diff options
| author | Shih-wei Liao <sliao@google.com> | 2011-06-15 12:38:29 -0700 |
|---|---|---|
| committer | Shih-wei Liao <sliao@google.com> | 2011-06-15 12:38:29 -0700 |
| commit | b460d4ba2d88c13480c2a6a8495de9d07a2c9f66 (patch) | |
| tree | d23c75f9822883e4e67805ca729ba339f13b2ffa /lib | |
| parent | ca65f4e8b7d5566d70a99273a4baabf4dbf08c05 (diff) | |
| download | external_llvm-b460d4ba2d88c13480c2a6a8495de9d07a2c9f66.zip external_llvm-b460d4ba2d88c13480c2a6a8495de9d07a2c9f66.tar.gz external_llvm-b460d4ba2d88c13480c2a6a8495de9d07a2c9f66.tar.bz2 | |
Fix the wrong encoding of vmov.32's immed8 value in LLVM's ARMInstrVFP.td.
Taking Renderscript Balls as example,
fc: eeb00a03 vmov.f32 s0, #3
120: edd00a02 vldr s1, [r0, #8]
128: ee600a80 vmul.f32 s1, s1, s0
VFPexpand in ARM's hardware will expand #3 above to 0x4e180000. This corresponds
to floating point value of 2.375. However, in the original Balls code it is 6.0.
As a result, Balls got smaller recently.
Note that 6.0 in F32 is 0x40c00000 and 6.0 in F64 is 0x4018000000000000! This leads
to how I fixed the bug.
Change-Id: I9a8e71771884dc27d5b9392abe7dfa541b45f357
Diffstat (limited to 'lib')
| -rw-r--r-- | lib/Target/ARM/ARMInstrVFP.td | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/lib/Target/ARM/ARMInstrVFP.td b/lib/Target/ARM/ARMInstrVFP.td index 840b880..a731731 100644 --- a/lib/Target/ARM/ARMInstrVFP.td +++ b/lib/Target/ARM/ARMInstrVFP.td @@ -1077,9 +1077,9 @@ def FCONSTS : VFPAI<(outs SPR:$Sd), (ins vfp_f32imm:$imm), // Encode instruction operands. let Inst{15-12} = Sd{4-1}; let Inst{22} = Sd{0}; - let Inst{19} = imm{31}; - let Inst{18-16} = imm{25-23}; - let Inst{3-0} = imm{22-19}; + let Inst{19} = imm{31}; // The immediate is handled as a double. + let Inst{18-16} = imm{22-20}; + let Inst{3-0} = imm{19-16}; // Encode remaining instruction bits. let Inst{27-23} = 0b11101; |