Subzero: fix alloca return address allocation
Only breaks in Om1 optimization mode because this code path is specific
to Om1. Broke after 01386d175 was merged because maxOutArgsSizeBytes()
now returns 32 bytes for the shadow store, so this is a rarely exercised
code path. Problem is that on x64, the new register being created needs
to match the size of the destination register, so 64 bits, not 32 as it
was assuming.
Fixes ReactorUnitTests Call_ArgsMixed, Call_ArgsPointer, IntrinsicTest_*
which were asstering with:
Assertion failed: !isScalarIntegerType(Dest->getType()) || (typeWidthInBytes(Dest->getType()) <= typeWidthInBytes(Source->getType())), file C:\src\SwiftShader2\third_party\subzero\src\IceInstX86Base.h, line 1308
Bug: b/145685985
Change-Id: I298edd5b4a286a242b351cc6267d70ecb37943fb
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/39169
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Antonio Maiorano <amaiorano@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
diff --git a/third_party/subzero/src/IceTargetLoweringX86BaseImpl.h b/third_party/subzero/src/IceTargetLoweringX86BaseImpl.h
index 5c312e1..becbbed 100644
--- a/third_party/subzero/src/IceTargetLoweringX86BaseImpl.h
+++ b/third_party/subzero/src/IceTargetLoweringX86BaseImpl.h
@@ -1570,7 +1570,7 @@
// Add enough to the returned address to account for the out args area.
uint32_t OutArgsSize = maxOutArgsSizeBytes();
if (OutArgsSize > 0) {
- Variable *T = makeReg(IceType_i32);
+ Variable *T = makeReg(Dest->getType());
auto *CalculateOperand = X86OperandMem::create(
Func, IceType_void, esp, Ctx->getConstantInt(IceType_i32, OutArgsSize));
_lea(T, CalculateOperand);
