| Message ID | 1398953269.20704.8.camel@localhost.localdomain |
|---|---|
| State | New |
| Headers | show |
On 05/01/2014 03:07 PM, Edward Nevill wrote:
> OK to push?
Yes, thanks.
Andrew.
diff -r 9d641fdeea4d -r f67f9b1b52ae src/cpu/aarch64/vm/nativeInst_aarch64.hpp --- a/src/cpu/aarch64/vm/nativeInst_aarch64.hpp Tue Apr 29 14:58:56 2014 +0100 +++ b/src/cpu/aarch64/vm/nativeInst_aarch64.hpp Thu May 01 14:57:36 2014 +0100 @@ -54,7 +54,7 @@ class NativeInstruction VALUE_OBJ_CLASS_SPEC { friend class Relocation; public: - enum { instruction_size = BytesPerWord }; + enum { instruction_size = 4 }; inline bool is_nop(); bool is_dtrace_trap(); inline bool is_call(); --- CUT HERE ---
Hi, I noticed while looking at the assembly output from C2 that the loops were not being aligned correctly as in this short code section 0x0000007f8d170a84: mov w0, wzr 0x0000007f8d170a88: nop ;*iload_1 ; - dhry::execute@9 (line 8) <<<<< Incorrectly aligned here ;; B6: # B6 B7 <- B5 B6 Loop: B6-B6 inner Freq: 4.99998 0x0000007f8d170a8c: add x12, x10, w11, sxtw #2 0x0000007f8d170a90: ldr w12, [x12,#16] 0x0000007f8d170a94: add w0, w0, w12 ;*iadd ; - dhry::execute@15 (line 8) 0x0000007f8d170a98: add w11, w11, #0x1 ;*iinc ; - dhry::execute@17 (line 7) 0x0000007f8d170a9c: cmp w11, w1 0x0000007f8d170aa0: b.lt 0x0000007f8d170a8c The reason for this is that instruction_size is incorrectly defined in nativeInst_aarch64.hpp. It was defined as BytesPerWord which is of course 8. I have changed this to '4'. The code works out the no. of nops to emit as padding/instruction_size which explains why it emitted too few nops in the above. I have applied the patch below and the code is now generated as 0x0000007f88d83cc4: mov w0, wzr 0x0000007f88d83cc8: nop 0x0000007f88d83ccc: nop ;*iload_1 ; - dhry::execute@9 (line 8) ;; B6: # B6 B7 <- B5 B6 Loop: B6-B6 inner Freq: 4.99998 0x0000007f88d83cd0: add x12, x10, w11, sxtw #2 0x0000007f88d83cd4: ldr w12, [x12,#16] 0x0000007f88d83cd8: add w0, w0, w12 ;*iadd ; - dhry::execute@15 (line 8) 0x0000007f88d83cdc: add w11, w11, #0x1 ;*iinc ; - dhry::execute@17 (line 7) 0x0000007f88d83ce0: cmp w11, w1 0x0000007f88d83ce4: b.lt 0x0000007f88d83cd0 OK to push? Ed. --- CUT HERE --- # HG changeset patch # User Edward Nevill edward.nevill@linaro.org # Date 1398952656 -3600 # Thu May 01 14:57:36 2014 +0100 # Node ID f67f9b1b52ae8b1778dacb49df641bb5b6e48da1 # Parent 9d641fdeea4d1772617b097fc231dcff8e4aa634 Fix instruction size from 8 to 4