
65ccb9 |
From 1331e4eec016a295949009b4360c592401b089f7 Mon Sep 17 00:00:00 2001
|
|
|
65ccb9 |
From: Richard Henderson <richard.henderson@linaro.org>
|
|
|
65ccb9 |
Date: Sun, 12 Sep 2021 10:49:25 -0700
|
|
|
65ccb9 |
Subject: [PATCH] tcg/arm: Reduce vector alignment requirement for NEON
|
|
|
65ccb9 |
|
|
|
65ccb9 |
With arm32, the ABI gives us 8-byte alignment for the stack.
|
|
|
65ccb9 |
While it's possible to realign the stack to provide 16-byte alignment,
|
|
|
65ccb9 |
it's far easier to simply not encode 16-byte alignment in the
|
|
|
65ccb9 |
VLD1 and VST1 instructions that we emit.
|
|
|
65ccb9 |
|
|
|
65ccb9 |
Remove the assertion in temp_allocate_frame, limit natural alignment
|
|
|
65ccb9 |
to the provided stack alignment, and add a comment.
|
|
|
65ccb9 |
|
|
|
65ccb9 |
Reported-by: Richard W.M. Jones <rjones@redhat.com>
|
|
|
65ccb9 |
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
|
|
|
65ccb9 |
---
|
|
|
65ccb9 |
tcg/arm/tcg-target.c.inc | 13 +++++++++----
|
|
|
65ccb9 |
tcg/tcg.c | 8 +++++++-
|
|
|
65ccb9 |
2 files changed, 16 insertions(+), 5 deletions(-)
|
|
|
65ccb9 |
|
|
|
65ccb9 |
diff --git a/tcg/arm/tcg-target.c.inc b/tcg/arm/tcg-target.c.inc
|
|
|
65ccb9 |
index 007ceee68e..34acfb522c 100644
|
|
|
65ccb9 |
--- a/tcg/arm/tcg-target.c.inc
|
|
|
65ccb9 |
+++ b/tcg/arm/tcg-target.c.inc
|
|
|
65ccb9 |
@@ -2477,8 +2477,13 @@ static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg arg,
|
|
|
65ccb9 |
tcg_out_vldst(s, INSN_VLD1 | 0x7d0, arg, arg1, arg2);
|
|
|
65ccb9 |
return;
|
|
|
65ccb9 |
case TCG_TYPE_V128:
|
|
|
65ccb9 |
- /* regs 2; size 8; align 16 */
|
|
|
65ccb9 |
- tcg_out_vldst(s, INSN_VLD1 | 0xae0, arg, arg1, arg2);
|
|
|
65ccb9 |
+ /*
|
|
|
65ccb9 |
+ * We have only 8-byte alignment for the stack per the ABI.
|
|
|
65ccb9 |
+ * Rather than dynamically re-align the stack, it's easier
|
|
|
65ccb9 |
+ * to simply not request alignment beyond that. So:
|
|
|
65ccb9 |
+ * regs 2; size 8; align 8
|
|
|
65ccb9 |
+ */
|
|
|
65ccb9 |
+ tcg_out_vldst(s, INSN_VLD1 | 0xad0, arg, arg1, arg2);
|
|
|
65ccb9 |
return;
|
|
|
65ccb9 |
default:
|
|
|
65ccb9 |
g_assert_not_reached();
|
|
|
65ccb9 |
@@ -2497,8 +2502,8 @@ static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg,
|
|
|
65ccb9 |
tcg_out_vldst(s, INSN_VST1 | 0x7d0, arg, arg1, arg2);
|
|
|
65ccb9 |
return;
|
|
|
65ccb9 |
case TCG_TYPE_V128:
|
|
|
65ccb9 |
- /* regs 2; size 8; align 16 */
|
|
|
65ccb9 |
- tcg_out_vldst(s, INSN_VST1 | 0xae0, arg, arg1, arg2);
|
|
|
65ccb9 |
+ /* See tcg_out_ld re alignment: regs 2; size 8; align 8 */
|
|
|
65ccb9 |
+ tcg_out_vldst(s, INSN_VST1 | 0xad0, arg, arg1, arg2);
|
|
|
65ccb9 |
return;
|
|
|
65ccb9 |
default:
|
|
|
65ccb9 |
g_assert_not_reached();
|
|
|
65ccb9 |
diff --git a/tcg/tcg.c b/tcg/tcg.c
|
|
|
65ccb9 |
index 4142d42d77..ca5bcc4635 100644
|
|
|
65ccb9 |
--- a/tcg/tcg.c
|
|
|
65ccb9 |
+++ b/tcg/tcg.c
|
|
|
65ccb9 |
@@ -3060,7 +3060,13 @@ static void temp_allocate_frame(TCGContext *s, TCGTemp *ts)
|
|
|
65ccb9 |
g_assert_not_reached();
|
|
|
65ccb9 |
}
|
|
|
65ccb9 |
|
|
|
65ccb9 |
- assert(align <= TCG_TARGET_STACK_ALIGN);
|
|
|
65ccb9 |
+ /*
|
|
|
65ccb9 |
+ * Assume the stack is sufficiently aligned.
|
|
|
65ccb9 |
+ * This affects e.g. ARM NEON, where we have 8 byte stack alignment
|
|
|
65ccb9 |
+ * and do not require 16 byte vector alignment. This seems slightly
|
|
|
65ccb9 |
+ * easier than fully parameterizing the above switch statement.
|
|
|
65ccb9 |
+ */
|
|
|
65ccb9 |
+ align = MIN(TCG_TARGET_STACK_ALIGN, align);
|
|
|
65ccb9 |
off = ROUND_UP(s->current_frame_offset, align);
|
|
|
65ccb9 |
|
|
|
65ccb9 |
/* If we've exhausted the stack frame, restart with a smaller TB. */
|
|
|
65ccb9 |
--
|
|
|
65ccb9 |
2.32.0
|
|
|
65ccb9 |
|