Several of the branchXX() macro assembler functions in MacroAssemblerARM64.h that take an absolute address materialize the address into the memory temp register and then load the result into the same memory temp register without invalidating the corresponding temp register cache. For example consider the code snippet below: [ 167] put_to_scope loc1, maxDepth(@id6), loc2, 65537 0x15b0d1494: ldur x0, [x29, #-24] 0x15b0d1498: movz x17, #24056 0x15b0d149c: movk x17, #11360, lsl #16 0x15b0d14a0: movk x17, #1, lsl #32 0x15b0d14a4: ldrb w1, [x17, xzr] 0x15b0d14a8: cmp w1, #2 0x15b0d14ac: b.eq 0x15b0d14fc 0x15b0d14b0: movz x17, #24064 <= Full address for x17 materialized here and next two instructions 0x15b0d14b4: movk x17, #11360, lsl #16 0x15b0d14b8: movk x17, #1, lsl #32 0x15b0d14bc: ldr x17, [x17, xzr] <= Load into x17, the temp register cache entry for x17 should be invalidated 0x15b0d14c0: cmp x17, x0 0x15b0d14c4: b.eq 0x15b0d14fc 0x15b0d14c8: movk x17, #24057 <= This move changes only the lower 16 bits of x17 assuming the prior materialized contents are valid. 0x15b0d14cc: ldrb w16, [x17, xzr] 0x15b0d14d0: cbnz w16, 0x15b0d2ba8 0x15b0d14d4: mov w16, #0x2 0x15b0d14d8: movz x17, #24056 0x15b0d14dc: movk x17, #11360, lsl #16 0x15b0d14e0: movk x17, #1, lsl #32 0x15b0d14e4: strb w16, [x17] 0x15b0d14e8: movz x1, #0 0x15b0d14ec: movz x17, #24064 0x15b0d14f0: movk x17, #11360, lsl #16 0x15b0d14f4: movk x17, #1, lsl #32 0x15b0d14f8: str x1, [x17, xzr] 0x15b0d14fc: movz x17, #24720 0x15b0d1500: movk x17, #11345, lsl #16 0x15b0d1504: movk x17, #1, lsl #32 0x15b0d1508: str x0, [x17, xzr] One fix is to invalidate the cache when the destination of an absolute load is memory temp register
Created attachment 218344 [details] Patch for landing
Comment on attachment 218344 [details] Patch for landing r=me
Committed r160056: <http://trac.webkit.org/changeset/160056>