From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Mon, 14 Jul 2014 23:40:43 +0000 (+0000)
Subject: R600/SI: Default to no single precision denormals.
X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=5b70c8ac7e6e0f7ff746fecda0610a90375291c8;p=oota-llvm.git

R600/SI: Default to no single precision denormals.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213017 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/R600/AMDGPUAsmPrinter.cpp b/lib/Target/R600/AMDGPUAsmPrinter.cpp
index 8c5fc846ed4..b2b7bf9af44 100644
--- a/lib/Target/R600/AMDGPUAsmPrinter.cpp
+++ b/lib/Target/R600/AMDGPUAsmPrinter.cpp
@@ -47,10 +47,18 @@ using namespace llvm;
 // precision, and leaves single precision to flush all and does not report
 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
 // CL_FP_DENORM for both.
+//
+// FIXME: It seems some instructions do not support single precision denormals
+// regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
+// and sin_f32, cos_f32 on most parts).
+
+// We want to use these instructions, and using fp32 denormals also causes
+// instructions to run at the double precision rate for the device so it's
+// probably best to just report no single precision denormals.
 static uint32_t getFPMode(const MachineFunction &) {
   return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
          FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
-         FP_DENORM_MODE_SP(FP_DENORM_FLUSH_NONE) |
+         FP_DENORM_MODE_SP(FP_DENORM_FLUSH_IN_FLUSH_OUT) |
          FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
 }
 
diff --git a/test/CodeGen/R600/default-fp-mode.ll b/test/CodeGen/R600/default-fp-mode.ll
index 214b2c20369..4488bdb7f44 100644
--- a/test/CodeGen/R600/default-fp-mode.ll
+++ b/test/CodeGen/R600/default-fp-mode.ll
@@ -1,7 +1,7 @@
 ; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI %s
 
 ; SI-LABEL: @test_kernel
-; SI: FloatMode: 240
+; SI: FloatMode: 192
 ; SI: IeeeMode: 0
 define void @test_kernel(float addrspace(1)* %out0, double addrspace(1)* %out1) nounwind {
   store float 0.0, float addrspace(1)* %out0